back to article IBM's Power7 servers imminent

IBM is planning to launch its Power7-based Power Systems servers sometime before the end of March. The new arrived Tuesday afternoon from chief financial officer, Mark Loughridge, during a conference call discussing Big Blue's fourth quarter and full year 2009 financial results. Previously, IBM had said that Power7 systems …

COMMENTS

This topic is closed for new posts.
Black Helicopters

p520 within the next three weeks.

But you don´t have this from me.

0
1

Cool!

It will be exciting to see real benchmarks on POWER7. As the POWER6 has peak 36 GFlops according to

http://www-06.ibm.com/jp/e-site/ssp/d6/d687868_BasicLecture.pdf

but scores lower than the 11 GFlops Niagara T2 on SPECfp_rate2006

http://www.sun.com/servers/coolthreads/t5440/performance.xml

we see that POWER6 reaches one third of it's theoretical max floating point performance.

I agree that POWER7 has good specifications, but it will be interesting to see if this POWER7, (which is a POWER6 derivative), behaves just like a POWER6, reaching only one third of it's theoretical performance in reality. The benchmarks will reveal this. Can't wait.

Of course POWER7 has it's strengths and will be fast on some workloads. There is no denial on that. But I want to see benches. Because one Niagara T2 is ~5 times faster than a POWER6, on certain multi threaded work (for instance SIEBEL v8), it will be fun to see if this new 8core high clocked POWER7 can catch up with the Niagara T2, which is also 8 core, 1.4GHz - on multi threaded work. But if 8core POWER7 is only twice as fast as the dual core POWER6, it will not happen. Then Niagara T2 will be faster on some work loads, and POWER7 will be faster on other loads.

2
2
FAIL

Kebabbert comparing again???

Kebabbert, please don't compare Sun Niagara with any IBM POWER processor.

They are different beasts.

Niagara do not scale to 32 sockets as POWER does.

The one that you could compare should be Sun ROCK processors , that will never appear in the market since it was cancelled.

So, keep your Niagara comparisons to x86 Nehalem.

They scale about the same!

Cheers!!!

PS: If you are willing to compare Sun against IBM, please use SPARC64 from Fujitsu instead.

0
1
Pint

Sigh...

Kebabbert. SPECfp_rate2006 is has nothing to do with GFLOPS. As the benchmark does not measure FLoatingpoint Operations Per Second, hence FLOPS.

And POWER7 is perhaps more like POWER5 than it is to POWER6 in design strategy.

And the predictions on how much faster POWER7 is compared to POWER6, then the factor that has been thrown around are somewhere between 5-5,5 times per chip i 'chip performance'

1 Thread:

SPECfp_2006:

POWER6 24,9

T2+ n/a

SPECint_2006:

POWER6 21,6

T2+ n/a

32 Threads

SPECfp_2006rate

POWER6 544,0

T2+ 68,5

POWER6 is 7,9 times faster

SPECint_2006rate:

POWER6 542

T2+ 97

POWER6 is 5,6 times faster

64 Threads

SPECfp_2006rate

POWER6 602,0

T2+ 133

POWER6 is 4,5 times faster

SPECint_2006rate:

POWER6 832

T2+ 183

POWER6 is 4,6 times faster

128 Threads

SPECfp_2006rate:

POWER6 2184

T2+ 270

POWER6 is 8.1 times faster

SPECint_2006rate:

POWER6 2155

T2+ 360

POWER6 is 6 times faster.

1 Chip

SPECfp_2006rate

POWER6 58,0

T2+ 68,5

Niagara is 1.2 times faster.

SPECint_2006rate:

POWER6 60,9

T2+ 97

Niagara is 1.6 times faster.

2 Chip

SPECfp_2006rate

POWER6 116

T2+ 133

Niagara is 1.1 times faster

SPECint_2006rate:

POWER6 122

T2+ 183

Niagara is 1.5 times faster

4 Chips

SPECfp_2006rate:

POWER6 222

T2+ 270

Niagara is 1.2 times faster

SPECint_2006rate:

POWER6 263

T2+ 360

Niagara is 1.4 times faster

8 cores

SPECfp_2006rate

POWER6 222

T2+ 68,5

POWER6 is 3.2 times faster.

SPECint_2006rate:

POWER6 263

T2+ 97

POWER6 is 2.7 times faster.

16 cores

SPECfp_2006rate

POWER6 544

T2+ 133

POWER6 is 4.1 times faster

SPECint_2006rate:

POWER6 542

T2+ 183

POWER6 is 3 times faster

32 cores

SPECfp_2006rate:

POWER6 602

T2+ 270

Power6 is 2.2 times faster

SPECint_2006rate:

POWER6 832

T2+ 360

Power6 is 2.3 times faster

So basically Niagara has a slim lead on POWER6 only when comparing chips to chips, and at ithe best result is a 60% lead and that is only against lower clocked POWER6 on SPECint_2006rate. On specfp_2006rate (16 way) each 8 core, 32 threaded T2+ chip is only 20% faster than 2 core 4 threaded POWER6+ chip.

Now you think that a POWER7 chip with four times the cores 8 times the threads, and fatter cores are only going to match a T2+ chip ?

Sorry but you are not being serious.

// Jesper

1
0
Silver badge
Happy

RE: Mauro & Jesper

RE: Mauro

"......So, keep your Niagara comparisons to x86 Nehalem. They scale about the same!...." Wash your mouth out! Nehalem scales higher, in fact x86/64 has already scaled higher for years, especially in such beasts as the Unisys boxen (the first generation ES7000 scaled to 32 sockets in 1999). But, Nehalem is probably a more valid comparison as x64 is the biggest competitor to Niagara in the webserving niche Niagara is struggling to survive in. But I agree that since Power can scale so much more than Niagara that it is silly to compare them.

RE: Jesper

".....Sorry but you are not being serious." The sad thing is Kebabfart actually is serious, but the good news is no-one else takes him seriously. Please don't remind him that to get T2 to multi-socket in T2+, where it finally made it to quad-sockets (the highest Niagara can scale to), they had to chop out half the memory controllers and halve the memory bandwidth, which makes T2+ even more of a performance dog than T2.

We are looking forward to getting a demo Power7 system to play with as our app teams are projecting a significant performance boost over our Power6+ systems. We have zero interest in any current Sun servers. As far as we're concerned, benchmarks are just performance indicators and it's much more important how the kit actually performs in our enviornment, with our stack, etc, etc (stop me if you're heard this before....). I don't have a delivery date yet, so looks like time to go harangue our IBM salesgrunt.

0
2
Happy

Power7 will be huge

And I am not talking chip size...no clue about the size nor do we care. It will be 8 cores with 4 simultaneous threads. Unlike the T2 which is simplified sparcII cores with alternating threads...these are the real deal. We think we can consolidate over 250 old UltraSPARC database servers per power7 box. We refuse to buy Fujitsu SPRAC64 chips since they will also be end of life after SPARC64 VII.

0
0

Slow but scalable?

Mauro M

Could you please explain to me, what is the point about scaling to many CPUs? I really dont get it. If you can scale to lots of POWER6, but still all of them together is slower than a few Niagara T2, so what is the point of having lots of POWER6? To burn more power?

I understand that scalability is good if you get good results. But no matter how many POWER6 you collect, you will STILL be slower than a few Niagara T2. So what is the point of collecting many POWER6, when the result is still slower? Please, explain that to me.

On SIEBEL v8, three IBM Power P570 servers with 12 of the 4.7GHz POWER6 CPUs is half as fast, as one SUN T5440 with 4 of the 1.4GHz Niagara T2. So assume you double the amount of POWER6, then you will get 24 of them. And then finally, those six P570 will match one SUN T5440. One P570 costs $413.000 and one T5440 costs $76.000.

Let us assume that POWER6 is even slower, then you have to collect, say, 50 of them to match one SUN T5440. So, yes, POWER6 scales better. BUT IT IS ALSO SLOWER. So what is the benefit of being able to collect lots of slow POWER6? If I can scale thousands to of 2MHz Z80 - what is the point? To burn more power? To burn more money? Are you weird? WHAT IS THE POINT OF SCALING SLOW CPUs, WHEN THEY DONT GIVE YOU ANYTHING??? Do noone of you think, at all?

.

Jesper Frimann,

"...So basically Niagara has a slim lead on POWER6 only when comparing chips to chips..." Are you joking? Who does NOT compare a CPU vs another CPU? Who compares ALU vs ALU? No one. If you try to establish which CPU is fastest, isnt it appropriate to compare... CPU to CPU? Or should we instead compare ALU to ALU, and then infer which CPU is fastest? Are you joking?

Regarding the POWER7, I am asking if the POWER7 will also achieve only one third of it's theoretical performance just as POWER6 does. Is that not a fair question?

Also, I have heard that POWER7 is basically a couple of stripped down POWER6. But you claim it is more similar to a POWER5? Do you have any links so I can read more on this?

Also, earlier I read that the IBM P595 which had the earlier TPC-C record costed $17 million. But you claimed that cost was not only the machine, but everything (clients, etc). Do you have any links on that, too?

.

Mattie Pattie Laddie,

Could you try to explain to me, what is the point of scalability when the CPU sucks? Why all the fuss about scalability? I mean, I prefer a few CPUs that are several times faster than lots and lots of slower CPUs, that still can not catch up! Who does not agree? If you can use 4 CPUs to match lots of other CPUs, which solution do you choose? Obviously the slower solution with many times more CPUs?? But hey, maybe you do. If you can claim something weird like "in my opinion that 2m guy is SHORTER than the other 1.5m guy", then you surely can choose a solution which uses way more CPUs, and costs 10 times more! Strange logic.

.

Allison Park,

You know, SUN migrated 251 Linux DELL dual cpu servers, than run 700 instances of MySQL down to 24 SUN T5440 machines. I bet you can consolidate way more than 240 old IBM servers on a modern SUN machine, maybe M9000. Dont you think?

.

Anyway, I am just trying to say that I would like to see benches and real numbers on the POWER7. Just like all you do. And see if it achieves as high as one third of it's theoretical numbers, just like the POWER6 does. Then I would like to pit it against the UltraSPARC 8-core Venus and see how much faster the Venus is. Let us continue when we see first benches on the POWER7. Let us see who is right. (Maybe IBM will refuse to publish benches, just like IBM refuses to publish benches on their dog slow Mainframes).

The point is, a traditional legacy 2.5GHz CPU idles 50% under _full load_ because of cache misses. A much higher clocked CPU, will be punished much more severely by cache misses. It will idle even more than 50% under maximum load. That is the reason a 5GHz POWER6 sucks, because 3 to 4 of the GHz vanish, they are spent on waiting on data - even under max load. Sure, the theoretical numbers may look great, but hey, on a high clocked CPU you will NEVER reach them. As evidenced by POWER6. One third is just really really bad. POWER7 is also legacy construct. On the other hand, the Niagara T2 only idles 5-10% under full load, all its clock cycles are doing useful work. Why do you the much slower clocked Niagara T2 can outclass a much higher clocked CPU, on certain work loads? Because of efficiency. That is why. T2 is much more efficient than the old legacy designed CPUs. No use of big cache, nor complex prefetch logic. T2 is a small lean fast CPU. Which is evidenced by benches.

1
3
Silver badge
FAIL

somebody didn't get the memo

>You know, SUN migrated 251 Linux DELL dual cpu servers, than run 700 instances of MySQL down to 24 SUN T5440 machines. I bet you can consolidate way more than 240 old IBM servers on a modern SUN machine, maybe M9000. Dont you think?

Whats the point of buying a modern Sun machine when in a years time they will only sell x86_64 at a retarded high margin? Sparc is dead and everyone in the industry knows it. In fact migrating apps from SPARC to x86 for my current Sun fanboi employer (even they know the writing is on the wall) is paying a lot of my salary right now.

0
0
Pint

What the...

"On SIEBEL v8, three IBM Power P570 servers with 12 of the 4.7GHz POWER6 CPUs is half as fast, as one SUN T5440 with 4 of the 1.4GHz Niagara T2. So assume you double the amount of POWER6, then you will get 24 of them. And then finally, those six P570 will match one SUN T5440. One P570 costs $413.000 and one T5440 costs $76.000."

Yes, we do think, yes many of us do have university degrees in Math, Computerscience, physics whatever. And many of us do have many years of experience in the business, opposed to you. Now I hav already on this forum, made a walkthrough of the Siebel bench

mark you are mentioning and showed you that yes, the T5440 is a very good one. BUT no where in any way just nearly as good as you keep trying to make it. You are maing a fool of yourself. Sorry.

"...So basically Niagara has a slim lead on POWER6 only when comparing chips to chips..." Are you joking? Who does NOT compare a CPU vs another CPU? Who compares ALU vs ALU? No one. If you try to establish which CPU is fastest, isnt it appropriate to compa

re... CPU to CPU? Or should we instead compare ALU to ALU, and then infer which CPU is fastest? Are you joking? It is very relevant. Why ?

CAUSE YOU USE CORES WHEN YOU DO CALCULATE SOFTWARE LICENSES!!! HOW MANY TIMES DO WE HAVE TO TELL YOU.

And thread level performance is also important, cause single threaded throughput is still a very important factor. One factor that all to many CPU's aren't really doing enough to adress IMHO. Damn.

Try doing some math here:

http://www.oracle.com/corporate/contracts/library/processor-core-factor-table.pdf

or

http://www-01.ibm.com/software/lotus/passportadvantage/pvu_licensing_for_customers.html

"Regarding the POWER7, I am asking if the POWER7 will also achieve only one third of it's theoretical performance just as POWER6 does. Is that not a fair question?"

*SIGH* Sure it is, but stop mixing the benchmarks up and using GFLOPS when talking about specfp_2006rate. It makes no sense. And you still don't understand virtualization. Which is perhaps the most important thing forcing up utilization and thus making sure that a system is well utilized.

"Also, I have heard that POWER7 is basically a couple of stripped down POWER6. But you claim it is more similar to a POWER5? Do you have any links so I can read more on this?"

Yes, if you read this presentation:

http://www.ibm.com/developerworks/wikis/download/attachments/104533501/POWER7%2B-%2BThe%2BBeat%2BGoes%2BOn.pdf

Note that POWER7 is a OO core, with no less that 12 execution units.. 12.. it's a fat core. It also has 4 way SMT, not coarse grained multithreading like Niagara.

"Also, earlier I read that the IBM P595 which had the earlier TPC-C record costed $17 million. But you claimed that cost was not

only the machine, but everything (clients, etc). Do you have any links on that, too?"

Yes, it's here:

http://www.tpc.org/results/individual_results/IBM/IBM_595_20080610_ES.pdf

Mattie Pattie Laddie,

Could you try to explain to me, what is the point of scalability when the CPU sucks? Why all the fuss about scalability? I mean,

I prefer a few CPUs that are several times faster than lots and lots of slower CPUs, that still can not catch up! Who does not

agree? If you can use 4 CPUs to match lots of other CPUs, which solution do you choose? Obviously the slower solution with many times more CPUs?? But hey, maybe you do. If you can claim something weird like "in my opinion that 2m guy is SHORTER than the ot

her 1.5m guy", then you surely can choose a solution which uses way more CPUs, and costs 10 times more! Strange logic.

The whole point is that it doesn't suck, power6 is damn fast. The project I am responsible for has no less that 16 power 570'es and a few p595's. And damn they are fast, and actually quite forgiving due to their high Ghz. Again if there is a CPU core that sucks then it is the T2. Now both Itanium and POWER has managed to keep One of the reasons why Itanium still sells fairly well is that it's single threaded performance actually is pretty good. And it will, IMHO, get better with Tukwila. And you still can't get it into your thick head that the one machine with the myriad of CPU cores is the T5440. It has bloody 32 cores with no less than 256 threads. Man that is half the threads of a maxed out M9000.

.

"You know, SUN migrated 251 Linux DELL dual cpu servers, than run 700 instances of MySQL down to 24 SUN T5440 machines. I bet you can consolidate way more than 240 old IBM servers on a modern SUN machine, maybe M9000. Dont you think?"

yeah Yeah, you keep quoting that story everywhere U go. First it is something you picked up from BMSEER here:

http://blogs.sun.com/bmseer/entry/mysql_consolidation_on_sun_sparc

And you couldn't even get the machine right, it wasn't a T5440, it was a T5220. Do you bloody research before wasting our time.And just quoting BMSEER isn't really particular serious, ever wondered why he/she isn't blogging anymore ?

He/she got fired, cause Oracle took over the business.Now you don't see me quoting, without using my head, some of the IBM/HP/DELL marketing slides. And the M9000 has even worse per

core performance with the quad core chips, compared to the T5400 when using large Core count. Damn..

Anyway, I am just trying to say that I would like to see benches and real numbers on the POWER7. Just like all you do. And see if it achieves as high as one third of it's theoretical numbers, just like the POWER6 does. Then I would like to pit it against the UltraSPARC 8-core Venus and see how much faster the Venus is. Let us continue when we see first benches on the POWER7. Let us see who is right. (Maybe IBM will refuse to publish benches, just like IBM refuses to publish benches on their dog slow Mainframe)

Again you are just rampling. There will be no SUN Venus box. Read their roadmaps. And yeah lets see the benchmarks and lets see a 128 Node Power7 box benchmark with DB2 Purescale.. now that would be fun. And be sure that POWER7 will blow anything that SUN can come up with out of the water and still have plenty of firepower left. Just wait and see.

"The point is, a traditional legacy 2.5GHz CPU idles 50% under _full load_ because of cache misses. A much higher clocked CPU, will be punished much more severely by cache misses. It will idle even more than 50% under maximum load. That is the reason a 5GH

z POWER6 sucks, because 3 to 4 of the GHz vanish, they are spent on waiting on data - even under max load. Sure, the theoretical numbers may look great, but hey, on a high clocked CPU you will NEVER reach them. As evidenced by POWER6. One third is just really really bad. POWER7 is also legacy construct. On the other hand, the Niagara T2 only idles 5-10% under full load, all its clock cycles are doing useful work. Why do you the much slower clocked Niagara T2 can outclass a much higher clocked CPU, on certain work loads? Because of efficiency. That is why. T2 is much more efficient than the old legacy designed CPUs. No use of big cache, nor complex prefetch logic. T2 is a small lean fast CPU. Which is evidenced by benches."

Again you have no idea what you are talking about. Have you ever made a solution or even been logged into one of the boxes you are talking about ? or are you just been overdosing on BMSEER. No wonder why the rest of the posters on this page, simply ignore you.

And yes the T2 is efficient.. it is .. it is a GREAT CPU for running many light independent threads which share the same code, so you can benefit from shared libraries.. but.. well... Xeons are just better at that. Sorry. And much chearper.. and has more RAS features and and and..

// Jesper don't wanna waste more time.

1
0
Anonymous Coward

Yo Matt!

Let me introduce you to a rather old computer concept: Memory interleaving!

Those links you claim to be stolen on a T2 for scalability *are* links to memory.

And if the local data is in another nodes $, this is the memory link for that line!

Duh!

1
0
Silver badge
FAIL

RE: Slow but scalable?

And Kebabfart displays yet another gross misunderstanding of the marketplace. Most mainframe buyers originally bought mainframes because it was the only platform that scaled large enough to do the job, or the only platform that ran the required application, or the only platform that offered the high levels of up-time required. They usually keep buying beacuse of the latter two reasons nowadays. Any attempt to compare any of the current Niagara servers to mainframes will just get you laughed out of the room.

As for all your waffling about cache misses, you forget that cache misses are just one factor in slowing down a CPU, but memory bandwidth is even more important. Which means the memory-crippled T2+ really is a step backwards over even the T2, which makes it even less of a competitor to real CPUs like Power, Itanium, Opteron or Xeon.

You say you'd rather have one CPU that can do the job rather than many that struggle, and history shows that Sun's own customers agreed. That's why Sun was forced to bring out the SPARC64-based M3000, beacuse the Niagara kit just couldn't replace the old SPARC boxes to the customers' satisfaction. In short, Sun's own customers preferred one real CPU, even a SPARC64, to any amount of multi-cored Niagaras!

/SP&L.

0
2
Happy

Matt Bryant

Ha! Matt still trying to convince anyone that will listen that he has a job in IT... Too funny.

1
0
Unhappy

Re: Kebabbert

"Also, earlier I read that the IBM P595 which had the earlier TPC-C record costed $17 million. But you claimed that cost was not only the machine, but everything (clients, etc). Do you have any links on that, too?"

You are always so quick to quote benchmark numbers, yet you've never been to the TPC website, and looked at a disclosure!? Shocking!

http://www.tpc.org/results/FDR/TPCC/IBM_595_20080610_fdr.pdf

Up to this point, I thought your posts were full of your own insights.

0
0
FAIL

@Matt / @Keba

Matt, I was referring to the available x86 Nehalem (5500 series) that is a dual processor architecture (see http://www.intel.com/Assets/en_US/PDF/prodbrief/xeon-5500.pdf and learn more about it).

Keba, I will not discuss this point with you anymore... You are correct, POWER6 cores are slow, that's why T2 machines are real blockbusters and Sun is making a lot of money from it. Besides that, Sun employees are really happy and there are no job cuts at this company. and the market really enjoys Sun servers roadmap with no delays or cancels. What a great company to invest your IT money on!

1
0
WTF?

Kebabbert, you can't be serious???

Ok, is it me or do others think he is actually serious??

We have a variety of servers in my shop and I administer the UNIX/Linux boxes. We have a couple of T2 servers along with a couple of P6 570 servers. No way is the T2 faster than the P6 570 servers (we are running Cache database on the 570's). Don't get me wrong, but the T2 is a different class machine. It is a great web server and a good small application box, but not even close to the 570.

So he must be joking or drinking the Sun kool-aid. I hope he is joking.

0
0
Flame

TheRegister Players

Matt Bryant - Long time contributor - Loves HP, but every post he has come to respect and might end up being a huge Power fan.....You have to respect him for that...and that is a lesson to the rest of you Sun fans..

Kebabbert - Sun lover/employee who is fixated on a Siebel benchmark and I think was embarrased when Oracle published the tpc-c benchmark and said ti was stupid of larry

David Halko - Sun Lover who relies on Sun systems to his living so there is no better system

I feel bad for all the Sun folks that will lose their jobs in the coming weeks and months as Oracle will not stay in the hardware business. Larry spent $5.5B for Java, Solaris, killing MySQL and the maintenance stream.

Hopefully Larry's big talk to the joint company on Jan 27th will be uplifting...this place is a disaster. I hope i get letter #2...will take the summer off then find a nice place to hang my hat.

cheers and it was fun while the fun lasted.....which was about 2001

0
0

Half of peak performance, in reality?

asdf

"...What is the point of buying a modern SUN machine when in a years time they will only sell x86_64 at a retarded high margin?..."

Are you joking? Why do you say that SPARC has a high margin, but dont even mention POWER? That is ridicilous as the POWER cost 5-10x times more than SPARC, for less performance!

For instance, on SPECjAppServer(R)2004 benches we see that two Niagara machines with one 1.4GHz CPU each, gets 70% better performance than two IBM POWER servers. Just the cost of the licensed cores in the P570 used as the J2EE server is 1.6 times the list price of the two SUN servers, delivered, set-up and fully warranted for three years!

http://blogs.zdnet.com/Murphy/?p=1537

As I also shown, the SUN T5440 costs $76.000 and one P570 costs $413.000, and you need six P570 to match one T5440 on SIEBEL!

.

Jesper Frimann

No, lots of you IBMers here, dont have university degrees in math, Physics, etc - that is evident from your posts. Why is it evident? We talk about which CPU is the fastest, right? Niagara or Power6. SO WHY DO YOU KEEP TALKING ABOUT LICENSE COST OF CORES??? "HOW MANY TIMES DO I HAVE TO TELL YOU"??

I dont know how many times I have explained again and again, that when we talk about which CPU is fastest, then it doesnt matter whose core is fastest, or whose ALU is fastest. If we try to discuss which CPU is fastest, then we should discuss which CPU is fastest.

If you are discussing licensing prices, then your claim is relevant. But my post is about the performance of the POWER6 and POWER7 vs Niagara.

.

Regarding if POWER7 is a legacy designed POWER6 derivative or a POWER5 derivative, you posted a link. I have skimmed that link, but I would be glad if you pointed out on which page it says the POWER7 is a POWER5 derivative.

.

Regarding the $17 million cost of the IBM P595 system that earlier had the TPC-C record. I see that all clients + client's software adds upp to 0.5 million. I think it is safe to say that the rest of the $17 million sum, is only for the P595 server then. I mean, that is quite a wicked pile of money, dont you think? $16.5 million!! For one sucky POWER P595 server!!! Who spoke about SPARC having large margins? *chuckles* Actually, without the discounts, ONLY the P595 server including it's sofware costs $35 million!!!!!! "SPARC large margins" - eh? :o) :o) $35 million! Jesus! (Thanks for the link :o)

.

"...The whole point is that [POWER6] doesn't suck, power6 is damn fast. The project I am responsible for has no less that 16 power 570'es and a few p595's. And damn they are fast, and actually quite forgiving due to their high Ghz..."

You just dont know that a high Hz does severely punish cache misses?

http://forum.beyond3d.com/showthread.php?t=54889

"Also IBM really needs POWER7 to be very good. POWER6 is not very impressive despite its crazy specifications. For example, a 4 CPU POWER6 4.2GHz machine (total 8 cores) performs similar to a 2 CPU Nehalem 2.93GHz machine (also total 8 cores) in TPC-C. Although POWER6 is older but Intel is closing very fast and that really worries IBM."

http://www.theinquirer.net/inquirer/news/1531766/ibm-power7

"Yes, POWER6 was inefficient in Linpack and similar benchmarks for HPC - roughly 60 per cent achievable out of theoretical peak is really bad for a modern CPU"

Also, on SAP we see that AMD Opterons is faster than 5GHz POWER6

http://www.hpcwire.com/offthewire/Sun-Microsystems-Sets-Records-with-N ew-x64-Server-78779357.html

Eight AMD opterons is 2.7x faster than four POWER6. A very expensive POWER6 can not justify it's much higher price against one AMD opteron.

The POWER6 is inefficient, it NEEDS 5GHz to match one Intel/AMD CPU. Let us clock POWER6 down to 1.4GHz and see how good it is then.

So how about the POWER7? Will it also suck? It has great specs, yes. Maybe it will even be faster than Nehalem on some things!

"I am sure Power7 will be the fastest processor around, probably faster than Intel's Nehalem in some benchmarks,"

http://www.eetimes.com/news/semi/showArticle.jhtml?articleID=219400955

What you dont understand, is that SOLARIS exists on Niagara, on SPARC and on x86! There will always be a good machine for any work load. Whereas AIX only exists on the POWER7 which is good for some things (single threaded work, where all data fits into a small cache - this is not characteristic for server work load), but sucks on other things (server workloads with lots of clients where all client's data does not fit into cache -> cache miss) where T2 excels. SOLARIS wins no matter the outcome, no matter which CPU wins.

.

Regarding the consolidation of 251 Dell Linux servers down to 24 SUN T5220, even better. The T5440 has four CPUs. The T5220 has only two Niagaras. Probably it would suffice with just 12 of the T5440. One POWER6 have trouble beating one Intel or AMD. I bet POWER6 could not consolidate all these dual cpu intel Linux servers. Because POWER6 is good at single threaded work, but not on this work.

.

"...Again you are just rampling. There will be no SUN Venus box. Read their roadmaps...." Bull shit. We both know that Oracle will decide. Oracle has not presented any decisions yet. You are just rampling, here. Maybe there will be a Venus box.

.

"...Again you have no idea what you are talking about...[high clocked CPUs are more punished by cache misses than low clocked CPUs]" Let me inform you, it IS true that high clocked CPUs are more punished than low clocked CPUs. The RAM is slow, but CPUs are fast. That means if CPU get a cache miss, then it has to idle many cycles, waiting for slow RAM. That is the reason POWER has big caches. And that is the reason they ramp up memory bandwidth so much. POWER tries to avoid cache misses at ALL costs, because misses are sooo costly for a high clocked CPU.

You need a radical design to circumvent this problem, which Niagara T2 does. It does not need large cache to be several times faster than POWER6. Nor high bandwidth.

.

It reminds me of when I showed links that Niagara is 13 times faster than a IBM CELL on large workloads (string pattern matching). Then you chimed in and said something like "that is not relevant, the CELL is 70% faster on small workloads!". Jesus. Who is interested in small workloads where everything fits into the cache? Real life workloads are always big and dont fit into the cache. Hence you get cache misses, and it start to thrash. Which is a BAD thing. Which is evidenced that CELL looses performance really fast, down to less than 5% of it's former peak, when it's cache starts to thrash. Maybe POWER6 has equal behavior.

.

Mattie Pattie Laddie

When you talk about Mainframes, yes the old customers keep bying - because of vendor lockin. But we talk about CPU performance, they suck big time. If you could use a Nehalem to run native Mainframe code, you would need 14 Nehalem to match 64 of the "uuuuuberfaaast" Mainframe CPUs. So new customers are not likely to buy them. Just look at the wikipedia article on "turboherkules".

"...In short, Sun's own customers preferred one real CPU, even a SPARC64, to any amount of multi-cored Niagaras!..." This is not true. Niagara is growing much faster than SPARC64, something like 30-40% each year. So this is a pure lie. Again.

.

Mauro M.

Thank you, for agreeing that POWER6 is slow. You must have finally read my links, showing benchmarks where we see that POWER6 is slower than Intel/Niagara T2s. We both agree and there is nothing else to say.

.

Anonymous Coward,

Yes, of course you have some T2 servers and some P570 servers. And of course the P570 servers are much faster. Yes. And on MY job, which is in a large investment bank/stock exchange/nuclear plant/pick your fancy, we see that one T2 machine is several times faster than P570 on multithreaded workloads. What I am trying to say, I dont believe you. I bet you are the person always writing that he works on a large investment bank/exchange/whatever and they are now migrating from his beloved SPARC to POWER. Just lies.

I will do the same thing from now on, to balance all your lies. I will write similar posts like you do, claiming that I work on a large investment bank/exchange/whatever sounding good, and we are migrating from POWER to SPARC.

I have actually read here in the comments (now, this is actually true that I have read this) that there are 4 solaris machines shipped for every AIX machine. If this is true, then no way people are migrating from POWER to SPARC so much as falsely claimed here.

I dont see the point in lying? I rely on official verifiable benches, you IBMers FUD and lie and twist everything. We discuss which CPU is fastest, you try to move focus away, to core pricing. Jesus. You always write the SAME thing: "work at a large bank/telecom company/etc which are loving SPARC and now migrating to POWER". No phantasy at all. The same lie all the time.

.

Anonymous Coward,

Why would I be embarrased because Oracle has the TPC-C record right now (via SUN gear)?

1
1
FAIL

@Keba

Kebabbert, I was not agreeing with you, I was only being ironic.

Only you thinks that Niagara performs well against high-end processors.

But, the market speaks for itself.

Die slowly with Sun

Cheers...

0
1
WTF?

The ravings of a lunatic

In a world of blind people the one eyed man is King, unfortunately the world is not blind so the one eyed man is just, well, one eyed.

Peeps you must remember the rules when talking to our Special friend Mr K - he's very special... about his "debates". They often turn into mass debates......

1) You are not allowed to express opinion or personal experience despite your industry knowledge or accumen (I wonder if Larry expressed his Opinion what the reaction would be??)

2) You are only allowed to use SUN friendly benchmarks from a SUN approved source (i.e. BMSeer)

3) You are not allowed to use any irony or wit (its deadly serious this IT stuff)

4) All people who use non Siebel related benchmarks that do not show T2 Niagara CPU's in a good light to back up their own arguements will be shot. These are plainly lies subliminally implemented by the IBM thought Police.

5) He's always right - the rest of us are completely brainwashed by IBM or HP and he is the paragon of the IT world - all hail

Now Mr K tootle off and play World of Warcraft, watch Avatar or whatever - its obvious you don't know how to work with people.

1
1
Thumb Down

Kebabbrain (1 of 2)

Kebbabbbbbbrain this is getting tiresome.

"For instance, on SPECjAppServer(R)2004 benches we see that two Niagara machines with one 1.4GHz CPU each, gets 70% better performance than two IBM POWER servers. Just the cost of the licensed cores in the P570 used as the J2EE server is 1.6 times the list price of the two SUN servers, delivered, set-up and fully warranted for three years!"

http://blogs.zdnet.com/Murphy/?p=1537"

Again you don't even understand the benchmark. You wouldn't even be dangerous if you were clever. Again no clue about what you are talking about. Quoting murphy is like quoting BMSeer, that is where he gets his ammo.

Why don't you try to do some basic research about the benchmarks you are quoting rather than just quoting others. People who again get their information from BMSeer.

First this is the Benchmark where a T5XXX box has to shine, this is what it is designed for. So you cherry pick a benchmark Fair enough.

1) Java version the T5220 on the benchmark runs Java version 1.6.0_03 the POWER 570 runs 1.5.0. If you have a look at the JBB2005 benchmark you will see that for POWER going from 1.5->1.6 will give you a boost of aprox 25%. Perhaps even more.

2) You keep mentioning two power 570'es. Actually it is one POWER 570 and one p570. The p570 is an oldish POWER5+ machine.

3) Both POWER servers are of the 570 series. But none of them are more than a 1 CEC version. It's not a fully configured power 570/p570. It could just as well have been a POWER 520, the smallest non POWER blade server.

4) The T5220 machines uses 10Gbit network interfaces, the POWER solution uses 1Gbit hardware.

5) The T5220 database machine uses a the POWER database server actually uses 2Gbit connections.

6) The T5220 Database server uses 64GB of RAM, the POWER p570 used 8GB.

7) The T5220 APP server uses 32GB of RAM the POWER 570 uses 16GB of RAM.

8) The response times on the two systems is very different, on the POWER setup the MAX request time is only 43% of what it is on the T5220 setup. For example the Maximum response time for Browse is 22+ seconds on the T5220, which is 5 times worse than the POWER setup.

9) If you knew just a little about UNIX you would quickly see that the people who setup the T5220's have done a much better job than the ones setting up the POWER servers. The most basic network and AIX java tuning hasn't been done. It's right in the manual:

(http://publib.boulder.ibm.com/infocenter/aix/v6r1/index.jsp?topic=/com.ibm.aix.prftungd/doc/prftungd/java_tuning_aix.htm)

So ... yes... sure you can compare but please stick to the facts.

"If you are discussing licensing prices, then your claim is relevant. But my post is about the performance of the POWER6 and POWER7 vs Niagara."

Eh.. I try to educate you, it is hard, cause you constantly resist knowledge. But HW cost is only a small portion of the cost of a system. Let me cut it out in cardboard for you. Perhaps that will help you.

It's like buying a car sold to you by SUN that cost 1000$ but will only run 10 mile to the gallon, compared to buying a 10.000$ car from HP/IBM that will run 50 miles to the gallon. Which is kind of a bitch if you have to drive 40.000 miles a year and the gas is 8$ a gallon. Because after driving 14000 miles the cheaper SUN car starts to cost you money.

Yes there are people that does this kind of thing all the time, they also have massive credit card misuse and a lot of unpaid bills.

"Regarding if POWER7 is a legacy designed POWER6 derivative or a POWER5 derivative, you posted a link. I have skimmed that link, but I would be glad if you pointed out on which page it says the POWER7 is a POWER5 derivative."

I wrote it in my post. It's a question about filosofy. POWER6 is a speed devil processor, much like POWER4. POWER7 is a braniac, just like POWER5. POWER6 is an mostly inorder processor, POWER7 is a OO processor.

Regarding the $17 million cost of the IBM P595 system that earlier had the TPC-C record. I see that all clients + client's software adds upp to 0.5 million. I think it is safe to say that the rest of the $17 million sum, is only for the P595 server then. I mean, that is quite a wicked pile of money, dont you think? $16.5 million!! For one sucky POWER P595 server!!! Who spoke about

SPARC having large margins? *chuckles* Actually, without the discounts, ONLY the P595 server including it's sofware costs $35 million!!!!!! "SPARC large margins" - eh? :o) :o) $35 million! Jesus! (Thanks for the link :o)

Begezus, you can't even read. Man and me who is trying to convince the wife to buy a house in sweden. Hmm.. Nahh.. Swedish education system is good I know it.. I choose to belive in my swedish roots. You just turned to the last page and didn't understand what it said.

Now if you want expensive then try to look at your precious SUN TPC-C submission:

http://www.tpc.org/results/individual_results/Sun/Sun_T5440_TPC-C_Cluster_ES_122309_v2.pdf

On page 2, they haven't even let you buy the software.. you only lease it.. and the support is Metalink only, no upgrade protection, no nothing. Just lease of the software. And for that you pay 2.6MUSD each year. It's the 10 miles a gallon car all over again. Buy some terrible entry servers and put expensive software on them.. yeah.. it's money out of the window. And you don't get

it. The Oracle software costs just as much to lease for a year, as DB2 costs in five years if you buy the software, and that is with full support.

Now here is the real reason why Oracle bought SUN.

"You just dont know that a high Hz does severely punish cache misses?

http://forum.beyond3d.com/showthread.php?t=54889"

Ever heard about SMT ?

And if you think that 2 threads running at 5GHz is bad for the caches ?

How do you think that 8 threads running at 1.6GHz with even less cache ?

Let os again do the math.

One POWER core has 128KB L1 cache 4096 KB L2 cache and 16384 KB L3 cache. This means that it has:

12.8KB per thread/GHz L1 cache

410KB per thread/GHz L2 cache

1638KB per thread/GHz L3 cache.

One Niagara core has 24KB L1 cache 512 KB L2 cache and no L3 cache. This means that it has

1.9KB per thread/GHz L1 cache.

40 KB per thread/GHz L2 cache

0 per thread/GHz L3 cache.

Ever heard the expression that you should throw stones when you live in a house made of glass ?

"<Cut out some strange ramplings that didn't make sense>

Also, on SAP we see that AMD Opterons is faster than 5GHz POWER6

http://www.hpcwire.com/offthewire/Sun-Microsystems-Sets-Records-with-New-x64-Server-78779357.html

Eight AMD opterons is 2.7x faster than four POWER6. A very expensive POWER6 can not justify it's much higher price against one AMD opteron."

Again you can only figure out how to read marketing announcements from SUN. Damn.. I have to ask.. I have a large steel construction in Paris that I want to sell.. it's kind of a tower, it is very nice, you want to buy ?

It's 48 cores and the only really suitable benchmark to compare it to is the power 550 which does around 40% of the 10.000 users with a 1/6 of the cores. So it's still 2.5 times faster per core. But ofcause it starts to fall behind now as it only has 2 cores per chip. Nothing strange about that.

And again nice cherry pick, take the newest just announced x86 AMD server and here in 2010 compare it to a machine from 2007, with it's replacement being announced right around the corner. A POWER7 box with half the sockets will outperform the X4640 by a factor of 2. Again cheap marketing shot, without any real thought behind it.

You are stating the obvious, that given the time with twice the number of sockets and 6 times the cores a SUN x86 system can outperform a POWER system.

...

1
0
Pint

Kebabbrain (2 of 2)

....

And why do you all of a sudden choose a x86 server. Why not use the T5440 ? Why cause a 256 Thread 32 Core 256GB T5440 does a medicore 4720 SAP users on the same benchmark.

SUN T5440 with 32 cores and 256GB RAM

4720 USERS/18 users per thread/148 users per core.

SUN X4640 with 48 cores and 256GB RAM

10000 USERS/208 users per thread/208 users per core.

Power 550 with 8 cores an 64 GB RAM

3752 USERS/232 users per thread/464 users per core.

"The POWER6 is inefficient, it NEEDS 5GHz to match one Intel/AMD CPU. Let us clock POWER6 down to 1.4GHz and see how good it is then."

It will most likely be very good, and very efficient. But how come that all the benchmarks that the T5440 does it has many many times the Memory of any other system that it is compared to ? It's a factor of 3 on the above SAP benchmark, a factor of 8 on the SPECjAppServer(R)2004 benchmark you mentioned...

And what do you think that the Oracle price is for the DB on the above SAP benchmark as it is a 2-tier benchmark ?

It's bloody 24 licenses versus 8 for the power 550.. and they do cost 40KUSD+ a piece...

"So how about the POWER7? Will it also suck? It has great specs, yes. Maybe it will even be faster than Nehalem on some things!

<Deleted strange stuff>"

See the cache calculations above. 'nuff said.

"Regarding the consolidation of 251 Dell Linux servers down to 24 SUN T5220, even better. The T5440 has four CPUs. The T5220 has only two Niagaras. Probably it would suffice with just 12 of the T5440. One POWER6 have trouble beating one Intel or AMD. I bet POWER6 could not consolidate all these dual cpu intel Linux servers. Because POWER6 is good at single threaded work, but not on this work."

You have no idea of what you are talking about. There is a think called overcommitment on the POWERVM hypervisor. On my little serverfarm we run with a overcommitment factor of 3. Hence we use a 16 way POWER 570 as it was a 48 way server. On our internal machines, we use a factor of 5, hence a 16 way machine becomes a 80 way. machine. I don't expect you to understand, but that is how it is.

"Bull shit. We both know that Oracle will decide. Oracle has not presented any decisions yet. You are just rampling, here. Maybe there will be a Venus box."

Well lets see then. But they had better do a better job than they did going from 2->4 cores. On SPEC OMPL2001, the scaling was an amazing 18% for double the number of cores. And an amazing 13% if you correct for an increase in GHz.

http://www.spec.org/omp/results/res2007q2/omp2001-20070417-00252.html

and

http://www.spec.org/omp/results/res2008q3/omp2001-20080714-00308.html

"You need a radical design to circumvent this problem, which Niagara T2 does. It does not need large cache to be several times faster than POWER6. Nor high bandwidth."

BLEH it does it by magic powder right. The T2 design is good, but not ground breaking. And it still has less cache, less memory bandwidth per 'click' than POWER6.

....

"It reminds me of when I showed links that Niagara is 13 times faster than a IBM CELL on large workloads (string pattern matching). Then you chimed in and said something like "that is not relevant, the CELL is 70% faster on small workloads!". Jesus. Who is interested in small workloads where everything fits into the cache? Real life workloads are always big and dont fit into the cache. Hence you get cache misses, and it start to thrash. Which is a BAD thing. Which is evidenced that CELL looses performance really fast, down to less than 5% of it's former peak, when it's cache starts to thrash. Maybe POWER6 has equal behavior."

Heh, what does cell has to do with POWER6, they are two radical different designs. And what I said was that you cherry picked the results that fit you. And threw the ones that you didn't away. You didn't bother to paint the whole picture. You only mentioned the stuff that fit your purpose. Or rather that is giving you to much credit, cause you got it from some marketing material where you get all your claims from. Again as you do now. You paint everything black and white. Things are grey in grey, in the real world. And I have maintained the whole time that the T5XXX is an excelent box if you use it for what it is designed for. And I still mean that. You have to do research yourself. Stop echoing Snoracle marketing material.

"When you talk about Mainframes, yes the old customers keep bying - because of vendor lockin. But we talk about CPU performance, they suck big time. If you could use a Nehalem to run native Mainframe code, you would need 14 Nehalem to match 64 of the "uuuuuberfaaast" Mainframe CPUs. So new customers are not likely to buy them. Just look at the wikipedia article on "turboherkules"

You don't get it, there is a big difference when having a z10 running zOS with DB2 in Parallel Sysplex, with CICS and exits written in PL/1 or COBOL perhaps with embedded assembler, it all running on a pair of 32 core server with 95% average utilization.

It is almost impossible for any Power/Sparc/Itanium/x86 box to compete with this. Both with regards to throughput and availability. Try to think about why.

" This is not true. Niagara is growing much faster than SPARC64, something like 30-40% each year. So this is a pure lie. Again."

Yeah, and it's always nice to make predictions based upon a graph that consists of 2 points. Yeaaaahhh... go go.

"Yes, of course you have some T2 servers and some P570 servers. And of course the P570 servers are much faster. Yes. And on MY job, which is in a large investment bank/stock exchange/nuclear plant/pick your fancy, we see that one T2 machine is several time

s faster than P570 on multithreaded workloads. What I am trying to say, I dont believe you. I bet you are the person always writing that he works on a large investment bank/exchange/whatever and they are now migrating from his beloved SPARC to POWER. Just lies."

Oh, so you actually have a power 570 box at work ? What is the serial number then ? What is the output from a 'lsdev -Cc processor', not that we don't believe you...

"I have actually read here in the comments (now, this is actually true that I have read this) that there are 4 solaris machines shipped for every AIX machine. If this is true, then no way people are migrating from POWER to SPARC so much as falsely claimed here."

I think it was this that you refered to was someone that posted this:

IDC Server Tracker WW Q309 - M9000 (Fujitsu & Sun)

2008Q2 2008Q3 2008Q4 2009Q1 2009Q2 2009Q3

183 129 221 159 165 90

IDC Server Tracker WW Q309 - Power 595 (IBM):

2008Q2 2008Q3 2008Q4 2009Q1 2009Q2 2009Q3

149 211 545 246 380 328

But as to number of servers, then yes SUN is shipping more UNIX servers than anybody else. But most of those are entry level servers, and SUN is in free fall with regards to revenue. In Q3 2009 SUN's server revenue from servers running Solaris fell an amazing 39%, according to IDC, in the quater before Q2 that it was 40% and in Q1 it was 27%. You have HP and IBM slaughtering SUN in the markedplace, and it's going to continue. Why ? Cause SUN has cut so deep in R&D that it will take a long time for them to recover. And it's not like Fujitsu is stepping up.

http://regmedia.co.uk/2009/09/11/sun_sparc_roadmap.jpg

Do you really think that a Chip with just more threads now at 128 on a chip, at the same GHz, produced by a company that has never before produced a serious server chip (TSMC) will be enough to stop POWER7 and an enhanced Nehalem?

"I dont see the point in lying? I rely on official verifiable benches, you IBMers FUD and lie and twist everything. We discuss which CPU is fastest, you try to move focus away, to core pricing. Jesus. You always write the SAME thing: "work at a large bank/telecom company/etc which are loving SPARC and now migrating to POWER". No phantasy at all. The same lie all the time."

Ehh.. You don't rely on benchmarks. You haven't quoted one single official benchmark, you have quoted marketing sites, forums where people positive to your case have put their interpertations on the benchmarks. That is not the same thing. Again this might work in your little local backyard newsites or Boards, but not here.

"Why would I be embarrased because Oracle has the TPC-C record right now (via SUN gear)?"

Jup, and badly scaling cluster where they have had to only lease you the software to keep prices down, and only give you websupport. Again if you don't examine what you get when you put down the big bucks, you get cheated.

But I forget you have trouble reading TPC-C disclosure repports. How long do you think it will take before IBM puts out a benchmark using DB2 purescale ? It will blow the Snoracle benchmark out of the sky.

You, really should stop trying to play with the big boys, when you get easily bruised. And damn it takes 2 posts to adres all your SUN marketing FUD.

// Jesper

0
0
Happy

Hey Jesper

slow day at work? ;)

0
0

It's got all the right ingredients.

This conversation reminds me of racing golf carts around in first gear until their engines explode.

0
0
Silver badge
WTF?

unix vendor sales scum

>asdf

"...What is the point of buying a modern SUN machine when in a years time they will only sell x86_64 at a retarded high margin?..."

Are you joking? Why do you say that SPARC has a high margin, but dont even mention POWER? That is ridicilous as the POWER cost 5-10x times more than SPARC, for less performance!

Of course all the non Intel Unix vendor chips are sold at a retarded margin and is why they are slowly disappearing one by one (next up is SPARC to be followed by Initium, etc.) Power is only going to last longer because IBM hangs there hat on insisting on still being an IC maker and all the video game console contracts. My point was not to defend Power per say but to state the bleeding obvious that SPARC is dead man walking and in many ways this is a bad thing (Intel only one with money and volume to keep advancing, with each generation more expensive than the last).

0
0

Part 2

Part2

A sidenote: regarding the SPECjAppServer(R)2004 benchmark, you say

"...2) You keep mentioning two power 570'es. Actually it is one POWER 570 and one p570..." No I dont mention that. Read my post again. But slowly.

To answer on your part 2:

Why am I suddenly choose a x86 server? Why not the T5440? It is because I want to prove that POWER6 has problem keeping up, even with a x86. As I have proved, POWER6 is utterly crushed in some workloads by the Niagara (multithreaded server workloads). Now I want to show that on POWER6's traditional strengths: few cores, single threaded work - it gets beaten there too (by x86). Ergo, on multithreaded work, POWER6 gets beaten by Niagara. On few threaded work (hardly representative for a big server) it gets beaten by x86.

So, POWER6 gets squeezed by Niagara and x86. And people here say that you should not buy SPARC because it is 2-3x more expensive than x86, but they fail to mention that POWER is 5-10x more expensive than SPARC for less performance! So the question is: why buy a slow POWER6? Sure, it is fast on the things it is designed for (small workload that fits into the cache), but is sucks on everything else. Achieves far from it's specs. Just like the POWER7 will do.

Then you ramble about pricing on Oracle on the SAP, etc. Let me say this again, my dear thick headed friend: We are discussing which CPU is fastest. Ok? Not which core is fastest.

Actually, I read about some old CPU having extremely fast ALU. Therefore that CPU is faster than POWER6, yes? Because it had a better ALU. Is this true? No. Why do you keep rambling about one POWER6 core being faster? Sure, it is. But we are talking about the whole CPU, not parts of it, right? Thick. Headed. Slow. To. Understand. (Heh, and you talk about you educating ME. It should be I that educates you)

.

Kebabbert: "So how about the POWER7? Will it also suck? It has great specs, yes. Maybe it will even be faster than Nehalem on some things!"

Ignorant person: "...See the cache calculations above. 'nuff said..."

You have heard this before, when you were a kid in school: your logic is flawed. We all know that Niagara does NOT need a big cache to keep up and surpass faster CPUs. It would be impossible for Niagara to keep up and surpass a 5GHz CPU, if it suffered from small cache. But we see in benches that Niagara surpasses 5GHz CPUs. So, your cache calculations prove NOTHING. Niagara has a radical new design which makes cache misses unimportant. As wee see from the evidence (benches).

.

When you talk about scaling on SPEC OMPL2001, maybe you dont know that POWER6 scales extremely bad, and achieves far from it's specifications?

.

"...The T2 design is good, but not ground breaking. And it still has less cache, less memory bandwidth per 'click' than POWER6..." Let me ask you: can the Niagara keep up and surpass the POWER6 on certain workloads? If yes, how can that be if Niagara suffers from small cache and RAM bandwidth? Maybe it does not suffer from small cache and RAM bandwidth? Look at the proof for Gods sake! Do you expect a Pentium3 at 1.4GHz to keep up and surpass a 5GHz POWER6??? No. So riddle me this: how can a 1.4GHz Niagara surpass a 5GHz POWER6 on some workloads, if Niagara suffers??? Give me a good answer. Dont talk about pricing here, please? Nor your moms car. Just answer me that! AND DONT STEER AWAY FROM THE QUESTION! Jesus. Worse than a politician.

.

Regarding the CELL, yes it is cherrypicked benchmark by me, and I read some marketing material on this. You obviously dont know that IBM did that benchmark, in an attempt to show how fast the CELL is! I did not force IBM to do that bench. IBM chose it. Then SUN tried the same bench on Niagara, and on big workloads (where things do not fit into cache) it turns out the Niagara is 13x faster than CELL. Success! Of course SUN posts that bench! And yes, if Niagara were slower than CELL, then SUN would not have posted that bench. That is true. But I expect IBM to act exactly the same. But I dont expect Niagara to be slower on large workloads, because that is where the Niagara shines. On big workloads. On small, single threaded work it sucks. Everyone knows that. But on big workloads... Man, it crushes. Every bench where Niagara is fastest, is on LARGE work loads. When the workload fits into a cache, the Niagara always looses to POWER6, to x86, etc. How can Niagara only win on large workloads if it is suffers from small cache? Maybe you are wrong, maybe it does not suffers.

.

No, I dont have IBM power 570s at my work. Which happens to be at a big large finance firm. Actually. It IS true. You dont have to believe me on this. But they pick only the best and most well educated. If my CV is true, is is highly likely I work at something prestigious, yes? Maybe at a large world famous financial institution? And we dont have any IBM POWER machines. We have Solaris / Linux / OpenVMS. One of our well known, world famous large finance systems, will migrate from Solaris to Linux, now. Man, it sucks. Solaris offers more performance on same hardware, than Linux. Traditionally, SUN has been targeting Finance and Telecom. There are lots of Solaris there.

.

About SUN is shipping more UNIX servers than anybody else. Yes, the SUN user base is vast. Roughly 20 million Solaris licenses have been shipped today, world wide. Only Solaris 10 has shipped 11 million licenses:

http://www.sun.com/software/solaris/success_stories.jsp

Also, OpenSolaris has been downloaded several million times. Old Solaris licenses are several million. I wonder if AIX licenses even have been shipped into the 100.000 yet? They are so expensive, and run only on POWER. Solaris runs on many different CPUs, and is open and free.

1
1

asdf

Ok, then I understand you. You mean that POWER/SPARC/Itanium/etc are too expensive compared to x86 and they will slowly die? Maybe you are correct. But remember that x86 is really buggy and it is hard to get really reliable computers on x86. Lots of people dont let x86 into their data centers. If you really want reliable high up time, you can not use x86. That is a fact.

I agree that x86 offers most bang for the money. But, it is buggy and not reliable. So if you have really critical activities, you must use non x86.

It is like comparing a consumer grade harddrive costing $100 for 1000GB vs a Server harddisk 146GB costing $1000 - the server version will not die off to the consumer grade, despite having higher price. Quality costs.

(Unless you use ZFS and cheap discs, which actually are extremely reliable because ZFS protects against bit rot and silent corruption, because of end-to-end checksums. Ordinary checksums (hardware raid) does offer the same protection)

1
1
Silver badge
FAIL

RE: Kebabfart

Blimey! So many delusions in so many rambling posts!

Lets start with mainframes. Right off, let me state I am not a big IBM mainframes fan (though I do have a grudging respect for them), in fact I have worked at three companies where I spent a lot of time trying to get the management to ditch their mainframes for UNIX boxes. In each case there were UNIX versions of the software that was running on the mainframes, a migration would not have been too troublesome, and would have saved the companies money in the long term. I only succeeded at one. Mainframe buyers realise they pay a premium for their mainframes but they like the sense of security they get from their stability. Whilst you can argue all you like (as I did) that their sense of stability was largely illusionary, it is very hard to overcome that perception that a mainframe is a safe bet. You may compare this to the way some Slowaris shops refuse to believe anything but Slowaris will do what they want, despite the countless examples disproving their beliefs. That's if you get the chance to experience working in a Slowaris shop before Slowaris goes the way of the dodo.

Regarding x86 being "buggy", where do you get these hilarious ideas? You may have meant software that runs on x86, but x86 hardware such as Xeon or Opteron is not "buggy" at all! Vendors such as Unisys have been proving for years that you can make a very stable and reliable 24x7x365 platform with x86, basically by imitating mainframes. With such products as VMware's VMotion and other clustering tools , I can even make bog-standard M$ Windows Server apps on proprietary x86/64 servers that have better uptime than clustered SPARC Slowaris boxes, and at a much lower price! Your lack of experience, or simply your lack of willingness to consider other technologies, renders you unable to comtemplate the options available on the market.

/SP&L.

0
1

Part 1

Jesus, what happened with my part 1? Anyway, a quick recap.

Jesper Frimann,

Your long list of differences between SUN T5220 and IBM POWER on SPECjAPPserver bench is not valid. For instance, 1GBit NIC will suffice (you can drive a whole big exchange, all order traffic on 1GBit NIC!). You dont need 10GBit NIC. Hardly a concern. If IBM bench team saw that another 16GB RAM would have increased performance, then IBM would have added it. SUN X4640 has the SAP record right now. HP has less RAM, that is because then HP can use faster memory sticks than SUN.

Regarding your rambling about core licensing. Look man, I dont talk about price. I talk about performance. I couldnt care less about which core is fastest, nor the price of the cores. I talk about which CPU is fastest. You try to move focus to core pricing, not Performance of CPU vs CPU. This is getting tiresome. Why do you always focus on core, is it becuase you know that POWER6 will be crushed if we talk about CPU vs CPU?

It is like TPC-C bench. We see that four POWER6 is needed to match two Nehalem CPUs, and what is your conclusion? THAT POWER6 IS FASTER, BECUASE IT HAS FASTER CORES!!! You start to ramble about how many cores, etc. But it is not relevant how many cores when we talk about performance of CPU vs CPU. If we talked about performance of cores, then yes your objection is relevant. But now we are talking about CPU vs CPU! So answer me this; how many POWER6 is needed to match two Intel Nehalem on TPC-C? The answer is four. So how can POWER6 be faster than a Nehalem? Your logic is a bit strange?

The reason I talk about AMD opterons and Intel now, is to show how slow the POWER6 is. It can hardly keep up with x86. Nor Niagara. But it costs shitload more.

When we talk about pricing of IBM P595, you just start to rample here. It is not a fact that the P595 is extremely expensive? And the POWER6 has great specs, but in real life benches, we see that it can not catch up with Nehalem nor AMD Opterons, but it cost much more! So I expect the P595 with 32 POWER6s be surpassed by 32 Nehalem, for a small fraction of the price. And 32 Nehalems does not cost $35 million! That is just... outrageous. And all your ramblings about how fast the POWER6 is, sure, run some real life benches and see how bad it does fare against Intel Nehalem/Opterons/Niagara.

This is just a quick recap. I will wait for your answer on this question: "if you need four POWER6 to match two intel Nehalem on TPC-C, how can the POWER6 CPU be faster??" Explain that to me! Then, we you can not wriggle anymore, you have to admit the POWER6 is slower than Nehalem, we can continue the discussion and see exactly what is that you dont agree on.

0
1
IT Angle

RE: TheRegister Players

Anonymous Coward posts, "David Halko - Sun Lover who relies on Sun systems to his living so there is no better system"

Why did I get dragged into this one! LOL! I didn't even post anything on this thread, until now!

For the Register record... I used to sell Intel PC's, but when I saw my first Sun Workstation while at a regional computer users group meeting, which happened to be meeting at a local college, I was struck by them. When I later played on a NeXT cube, I have to tell you, I was far more impressed.

I had personally enjoyed Sun from the days I originally programmed on a Sun 3/50 in the SunView environment. The bundling of a graphical user environment, graphical debugger, graphical calendaring, graphical email, etc. was extremely innovative. IBM & Microsoft didn't have anything in comparison until years later. My draw to Sun is due to their innovation, even if some of their ideas are not taken up by the market.

This does not mean that I don't appreciate other vendors, technologies, and systems architecture - I enjoy the innovations in POWER & Itanium... as well as outsider systems vendors like Apple (since their acquisition of NeXT.) SMP Intel platforms (running various operating systems) had offered me a very good home, to earn money over the years.

Some of the applications that I find my expertise in, however, do not run under POWER or Itanium... they seem to be less multi-purpose than the SPARC or Intel platforms - but that does not mean that I don't appreciate the other architectures for what they are.

The applications that I worked with today were introduced to be under Intel SMP platforms, and I ran them under Intel for many years. After we migrated to SPARC, only then did I realize that apps were more solid under SPARC and we could press the SPARC hardware & Solaris much deeper into high utilization than the Intel hardware (with another UNIX's) could traditionally handle. Under light loads, Intel was snappier, under heavy loads, SPARC keeps on chugging. (I suspect it has to do with the unique way SPARC register windows work and innovation at the kernel level of Solaris... other people suggested Intel application programmers are just sloppy, but I don't necessarily subscribe to that agenda.)

All vendors bring something new to the party, which we call the Information Technology Industry.

I am not so happy when any competitor in the industry loses a foothold - it is a detriment to the advancement of technology, as a whole. The loss of DEC, Compaq, SGI, etc. pained me greatly... no one single vendor (Intel, IBM, HP, etc.) has done for the industry what competition from smaller companies had offered over the decades.

0
0

Mattie, you are fooled! Wake up!

Mattie Pattie Laddie

"...So many delusions in so many rambling posts!..." Jesus. You still dont know much about computer architecture. You are simply buying everything you hear from the salespersons. Please, study some computer architecture before you post again? It makes it easier for both of us.

Regarding Mainframes. One Nehalem EX gives you 400MIPS via software emulation. One Mainframe z10 CPU gives you 437 MIPS in native code. Software emulation is a factor 5-10x slower. One large Mainframe with 64 cpus give you 28.000MIPS:

http://en.wikipedia.org/wiki/Hercules_emulator#Performance

You need less than 16 Nehalem-EX CPUs to match 64 Mainframe CPUs. How much does one fully equipped Mainframe with 64 cpus cost? $50 million? How much does two 8-way Nehalem-EX machines cost? Less than $25.000? You do the math. Am I wrong, or am I correct?

Regarding x86 being buggy or not. Of course they are buggy! The x86 architecture has over 1000 instructions today! That is just hilarious. On x86 you need more transistors to decode it's instructions than a whole UltraSPARC CPU needs. That is why one RISC with 10 million transistors is faster than a CISC with 50 million transistors (10 million transistors are needed just to figure out which x86 instruction you just read and where the next instructions starts, 20 million transistors are never used but needed for backwards compatibility. 20 million are used for cache, etc). To create a x86 you need to support over 1000 instructions today. Bloat is a BAD thing!

Please study the horrible mess that x86 is, on Anandtech, because you never studied comp sci or arch in univ earlier.

http://www.anandtech.com/weblog/showpost.aspx?i=661

You are just totally ignorant. Many sysadmins would never let x86 into their datacenters, becuase it is buggy. And, no, the POWER6 is not that fast, nor the Mainframes. Just study some of the links/benchmarks vs AMD/Intel/Niagara I've provided, and google a bit. WAKE UP, MATTIE! You are fooled!

.

How can a 1.4GHz Niagara keep up and surpass a 5GHz POWER6, if Niagara suffers from small cache? That would be impossible. Dont you understand that something is wrong? If IBM were true, then a 1.4GHz CPU which suffers from a small cache, could in no circumstance, in no scenario, never ever, keep up, nor surpass 5GHz with a huge cache. That would just never happen. Or IBM is wrong, then Niagara doesnt suffer from small cache, and can keep up and surpass 5GHz in some scenarios. Is IBM right or wrong? What does benches and real life say? What does the evidence say? Just think a little bit? CAN YOU ANSWER THIS QUESTION, MATTIE? Or do you refuse to answer to this question? Are you afraid you will understand you are tricked by IBM?

.

Jesper Frimann,

I know you are reading this. Can you answer to my question? If you need four 5GHz POWER6 to match two 2.93GHz Intel Nehalem on TPC-C, how can you justify your conclusion that the POWER6 is the faster CPU? If I can only choose one CPU, which should I choose then? Which is fastest? I do not talk about cores, nor pricing of cores. I talk about performance of the whole CPU.

You dont see that your logic is a bit strange? Your logic is IBM logic - a strange one. For instance, IBM claims that IBM still has the TPC-C record today, because IBM has faster cores! That is just hilarious! Oracle/Sun does NOT have the TPC-C record today. IBM has! *chuckles* IBM logic - weird one. Jesper, you use exactly the same logic.

http://www.c0t0d0s0.org/archives/6012-IBMs-reaction-to-SunOracle-TPC-C-announcement.html

.

POWER6 does not scale good:

"Perhaps the challenge gets a little more clear, when we stay in the same shop: Almost 4 years ago IBM published a result of the p5-595. This system yielded with 1.9 Ghz Power5 CPU 100,512.30 QphH. 4 years of development, 256 GByte more memory and 2.6 times the frequency gives you just roundabout 50% more performance. Nice, but not that impressive, especially given the effort put into this CPU in regard of cycles."

I really really expect POWER7 to scale not good, as it is a POWER6 derivative. Sure, the specs are great, but in real life, it will suck. It will hardly catch up on a 1.4 Niagara on certain work loads. It will scale bad, just like all legacy constructed CPUs, which rely on high Hz and big cache.

0
1
Go

Nice work, Kebabbert

We are most pleased with your services in the last years. For every comment you made here or other places on the Internet, we sold another server. But now that Sun is soon to be extinct your effort is not needed anymore.

The check is in the mail.

Best regards and good luck in the future

All your friends at IBM

1
0
Thumb Up

Re: Nice work

Couldn't agree more,

I can't think of another individual who has polarised my opinion of SUN so much, his sheer arrogance and ability to alienate the reader is beyond belief.

I now have an irrational dislike of SUN simply because of Kebabmuppet, which is a bit unfair on SUN but as the saying goes People buy from People so tough!

Adios Kebabburp May the Schwarz be with you

0
0

Anonymous Coward

You are welcome, it is entirely on me, and my pleasure. :o) It feels good that I can influence people investment decisions. :o) My effort is not needed anymore? You are trying me to stop telling about IBM tech shortcomings and stop me from quoting official benchmarks and comparisons, here and other places on the internet? Why is that, you dont really like when start dispelling the IBM marketing with hard facts and benches? You would rather me be quiet? Well, if you buy me a christmas dinner, then maybe I will talk well about IBM instead? :oP Or if Oracle buys me a christmas dinner. Maybe Oracle values my dispelling of the competitors FUD on internet, here and there?

.

You know, sooner or later people are bound to think for themselves and question the "facts" that IBM gives them. Yes, the POWER6 is soooo fast, yes? Heck it is! People have not studied computer architecture, had they would, they would understand that IBM marketing is just false. IBMs greatest threat is people knowledgable in computer architecture who understand details.

Another lesson: basic computer architecture facts:

To reach high Hz, you need a deep CPU pipeline. Just like the Pentium4, it had high Hz and deep pipeline. It reached somewhere around 3GHz. If you have deep pipeline, then cache misses are severe and punishes performance greatly. The pipeline stalls - have you ever heard that? The POWER6 is just like Pentium 4, depends on high Hz to get some performance. I wonder: would the Pentium 4 crush the POWER6, if you clocked the P4 to 5GHz?

And if you have high Hz, then the CPU speed compared to RAM is very big. RAM is slow, and CPU is fast. You need to avoid cache misses, otherwise the pipeline stalls and the CPU chokes. You need two things

1) Big cache

2) High memory bandwidth.

When the data is available in cache, you need to transfer it as fast as possible to the CPU, hence the need for bandwidth.

We have evidence that POWER6 chokes very often, because it never reaches it's theoretical max. In real life, we see that it is really slow, for instance on LinPack, TPC-C, SPECfp2006, etc etc etc It reaches less than 50% of it's peak performance, which is evidence of pipeline often stalls and CPU always chokes. It DOES happen all the time, just look at the benches! POWER6 is like a fast car motor that needs lots of gasoline, but never gets any. It chokes all the time. Except, when you can fit the work load into it's cache, then it never chokes and can run full speed and crushes. But fitting data into it's cache never occurs for server work loads, which are large and dont fit into any cache. I expect POWER7 to exhibit the same behavior.

.

OTOH the Niagara which has small cache, it achieves high performance and beats 5GHz CPUs sometimes. That would be impossible if Niagara was crippled by small cache and low memory bandwidth. Hence, it does not suffer. Small cache is not a problem, because of it's new radical design. How else can you explain that 1.4GHz can beat 5GHz? It is a new radical design, yes?

0
1

Adam 61

Oh, so you are implying that you never had a dislike for SUN earlier, nor work at a POWER6 based company like IBM since many years? So all you anti SUN guys here, are anti just because of one person trying to dispell IBMs FUD here? Wow, great! Ok, so if I stop post facts here like "Mainframe CPUs give 437MIPS, while a Nehalem gives 400MIPS under software emulation (which is a factor 5-10x times slower!)", or "you need four POWER6 at 5GHz to match two Nehalem 2.93GHz according to official TPC-C benchmarks", and similar stuff, then all you will suddenly be SUN lovers? Oh yeah? Wow cool!

.

Ok, I can state the same thing as you do: Only because of you Adam 61, I am trying to dispel IBM FUD and marketing on the internet. If you, Adam 61, had not posted IBM biased things here, I would never had been trying to dispel IBM FUD. It is your fault, Adam 61, that I post benchmarks and white paper on several sites now. You have made me do this!

(Does this sound probable? Do you really believe me? No? So, why should we believe you? You probably work at IBM since many years, nothing I can say or do will make you SUN supporter. Quite dumb of you trying to imply such stuff, it only makes you look like IBM supporters states lies, hence "be wary of IBMers statements". You discredit them, not very smart of you)

0
1

Kebbaburp

Please do give me an example as to where I have posted any IBM FUD ETC, fair play if I have but I don't recall any.

The only things I have ever commented on are you inability to have a debate without ranting at people. Rude and obnoxious are two t

Personally I don't have a dislike of SUN, but I do dislike the zealots who are out there (please take a bow) who won't accept that there are viable alternatives out there and think that the IT world revolves around SUN.

I don't care much for IBM either - they do use FUD and selective benchmarking (i.e. certain workloads etc) - sound familiar - believe me dealing with them is a pain.

HP can be a pain as well by plodding on with old products that aren't as competitive as they should be in the UNIX and Storage space.

To me the vendors are like politicians lots of good stories but not always delivering on them

If you really want to get your message across in a constructive manner then stop being such a patronising arse

0
0

Adam 61

I am terribly sorry if you feel like this, but I suggest you just skip my posts.

I understand that you dont care about IBM lies, but I happen to care. I dont like lies nor FUD, such things must be dispelled with facts/benches/white papers/etc. Fair play is very important to me. I have been a Red Cross manager for a small group, for twelve years, on my spare time, as a volunteer.

There is another guy called Segedenum who also stated false claims about SUN, e.g. he said that ZFS requires several GB just to boot, etc etc. I showed links where FreeBSD developers admitted that there was a bug in the ZFS port so FreeBSD ZFS required much RAM. It took quite some postings before he realised he was wrong, but now he has stopped. I dont like when people are spewing out false things and FUDs, either deliberately or because of pure ignorance. Then I must tell them "no, you are wrong, it is like this instead: here are links and facts and benches"

Of course everybody has the right to an opinion. Anyone can say he hates Solaris/Niagara/etc, that is totally ok with me. Segedenum sometimes spews out Solaris rantings and SUN supporters start a heated discussion, but I PROTECT HIM! I say "he has the right to have an opionion, let him be. As someone said: I may not agree with you, but I can give my life for your right to express your viewpoint". To this SUN supporters have objected, but I have defended him! Everyone has the right to an opinion!

The difference is when someone LIES. If someone says "I like burgers more than kebab/kabob/kabop/etc" - there is nothing I can say. It is his opinion, and it is a fair opinion. But if he says "I like burgers more than kebab because kebabs are poisonous and you die" - then I have to object and tell him: "dont say so, it is not correct here are links and facts:"

There is a difference between an opinion and facts: OPINIONS ARE NEVER WRONG. FACTS CAN BE WRONG.

Maybe this is what Mattie tried to do earlier? He read my discussion with Segedenum and saw that I always protect opinions, so maybe therefore Mattie wrote "in my opinion Niagara is slower than POWER6"? Well, that is not an opinion, that is a false fact.

If Mattie had written "I prefer highly clocked CPUs to low clocked CPUs, and that is why I like POWER6 best" I would have said nothing, because it is his opinion. But if he says "Niagara is crippled by small cache" - that is not an opinion, it is a false fact because it is not true that Niagara is crippled, as evidenced by benchmarks, and the misunderstanding/lie should be corrected. If Jesper Frimann says "you need four POWER6 to match two Nehalem in TPC-C, and therefore the POWER6 is faster" - that is clearly wrong and must be corrected. He could have said "I like POWER6 for TPC-C better than Nehalem" - to that I must be silent. It is a fair opinion.

.

The thing is, on other sites there are some Microsoft people, and they can be a pain. But when I show links, they stop. The difference here is that Mattie, Jesper Frimann, etc still continues with their FUD and lies. I tried to explain again and again with links, benches, white papers, they are wrong - Niagara is NOT crippled, POWER6 is not that fast as they claim, as Nehalem easily beats POWER6, etc. But I just dont reach them. I have never seen this behaviour earlier. Normal people, when they see facts, they just stop. But these guys continue, despite having seen facts, therefore they _deliberately_ lie and FUD. It is not ignorance anymore, it is lies. Ok, if they express their opinion, they I would say nothing. They can hate Niagara as much as they want - it is ok with me, but when they lie about facts I must correct them. Lying is not fair play?

You accept their FUD and lies, and dont care, you say nothing. But when I dispel FUD and lies with facts, you care. The signal you send out, is that everyone should start to lie and FUD instead of using facts! They continue to say lies, and each time I react and try to dispel the lies - and you say you are tired of my repeated dispelling?? Why dont you become tired of their repeated lies? But, no, you are silent and accept them. If you want to end this, say to the liars and FUDers to stop, and then I will stop.

It is like a bully hits foreigners every day, and you dont care. One day, a foreigner starts to hit back and then you suddenly react: "hey, why did you hit back? I hate foreigners, because of you! You always hit back, I am tired of your repeated fighting back! I am into fair play!"

And in the midst of all this dispelling of lies and FUD, you come in and say "Hey! What are you doing? Why are trying to dispel the FUD and lies? I hate SUN because of this" and at the same time preaches fair play! Instead, if you are really into fair play, you should react against FUD and lies, instead of calling me an "arse". Why dont you call the FUDers and Liars "arse"? All your talk about fair play echoes hollow to me.

0
2
FAIL

Kebbablurb

Simples - you are confrontational, arrogant and aggressive in your posts- no wit & no humour

OK Matt is a bit obnoxious but usually in a very obvious and humourous fashion.

I don't mind you sticking up for what you believe in - perfectly fine with me, its the way you do it - why don't you think about that.

I don't get annoyed with FUD because I see it for what it is, every vendor does it, its part of the game - water off a ducks back is what it is.

Would love to chat more but have work to do

TTFN

0
0
Pint

Just a moment.

Damn what a lot of posts.

David Halko.

I used to dream of a SPARC 3 WorkStation. I grew up on BSD 4.2/4.3 and later SunOS and HPUX. So I know the dreams. I just got converted from being a diehard SPARC/HP9000 man to like POWER servers in the early 90ties.

Kebabbrain:

I'll deal with you later. There are more important things in my world right now. His name is Peter he weights 3020 Gram and is 53 cm long and was born the nigt to Monday.

Matt Bryant.

Jup, no big fan of the Mainframe and that bunch, as a Hardcore Unix Fanatic. They have kind of always been the arrogant bigbrother, who got all the cool stuff, and all the funds. But I agree with you one have to respect the Mainframe. Specially if one works with AIX and POWER cause there is a lot of Mainframe goodies

But it is not always a good idea to go off old Mainframe or the like platforms. I have seen some hardcoded HP3000 to Unix projects that didn't really turn out the way they should.

// Jesper

0
0

Adam 61

I admit my posts are full of benches and facts about Computer architecture and logic and stuff. And these guys where first attacking my credibility by saying "I know nothing", etc, so I just did the same to them, as a reaction. I see that it is ok if they attack me, but when I defend myself with exactly the same method It is not ok? Fine. And you are into... what was it again? Fair play?

Clearly you allow them to go on with their FUD and lies, and you even think it is acceptable and funny to read. Did you know that there are people that gets offended, even to things you think it is funny. People react to different things. I react to FUD, you think it is fun and acceptable. And when I react strongly and try to stop FUD, with the same methods as them, you react strongly towards me. Fine. Some people would say that you show some clear IBM bias here. And at the same time you talk about "fair play"?

Look, I react strongly to FUD, you think FUD is acceptable. You react strongly to my defense. People react strongly to different things. You dont decide what people should react to, to Lies or to my defense. There are people teaming with me, and people teaming with you. Who is correct? Who has the "right" to be most upset? Those dispelling FUD and lies, or those that react when I use the same methods as the FUDers?

If you showed some balance, and also criticized FUDers and liars while criticizing me, I would have listened to you. But to me, your posts are one sided just like Matties. Why should I listen to Mattie? Why should I listen you, with a clear IBM bias? Why do you call me an "arse", while protecting FUDers and liars and talk about "fair play"?

.

Jesper Frimann

Congratulations! :o) Is it your first kid? I only have a cat. But I am happy for you! Cool, man! :o) And out of respect for your family, I will say that we can continue this discussion some other time, so I will stop adressing you in this thread. Come back when you have time to continue our discussion! I will be there.

I think you know very much about Enterprise computers, that is clear. We have to meet somewhere in the middle, in our discussions. But for now, take care of your family, and do it well, man! :o)

0
0
Pint

Re:kebabbert.

Thanx for kind words. Nice to see that there is a nice person behind the all the spewing of FUD. And it is a show of good character that you can let us see that side :)= And nothing wrong with having cats, we have two norwegian forrest cats. Pseudo Children :)=

But enough small talk.

TPC-C benchmark.

You do know that the TPC-C benchmark is divided into a clustered and a non clustered part right ? The 384 Core Clustered T5440 is the fastest in the clustered version. The 64 Core POWER 595 is the fastest in the non clustered version. It is so simple... Not that you or the IBM or Oracle marketing departments seem to know this. But the trick is to see past the marketing bullsh*t that all companies make. Ok, some worse than others. And Oracle is one of the really bad boys in that depardement :)=

Nehalem

Is a really really great CPU. It is fast on the core level and

Core versus CPU versus CHIP versus Socket.

Now this whole area used to be pretty straight forward before SUN started mixing terms up. The old way was that a CPU equaled a Core and a CHIP could then contain several Cores and a socket could then have several CHIPS containing Cache and/or CPU's that might be sitting on a MCM (Multi Chip Module).

Now the general consensus today is:

Core which is also referred to as a CPU core.

CPU which is also referred to as a CHIP, which might contain multiple cores.

CHIP which is then a CPU and Perhaps even Cache.

Now I liked the older definition as it was more clear. But a lot of marketing FUD has unfortunately made the later the way things are perceived today.

Now what is important here.

1) Thread performance or rather throughput.

Thread performance is important cause a lot, let me correct that, most workloads depend on at least one thread running fairly fast.

2) Core performance or rather throughput.

This is important cause almost all commercial software is licensed on a per Core basis. This goes for Oracle and IBM. The two most important software vendors in the UNIX marked.

And with the price of software

3) Chip performance or rather throughput.

This is important cause the more you can push through a chip, the more you can push through a chip the less sockets you need. Less sockets means a simple machine which then in turn means a cheaper machine. Quite simple.

4) Other.

Things like virtualization capabilities, RAS features etc.

Now you have been talking about Nehalem, Niagara and POWER6. Lets have a look at them with regards to the above.

Single threaded throughput

Nehalem has an kick butt single threaded throughput, only matched by POWER6. It is hard to compare on SPECINT as all Nehalem uses auto parallelism, and no POWER6 does. Niagara simply has no submissions.

If we on the other hand look at per thread performance using SPECIntRate then POWER6 wins with a nosehair in front of Nehalem. And both Nehalem and POWER6 are around 12 times faster than the Niagara. (13 times on FP).

Single Core throughput

Again the picture for Nehalem and POWER6 is the same, but as Niagara uses 8 Threads per core then the values are a bit better. But both POWER6 and Nehalem are more than 3 times faster than Niagara.

Oracle is 0,75 licenses per core for Niagara T2

Oracle is 0,5 licenses per core for Nehalem

Oracle is 1.0 licenses per core for POWER6.

Here we can see that Nehalem clearly gives the most software bang for the buck, with POWER6 coming in at number 2 and being clearly the worst.

IBM licenses is 50 PVU for Niagara T2

IBM licenses is 70 PVU for Nehalem

IBM licenses is 80/120 PVU for POWER6

Here we can see that Nehalem and POWER6 are evenly matched in the lowend with POWER6 being more expensive in the higher end. Niagara is still the worst bang for the buck, but it is not as bad as it is with Oracle Licenses.

Chip Throughput.

Here the clear winner is Nehalem. On specintrate it does no less than 125 per chip, Niagara does around 90 and POWER6 does around 66. POWER6 is the oldest of the chips from 2007 and only holds 2 cores, hence it is no match for the 4core/chip Nehalem from 2009 and the x4 number of cores on the Niagara T2.

But there is a but here, Nehalem still only scales to 2 sockets, and if we have to go beyond 2 sockets then the picture is very different, and the T5440 only scales to 4 sockets, where as POWER6 goes all the way to 32 sockets.

And POWER7 will with it's 8 cores per CHIP, blow Niagara out of the water. And in the higher socket count I think it will also outclass Nehalem.

Now that was a pretty fair walk through of the +/- es of the chips.

// Jesper

0
0

Jesper

I love norwegian forest cats! I was strongly considering getting one or a Main coone. But now I have two home less cats. One I caught myself as a kitten and raised him. The other, I bought from a cat house for home less cats. :o) I would like to have children, later. It seems fun and way cool. I am happy for you. :o)

Regarding my FUD, you do know that I dont consider it as FUD? I link to benches and white papers, do they lie??? If you really want me to stop this "FUDing" as you call it, you have to disprove all benches and white papers and prove why they are wrong. Then I stop linking to them and I dont like to state my opinions (as they are worthless in a discussion of performance). Then I stop post at all.

.

Regarding TPC-C

Yes, I know that Sun uses a clustered solution. Everyone knows that. So what? How can IBM claim they STILL have the world record (because they have faster cores)? Look at the TPC-C list. Who is at nr 1? Sun or IBM?

If IBM wants the world record, well then they can cluster P595s and grab the record again. It's no big deal for me who has the record. The thing is, I dont like when IBM claims they still have the record. This is clearly false and a lie. Then I react. Fair play is important for me. The point is, I dont care who has the TPC-C record, I care who false plays and lies.

Here is an example where I criticize Sun, for comparing an outdated HP-UX machine to a new Sun machine. That is not fair, and only something I would expect from IBM. Read the comments. If Sun FUDs and lies, I must react. Fair play is important to me. If you see that Sun FUDs and lies, please show me links. I want to know.

http://www.c0t0d0s0.org/index.php?url=archives/6318-Peoplesoft-NA-Payroll-240k-benchmarks-with-16-streams.html&serendipity[csuccess]=true#feedback

.

Regarding Niagara vs Nehalem vs POWER6

I agree it is not fair to compare old POWER6s vs new Nehalem. I am only trying to show that the mighty superior POWER6 gets beaten by a modern x86. In other words, the POWER6 is maybe not that fast if it can be compared to a Nehalem at 2.93GHz?

Also, I dont understand

"...If we on the other hand look at per thread performance using SPECIntRate then POWER6 wins with a nosehair in front of Nehalem. And both Nehalem and POWER6 are around 12 times faster than the Niagara. (13 times on FP)..."

Later you write:

"...On specintrate [POWER6] does no less than 125 per chip, Niagara does around 90 and POWER6 does around 66..."

It seems you claim that Niagara is slowest on SPECintRate, and then you say that Niagara does 90 SPECintRate which is higher than POWER6 score of 66?

Anyway, I still dont like the talk about "POWER6 is the fastest CPU because it has faster cores". That is just strange. If you need four POWER6 to match two Nehalem, how can POWER6 be fastest? You also elegantly proved that Nehalem is faster than POWER6. And then you start to talk about pricing. Do we talk about performance, or pricing? If we talk about pricing, then I have to study that and return later into that discussion. I am talking about performance, not pricing. It may be true that POWER6 gives most bang for the buck, I havent studied pricing at all. I dont know. But I talk about performance.

.

Regarding POWER7, yes I really hope that it is faster than Niagara on every benchmark (including Niagaras strong points: multi threaded server work loads). If not, I will have much to say. I also want to see real life benchmarks, as the POWER6 scores far lower from it's specifications. If the POWER7 behaves identically bad, I will have much to say.

0
0
This topic is closed for new posts.

Forums