Feeds

back to article IBM touts Power Systems prowess on SAP tests

Server maker IBM is touting its ability to run ERP software from SAP as a means to promote its Power Systems servers running AIX, and it is using an SAP Sales and Distribution (SD) benchmark test to try to prove the value of its Power-based machinery. For whatever reason - and we all have our guesses on this - IBM has not seen …

COMMENTS

This topic is closed for new posts.
Anonymous Coward

per core pricing & SAP

Interetsing that IBM (or TPM) would seek to talk about "per core performance" off the back of a SAP benchmark, when SAP is one of the software products out there that isn't licensed on cores, sockets, server tiers or anything...

SAP is generally licensed "per user", so all this per core calculations are irrelevant.

and the whole "CPU performance" conversation is just so meaningless these days - show me a system that gets constrained by CPU... most apps these days tend to be IO or memory constrained...

0
0
FAIL

Sun have been at it too

Sun have been benchmarking too, boasting about how they beat IBM:-http://blogs.sun.com/BestPerf/entry/sun_t5440_world_record_sap

Mind you the most interesting that about that is the HP DL585 that came

second. A £15k DL585 was only slightly behind an £85k T5440. No idea

how much power6 machines are. But why bother with RISC at all any more

at those price/performance ratios?

0
0
Happy

Compare Power550 against Sun T5440 Midrange Peer!

Timothy Prickett Morgan writes, "The full list of SAP SD benchmark tests can be found here, and for the sake of this discussion, I am just going to compare the Power 550 to its midrange peers."

I am surprised the Sun quad-socket T2 processor was not compared in the article!

http://www.sun.com/aboutsun/pr/2009-07/sunflash.20090721.2.xml

"Running the SAP solution with Oracle Database, the SPARC Enterprise T5440 server surpassed the competing IBM System 550 server with four POWER6 processors by 26 percent with 3,752 SAP SD Benchmark users(1). This result also highlights the optimal performance of SAP ERP on the Solaris OS and the seamless multilingual support available for SPARC Enterprise systems running SAP applications."

Taking a closer look at the benchmarks

http://www.sap.com/solutions/benchmark/sd2tier.epx?num=100

IBM Quad Socket Power 6

6/16/2009 IBM 3752 0.97 1231000 20520 410330 AIX 6.1 DB2 9.5

SUN Quad Socket T2+

7/21/2009 Sun 4720 0.97 1550000 25830 516670 Solaris 10 Oracle 10g

Comparing Sun/Oracle SPARC/Solaris/Oracle10g against IBM Power/AIX/DB2, it seems like the OpenSPARC T2+ from Sun cleaned Power6's clock:

- OpenSPARC supported ~28% more users over Power6

- OpenSPARC supported ~26% more dialog boxes over Power6

- OpenSPARC supported ~26% higher SAP's over Power6

- OpenSPARC supported ~26% more line items processed per hour over Power6

Why was the quad socket OpenSPARC T2+ not considered a Power550 Midrange Peer, with the same number of sockets and higher performance on the same benchmark?

0
0
Happy

Anon: But why bother with RISC at all any more...?

Anonymous posts, "But why bother with RISC at all any more at those price/performance ratios?"

I think this is a very good question.

Usually, smaller chip vendors will do performance hopping, with the RISC vendors releasing a higher priced product with more dramatic performance increases than the more granular increases seen more often by the larger volume traditional vendors.

If vendors like IBM sees a good speed bump out of POWER7 released soon and SUN seesa good speed bump out of T3 released soon - your question will be answered with a completely different answer than what some people suggest today.

Let's see what the competing chip vendors offer in the next 1-2 quarters!

0
0

RE:Compare Power550 against Sun T5440

Well sure they are comparable, sort of.

I mean you would need a 42 inch monitor and a real small font to just to be able to see all the threads in top SMP view, on the T5440. I mean 256 Threads versus 16 for the power 550. That is a huge difference, and it's not a coincidence that the benchmarked T5440 has 256GB of RAM, all those threads need huge amounts of memory, to be able to do some work.

What I don't like about the T5XXX it is that you get the complexity of a high end box, with all those threads, without actually getting any of the benefits.

// Jesper

0
0

IBM: Who buys one core??

IBM knows that Power6 CPU is inferior to the SUN Niagara CPU by far. Niagara has many cores at 1.4GHz. IBM has few cores at 5GHz. IBM has faster cores, yes that is true. But does that really matter, when we discuss highest performing CPU? Maybe Niagara has faster floating point calculations (which is does) or maybe Niagara has faster ALU, or whatever. But to just focus on one part of the CPU is not really conclusive when we discuss the highest performing CPU?

When we discuss "fastest CPU", we have to focus on the CPU, not one core. But IBM is reluctant to do so, because Niagara CPU wins easily. That is the reason IBM tries to shift the discussion from "fastest CPU" to "fastest core". Which sane person would boast that Power6 is superior to Niagara by presenting an argument like this: "IBM Power6 has a faster core, ergo, the entire CPU is faster"? Some people would suspect this to be plain FUD and marketing?

To hide this fact, IBM always presents the number of cores in the benches. How many CPUs were used is hidden. To me, if one CPU with 16 cores wins over 8 dual core CPUs, which CPU is fastest? You clearly have to pit _one_ CPU against _another_ CPU. But you have to find out the nr of CPUs IBM uses, by yourself. If you do, you will find that Power CPUs are sloooow.

IBM: listen carefully, few fast cores are obsolete. Many cores is the new superior design. I know you IBMers have always mocked SUNs solution with many cores, but SUN's solution works better than yours, IBM. "Few fast cores are the way to go!!" - not quite.

The fun part is that for the price of one slow IBM Power server, you can buy several other higher performing servers. And these servers are not locked into proprietary AIX. Linux and Solaris are free. IBM can charge whatever they want for their solution. Before anyone consider buying a Power server, I would urge him to consider the competitors too and buy the solution with most bang for the buck.

PS. I think this article should also discuss Niagara CPU, because it discusses non Intel CPUs such as Power. Why not discuss non intel CPUs such as Niagara also? Niagara wins all these benchmarks easily.

0
0
Anonymous Coward

Re: IBM: Who buys one core??

Kebabbert, you might be right, T2 is very good at Java performance, but... nothing else. It's good for J2EE app server, but it's very poor (probably the worst) for database server.

Could you show those superior T2 performance results in OLTP processing (TPC-C)? Sun hasn't even published any. Are they ashamed? Even Oracle confirms that, since the processor factor licensing table shows Niagara is the slowest, then comes Intel and IBM are the fastest.

Single core performance does matter a lot as well. A lot of tasks are still single threaded and having dozens of threads will not speed them up at all.

> That is the reason IBM tries to shift the discussion from "fastest CPU" to "fastest core".

> IBM: listen carefully, few fast cores are obsolete.

Would you like to have on fast car or 16 slow ones? :-)

> IBM always presents the number of cores in the benches.

That's true - who cares how many sockets are inside the server? Admins? End users? Nope.

IBM has 100% right to count cores instead of sockets - they invented and were first to build dual core processor on single silicon chip.

0
0
Stop

FYI ... Sun is history

Why would you even want to buy a T2 or anything on Sun? Oracle will either sale the hardware business (if any takers) or just keep it around to migrate others to x86 boxes.

The T2 is a nice low end RISC box, but as a SUN customer (well we were until this year) we our planning a migration for our Solaris boxes to either Linux or AIX. I myself prefer going with Linux but upper management likes the idea of having a big throat to joke so looks like we will be going with IBM AIX p595 boxes and LPAR the many medium size applications.

0
0

Anonymous Coward

"Kebabbert, you might be right, T2 is very good at Java performance, but... nothing else."

Here we see some old benches where T2 outperformed Power6, on spec_int, ORACLE, SAP, Lotus Notes, etc.

http://johnjmclaughlin.blogspot.com/2007/10/utrasparc-t2-server-benchmark-results.html

-----------------------------

"It's good for J2EE app server, but it's very poor (probably the worst) for database server. Could you show those superior T2 performance results in OLTP processing (TPC-C)? Sun hasn't even published any. Are they ashamed?"

SUN has publicly stated that TPC-C is meaningless. It is favours pathological servers that no one use in real life. A while ago, IBM held the record. The server used many many Intel CPUs (not Power cpus) and 2TB RAM! And it used many short stroked hard drives! Who uses short stroked hard drives? How much does 2TB RAM cost? This config is really silly. And this IBM machine costed like 120 million USD, built for one purpose: TPC-C but nothing else. People shun away from TPC-C, they say "if you are good at TPC-C, then it proves that you are good at TPC-C and nothing else. Literally nothing else, certainly not average DB usage".

If you really think that a average DB admin has access to this machine, then you are wrong. TPC-C is really artificial. SUN could also build a server with many intel CPUs and compete in this artificial test, but why would they? It says nothing about real life applications.

Your point does not prove that Power6 is better than Niagara, which is not. This reminds of Top500 supercomputers, which are basically a bunch of nodes in a large network. On rank nr 5 we find the IBM Blue Gene super computer. It has.... dual core 700MHz PowerPC CPUs! Does a high ranking at Top500 mean that IBM makes better CPUs than SUN? Not quite.

------------------

"Even Oracle confirms that, since the processor factor licensing table shows Niagara is the slowest, then comes Intel and IBM are the fastest."

Come again? Do you mean that Niagara is slower than IBM? Your logic is really really unsound here. I can help you straighten up your logic: "Niagara is more expensive with Oracle, which Oracle factor licensing shows" but not "Niagara is slowest CPU, which Oracle factor licensing shows". It is helluva a difference between your statement and the correct statement.

---------------------

"Would you like to have on fast car or 16 slow ones? :-)"

It depends on the work load. If I have to transport 1000 persons from one place to another, then what would YOU choose? Niagara is like a huge slow bus, whereas legacy constructed CPU as Intel/Power6 are like a porsche. Imagine you have to transport 1000 persons. Which would be first to finish? Niagara or Power6? Niagara shines on multi threaded work under huge loads. It never chokes, which Power6 does.

-----------------------

"IBM has 100% right to count cores instead of sockets"

Yes of course. But if you talk about the highest performing CPU, the FASTEST CPU, the you should compare cpu vs cpu, not compare core vs core. Right?

------------------------------

"The T2 is a nice low end RISC box,"

Yes it is, and still it is faster than any other CPU. SUN has the best tech. :o)

---------------------------

Actually, I read lot of FUD here from e.g. Matt Bryant and Anonymous Coward. They lie, or are ignorant. Anyway, they FUD a lot. This makes me believe that all these stories about "we were a SUN shop, but now we switch to IBM Power" are lies. I bet all these stories are posted from a few people, with the same IP adress. If many people write so, then it shuns other people from considering SUN boxes. They think "many are switching from SUN to Power6, that must be for a reason, I should switch too".

Maybe all SUN lovers should write that to? Post as "Anonymous Coward" and write "We love IBM and Power6, but when we tried the new Niagara it smoked! And it is one fifth of the price. So we are now swapping our Power6 to Niagara. Sadly. I really liked Power6, but clearly Niagara is better and way cheaper. We just can not afford 413.000USD P570 vs a 76.000USD Niagara. And the Niagara smokes P570 too".

You know, for the price of one Power server, you get at least 4 or more SUN Niagara servers, which are higher performing too. In this financial crisis, who will buy Power servers that costs 413.000USD, and getting lower performance? Not many new investments are done.

Something is fishy.

--------------------

Here is a fun discussion with that Matt Bryant. I tell him several times that Niagara doesnt need large cache, because it has a new radical design that doesnt need large caches nor complex pre fetch logic, but still he refuses to understand. This radical design is the reason Niagara wins over legacy constructs as Power6 and Intel, in all benchmarks. I explain everything here:

http://www.theregister.co.uk/2009/07/22/sun_sparc_t_crank/comments/

0
0

SAP sizing

As somebody who regularly sizes systems for SAP and does performance tuning I can say that with SAP you need to balance number of CPU cores and CPU speed.

Lots of slow cores with multiple threads does not help that much as SAP is essencially single threaded (lots of individual processes but not multi-threaded).

CPU speed helps with single threaded performance. But not with throughput or number of users.

Be best balanced systems in my experience are highish core count with fast cores, i.e. x86-64 CPUs

0
0
Stop

Hey, Kebabbert, we were a SUN shop

Kebabbert,

Ok, we are still a SUN shop since we have many SUN boxes (M5000's, T5520's, v440's, v210's, and 4800's) still in our data centers. But when an older SUN box needs to get replaced we can not purchase a new T2+ or any other SUN server even though it would be fine. The reason is that upper management told us that they will not sign for any new hardware from Sun. So the T2 and T2+ could be a good fit for our older SUN equipment that we need to get replaced, but we have to look somewhere else and I'm sure we are not the only company.

Companies like mine have some major concerns about SUN hardware and for you to dismiss them proves that your head is in the sand or have too much of a bias opinion towards SUN. My company and others have two choices, wait to see what happens to SUN's hardware or replace older SUN equipment with HP or IBM. My company went with the second option of just going with IBM AIX (have it already in our shop as well) as a replacement (again upper management decision) for Solaris. So at the very least you have to agree that companies need to make a contingency plan since know one knows the fate of the SPARC chip. Which you have to agree hurts SUN since many companies will look at other options.

I'm sure you will not like this, but here is a good podcast about the roadmap for Sun customers after the Oracle acquisition by the Burton Group:

http://searchdatacenter.techtarget.com/tip/0,289483,sid80_gci1362140_mem

1,00.html?track=NL-832&ad=716328&asrc=EM_NLN_8764838

By the way, I posted anonymous NOT because I'm Matt Bryant (who just like you seems to have his own agenda) but because my fellow co-workers also read these articles.

0
0

Anonymous Coward

I understand that companies dont like uncertainty about SUN's hardware. My own company doesnt like SUN hardware, although it can make tremendous savings.

The only thing I am trying to say is this: SUN's tech is not as bad as Matt Bryant and other FUDers here say. In fact, it has several world records, for a fraction of the price. The sad thing is that marketing wins. And SUN always had lousy sales division. They couldnt sell water to one dying of thirst in the desert.

We see that Window wins big time over Unix and Linux. And that is not a testament that Windows is superior to Unix. Neither is it that Power6 is more expensive and slower than Niagara. Regardless what Matt Bryant and others here say. I just want to debunk their FUD and lies. That is all. I know SUN has lost the battle. But I dont like people lying.

Maybe Oracle can turn the tide. Then, maybe not. Anyway, SUNs Niagara CPU is a remarkable piece of technology. Says all those who knows a bit about CPU tech. Maybe that is why it sets world records all the time. There are many stories of inferior technology winning over superior technology. This is one of them.

0
0
Thumb Down

RE Kebabbert

>Here we see some old benches where T2 outperformed Power6, on spec_int, ORACLE, SAP, Lotus Notes, etc.

>

>http://johnjmclaughlin.blogspot.com/2007/10/utrasparc-t2-server-benchmark-results.html

Are you serious ? These are BMSeer quotes, do you seen any other people participating here using links HP or IBM marketing fud ?

These are chip versus chip comparisons. Yes they are perhaps interessting to you. But if you are talking Licinsing, setting

up real world applications or even comparing servers versus servers they are utterly meaningless.

Have you ever tried to setup a 4 threaded system versus a 64 threaded system to do the same workload ? It's a completely different

ballgame.

You just put yourself in the same category of people that you are pointing fingers at by using SUN's own FUD central.

>SUN has publicly stated that TPC-C is meaningless. It is favours pathological servers that no one use in real life. A while ago, IBM held the record.

>The server used many many Intel CPUs (not Power cpus) and 2TB RAM! And it used many short stroked hard drives! Who uses short stroked hard drives?

>How much does 2TB RAM cost? This config is really silly. And this IBM machine costed like 120 million USD, built for one purpose: TPC-C but nothing else.

>People shun away from TPC-C, they say "if you are good at TPC-C, then it proves that you are good at TPC-C and nothing else.

>Literally nothing else, certainly not average DB usage".

Who cares what SUN says. TPM-C have done more for Database performance, cause the hardware vendors actually learn something from this benchmarks, that goes

into their products, than SUN has done since they choose to stop making those benchmarks.

> If you really think that a average DB admin has access to this machine, then you are wrong. TPC-C is really artificial.

Sure TPC-C is artificial, it's a benchmark. But jup I have access to several, You name it POWER5 p595's with 2TB of RAM, SUPERDOMEs with 64xPA8900

or 128 way x I2 Montecitos, power 570 with 16 4.7GHz cores. And yes they are getting used some with quite large partitions. 56+ cores.

> SUN could also build a server with many intel CPUs and compete in this artificial test, but why would they? It says nothing about real life applications.

So why don't they ?

> Your point does not prove that Power6 is better than Niagara, which is not.

> This reminds of Top500 supercomputers, which are basically a bunch of nodes in a large network.

>On rank nr 5 we find the IBM Blue Gene super computer. It has.... dual core 700MHz PowerPC CPUs! Does a high ranking at Top500 mean that

> IBM makes better CPUs than SUN? Not quite.

You don't really get it do you ? You hail systems with lots of slow cores and lots of threads and then you scoff at Blue Gene, which kind of like uses

the same principle. I mean... you make less and less sence.

>"Even Oracle confirms that, since the processor factor licensing table shows Niagara is the slowest, then comes Intel and IBM are the fastest."

> Come again? Do you mean that Niagara is slower than IBM?

Sure it is. Don't confuse speed with throughput. A niagara Chip will deliver more throughput than a POWER6 chip. But on a core to core basis there

really isn't a match. On a throughput basis it's different there the 8 core 64 threaded niagara chip will outperform a POWER6 chip, if...

1) Your workload can be threaded to enough threads to exploid the 64 threads of the niagara. Not much good if you only have 4 threads.

2) All your bloody threads doesn't block eachother.

3) Your job isn't depending on serialization like the job I'm doing tuning on at the moment. 10 min with 1 thread, then 20 min with Many threads,

then 10 min with one thread and result. The job is heavily CPU bound.

Now who would run this job fastest ? This currently runs on a 3 Core POWER6 4.7Ghz Virtual machine.

Now on a niagara I would most likely get 80 min with 1 thread 14 min with many threads 80 min with 1 thread, oh yes sure I could start up a whole

development project to paralize the code, and pay more for my Oracle License. Welcome to the REAL world.

>It depends on the work load. If I have to transport 1000 persons from one place to another, then what would YOU choose?

>Niagara is like a huge slow bus, whereas legacy constructed CPU as Intel/Power6 are like a porsche.

>Imagine you have to transport 1000 persons. Which would be first to finish? Niagara or Power6?

Wrong again, Niagara is the 64 small slow minibusses, where you need 64 drivers, power6 are 4 fast Large busses using only 4 drivers. So sure

>Niagara shines on multi threaded work under huge loads. It never chokes, which Power6 does.

What are you talking about ? Try slamming your 64 minibusses and 4 Large busses onto a public road, and see who messes up. The sheer

complexity of having 64 threads versus 4 is exactly why the niagara chokes.

>"IBM has 100% right to count cores instead of sockets"

>

>Yes of course. But if you talk about the highest performing CPU, the FASTEST CPU, the you should compare cpu vs cpu, not compare core vs core. Right?

You seem to forget that before Intel and SUN started calling multi core, chips for CPU's, a CPU was equal to what we today call a core for.. well 60 years

or so. And it worked it confused people like you. So why don't we start to call a pig for chicken. It will do wonders for the PIG meat sales.

> Maybe all SUN lovers should write that to? Post as "Anonymous Coward" and write "We love IBM and Power6, but when we tried the new Niagara it smoked!

>And it is one fifth of the price. So we are now swapping our Power6 to Niagara. Sadly.

>I really liked Power6, but clearly Niagara is better and way cheaper. We just can not afford 413.000USD P570 vs a 76.000USD Niagara.

>And the Niagara smokes P570 too".

Well I've seen people throw away their new Niagara boxes, and replace them with partitions on old POWER5 partitions on p595's. Why ? It made sence from

a licing perspective, and running RAC on the Niagara boxes performed so badly that they actually tried to move it to the older v880 test machines to get

some throughput.

And if people are stupid enough to only focus on the cost of the hw box, then let them I don't care. When they mess up their IT infrastructure, then come

to us and we'll throw the lot onto a large POWER server and we'll give you more performance, better uptime at a lower cost.

// Jesper

0
0

Jesper

"Are you serious ? These are BMSeer quotes, do you seen any other people participating here using links HP or IBM marketing fud ?"

I have no problem with such links, as long as they are published white papers and I can look it up myself. In fact, I have studied articles from IBM, and found weird claims there: "One Power6 core is faster than one Niagara core, ergo the Power6 cpu is faster". Now that is clearly weird. Or, when IBM states that one Mainframe can consolidate 1.500 x86 servers, which assumes that all servers idle at a few percent, whereas the Mainframe is 100% loaded. That is also clearly weird. Then I can state that my laptop can consolidate 100 of the x86 servers (if they idle). But noone would believe that my laptop can do the job of 100 x86 servers.

If you wish, I can post white papers from other sites, e.g. Oracle's site. One SUN T5440 gets 14.000 SIEBEL benchmarks, vs three IBM power570 gets in total 7.000 SIEBEL.

----------------------------

"You just put yourself in the same category of people that you are pointing fingers at by using SUN's own FUD central."

You have understood nothing. Let me say this, It doesnt matter who says what, as long as it is true. If IBM says something, as long as it is true, I accept it. I dont care who says what. I look at the content. When you write research papers, it doesnt matter who says what, as long as it is true. You have understood nothing.

------------------------------

These are chip versus chip comparisons. Yes they are perhaps interessting to you. But if you are talking Licinsing,"

Who is talking about licensing? What I talk about, is which CPU is fastest. Niagara or Power6. I couldnt care less about which part of a CPU is faster than another small part on another CPU. I want to know, which CPU is fastest? Who wins all benchmarks?

-------------------

"If you really think that a average DB admin has access to this machine, then you are wrong."

"But jup I have access to several, You name it POWER5 p595's with 2TB of RAM, SUPERDOMEs with 64xPA8900 or 128 way x I2 Montecitos, power 570 with 16 4.7GHz cores. And yes they are getting used some with quite large partitions. 56+ cores."

Are you serious? Just because YOU have access to all these multi million USD machines, do you really think that the average DB admin has access to machines with 2TB RAM? Do you really believe that the average DB admin is affected by TPC-C results? Those machines does not exist out in the field. TPC-C machines are unnatural and pathological. That single IBM TPC-C machine I talked about, costed like 120 million USD! Look, I can promise you that the average DB admin has not access to 120.000.000 USD machines. You are wrong.

---------------------------

"You don't really get it do you ? You hail systems with lots of slow cores and lots of threads and then you scoff at Blue Gene, which kind of like uses the same principle. I mean... you make less and less sence."

Look, I am nocking Blue Gene which uses 700MHz PowerPC CPUs, not because the CPUs are slow. But to those that say "Niagara has no rankings at top500, hence it must be slow". I am trying to show that their claim is wrong, because having a ranking at top500 is no evidence that the CPU is superfast. I understand if you think that I make less sence, if you dont understand anything.

--------------------------------

"Sure it [Niagara is slower than Power6] is . Don't confuse speed with throughput."

You havent studied higher math, that is obvious. I suggest you study measure theory. When you measure something, you have to impose a metric. It preferably should be an Euclidean metric. But the point is, the metric is NOT defined here. Again, you have not understood anything. Geez.

------------------------------

"On a throughput basis it's different there the 8 core 64 threaded niagara chip will outperform a POWER6 chip, if...

1) Your workload can be threaded to enough threads to exploid the 64 threads of the niagara. Not much good if you only have 4 threads.

2) All your bloody threads doesn't block eachother.

3) Your job isn't depending on serialization like the job I'm doing tuning on at the moment. 10 min with 1 thread, then 20 min with Many threads,"

You know, for a server CPU, the workload are multi threaded. Client - Server, eh? One server serves many clients, eh? This translates easily to multi threads, where each client is one thread? You know, the server software are not typically P-complete, they tend to NC-complete because of the many threads. Didnt you know that?

----------------------

It feels like I could spend the rest of the day pointing out all errors or misunderstandings in the rest of your single post. Alas, I dont have neither time nor ork to do that, so I stop here. Why should I spend more of my precious time on some meaningless post where the author has misunderstood everything?

0
0
Silver badge
Happy

RE: Kebabbert & Novatose

Good to see the whole Sunset thing hasn't got old Novatose down, he's still furiously pluggin' away with his cherry-picked benchmarks. And now he has an ickle side-kick in wet-behind-the-ears Kebabbert, our junior league contender from some continental outback uni where they probably still wear lederhosen. Maybe the two should get together, they seem to have a lot in common, especially the ability to type with their heads firmly subterranean.

I especially enjoyed Kebabbert mentioning the other thread he has been amusing us all in, http://www.theregister.co.uk/2009/07/22/sun_sparc_t_crank/comments/, as that's also the thread where, after telling us all how good and smart he is, he has to admit he can't understand why his own bosses won't buy Sun! What he can't understand is that there is a lot that goes into the selection of a platform that gets to run an enterprise application, and "clever" CPU design is not very high on the agenda. Especially when that "clever" CPU design runs like a dog with the type of applications businesses actually want to use.

As for TPC-C, Sun were quite happy to brag about TPC benchmarks for years when they had a design that could compete, but as UltraSPARC got further and further behind Sun got quieter and quieter, and then they just switched to FUDing a benchmark they used to use as a sales tool. It is very true that TPC-C is no real guide as to how your application will run in your environment with your stack, but Sun were quite happy to sing its praises when they had a song to sing. But I think you'll find that I have repeatedly advised in my posts that you ignore any vendor bench as anything other than a rough guide and instead insist on having a shoot-out in your environment, with your data, your tools and your application stack, as it is the ONLY way to get a realistic view of how that platform will perform in real use. It's also a good way to screw a few more points off the price ;)

But I guess you'll just call that advice more "lying" - how funny! It shows a major lack of both maturity and just plain manners that you immediately accuse anyone that disagrees with you of being a "liar", I do suggest that you moderate your approach, especially as you may end up working with really stressed individuals like DBAs, and they may not take too kindly to your tone. Well, you may end up working with DBAs if you work with something other than Niagara!

And as for short-stroked disks, you'd be amazed at some of the tricks I've seen used to try and squeeze a few more transactions out of a system! I have seen an EMC disk array populated with the maximum number of disks and the maximum number of interface ports, all to try and get a 20GB database to run that little bit faster. I reckon they were using less than 2% of the available disk space, but because their business depended on making that database fly, they paid for the massively under-utilised storage, and over-spec'ed server cluster, and dedicated SAN and LAN links. Sometimes the business requirements really do mean you end up with what is technically a poorly-utilised design, but if it's what the business wants and will pay for then it will happen.

Anyway, isn't it about time we started referring to it as Oracle Niagara? Oh, hold on a sec - Larry still hasn't committed to actually making any more Niagara, so all the hot air about T3 is likely to remain just that, just as it did with Rock.

/SP&L <= Kebabbert may need one of the older Sunshiners to explain this for him.

0
1

Matt, Oh Matt...

Have you now finally understood why a Niagara CPU doesnt need a large cache?

A server CPU such as Niagara serves hundreds/thousands of clients. It is impossible to hold all that data in a cache. The cache gets emptied and refilled all the time. The Niagara has taken another route, as Ive told you many times. It doesnt fight cache misses, instead it has fast access to RAM and works with another thread while waiting for data.

Only a ignorant person would claim that "OK, again just for you - Niagara doesn't have enough cache." The worst thing is that Ive tried to explain this many times, but still you insist that Niagara doesnt have enough cache. This is ridiculous. If Niagara can not function properly because of a small cache and is crippled, why does it win all these benchmarks that Ive posted?

The Power6 can neither hold an entire server work load in it's cache. It gets emptied/refilled all the time. Hence the Power6 sucks badly as a server. It's more suited as a desktop PC, where it can hold all data in it's cache. Then I expect it to be fast. But as a server it is sloooow and bites the dust. Three P570 powerservers together, gets half the SIEBEL score as one SUN T5440. That is ridiculously bad. And one P570 costs 413.000 USD and one T5440 costs 76.000 USD.

You can say what you want, the fact is, Niagara wins all these benchmarks. Isnt it true? Yes or No. Can you answer my question? Another question; can you stop lie about Niagara being slow and having a too small cache, when it in fact wins? The cache is not too small, as Ive told you.

0
0
Happy

RE:Compare Power550 against Sun T5440 #

Jesper Frimann Posted, "I mean you would need a 42 inch monitor and a real small font to just to be able to see all the threads in top SMP view, on the T5440. I mean 256 Threads versus 16 for the power 550. That is a huge difference"

That is only if you are interested in watching threads spin - I would ask for a 56" flat panel, if I were you!!! ;-)

It is kind of like moving sand - if it costs less money to supervise 20 people with shovels than it does to rent & use a back-hoe... the choice may be made by economics.

Jesper Frimann Posted, "it's not a coincidence that the benchmarked T5440 has 256GB of RAM, all those threads need huge amounts of memory, to be able to do some work."

That is a very interesting observation! I wonder what the memory utilization was like between the platforms? I guess that begs another question - what is more expensive: more memory chips or different CPU architecture? There are advantages and disadvantages to every approach - licensing of the applications in the environment may also have something to do with the economics.

Jesper Frimann Posted, "What I don't like about the T5XXX it is that you get the complexity of a high end box, with all those threads, without actually getting any of the benefits."

Some advantage from a high-end box achieved in the T2 line include: high throughput, partitioning at the firmware level, partitioning within the OS level, good lights-out management.

Some advantages from a low end box achieved in the T2 line include: low power consumption, fewer overall components (to fail), less rack space consumed, lower HVAC costs, less expensive & widely available industry standard peripheral components.

The servers based on the T2 line of servers are very interesting, indeed. Not good for every application, but very good for quite a few.

0
0
Happy

Various anonymous posts

Anonymous posts, "IBM has 100% right to count cores instead of sockets"

Indeed - and application vendors who license may choose Multi-Chip Module, Core, or Socket. Those application vendors have the right to count however they wish and the customer who chooses a vendor (like IBM) may have to pay a higher premium, regardless of how IBM wants to count their performance benchmarks!

Of course, Sun has the same problem, which is how some application licensing is cheaper under OpenSPARC at 4 sockets than Power and other application licensing is cheaper under Power.

Anonymous posts, "when an older SUN box needs to get replaced we can not purchase a new T2+ or any other SUN server even though it would be fine."

Seems like someone is not making a reasonable business case. One just has to make the business case to the right person.

Anonymous posts, "Why would you even want to buy a T2 or anything on Sun? Oracle will either <incomplete speculation>... The T2 is a nice low end RISC box, but as a SUN customer (well we were until this year) we our planning a migration for our Solaris boxes..."

Some reasons to buy Sun or OpenSPARC T2 platforms: maintaining existing business processes, streamlining existing business processes, transparently consolidating existing platforms (Solaris 8 or Solaris 9 migration assistant onto Solaris Containers), gaining application throughput, gaining flexibility.

Wasting money on platform architecture migration due to speculation is not incredibly wise - especially when Oracle [in a statement released to the media by the CEO] claims they make most of their money under Solaris!

It makes no fiscal sense that Oracle would kill their cash cow (Sun Solaris) and even less sense that any management in any company would assume that to be the case.

0
0
Stop

Kebabbert

It is very seldom I say this about someone posting here.

But you have absolutely no clue about the things you are talking about.

And your troll techniques are kind of infantile.

// Jesper

0
0
Badgers

RE: Mr Doner

"It doesnt fight cache misses, instead it has fast access to RAM and works with another thread while waiting for data"

So if your app doesn't have lots of threads, then it does have to wait around on a Niagara CPU because fast access to RAM won't be quicker than Cache.

So I guess the question is how many threads would typically be used in the majority of mainstream applications, and does an ISV have to put in extra work (and how much effort is this) on an app that needs to be "optimised" to work with Niagara over and above Power or Itanium.

PS the comparing core to core performance is relevant for two reasons:

1. Lots of software is licensed per core

2. HP and (I think) IBM don't have to enable all the cores in a CPU,

0
0

Jesper Frimann

You havent understood anything which your remarks show. Begone yourself.

------------------------------------

Mr Doner,

"So if your app doesn't have lots of threads, then it does have to wait around on a Niagara CPU because fast access to RAM won't be quicker than Cache."

Yes that is true. And it is hard to parallellize an application.

Fortunately that is not needed, when talking client - server. A server serves many clients. That is many threads. Client - server software is already multi threaded. No need to rewrite server software. Hence, Niagara does fine for server work. Which benches show.

A desktop CPU such as Power6, can never hold all thousands of client's different data sets in a cache. A desktop CPU is very sensitive to cache misses and will suck badly at server usage. The speed of a desktop CPU comes from it's ability to hold all data in it's cache. If it can not do that, it can not work. Therefore it looses server benchmarks with many clients (aka threads).

A server CPU such as Niagara, is not sensitive to cache misses, thanks to it's new and radical approach. It doesnt try to fight cache misses, as Power6 does. Niagara works around cache misses by masking them. Therefore Niagara smokes desktop CPUs such as Power6.

0
0
Happy

RE: Mr Doner #

Adam 61 posts, "So if your app doesn't have lots of threads, then it does have to wait around on a Niagara CPU because fast access to RAM won't be quicker than Cache."

Yes, absolutely right. Not necessarily the right CPU for the job if the application or environment on which it sits is not well threaded.

Kebabbert posts, "Yes that is true. And it is hard to parallelize an application."

The concept that environments with common applications would not benefit from highly threaded hardware is really a myth propagated by DoS trained folks.

With DNS cache lookups, async I/O, file system syncs, multi-threaded NIC cards, VPN encryption, HTTPS encyption, compression file systems, web browsers, background processes, software update downloads, virus checking, signature checking downloads, de-duplication, backups, RSS feeds, internet radio, MP3's playing, etc. - the common user benefits tremendously as hardware become more highly threaded with a generally more responsive platform.

Even my Windows XP desktop has 74 processes running, never mind the thousands of threads!

Kebabbert posts, "A server CPU such as Niagara, is not sensitive to cache misses, thanks to it's new and radical approach."

Unfortunately, the approach is not new any longer. Sun has pioneered it, proved it, and the rest of the market has confirmed Sun's approach by scraping their old road maps and emulating Sun CoolThreads (although the market is still years behind, or less if it is a cost/performance ratio.)

I hope for Oracle that Sun has some other major development breakthrough up it's sleeve, that the rest of the market will have to backwards-engineer.

Proximity computing has extremely high potential, as a new breakthrough technology from Sun - the power savings from not having to have chips actually get soldered to a board or seated in a socket is tremendous, never mind the benefits of not having to manufacture multi-chip modules & the loss that is incurred in the manufacturing process to both chips when something goes awry.

0
0

Correction

That IBM tpc-c machine I talked about, was the "IBM Power 595 Server Model 9119-FHA" which uses 128 of the intel Xeon 2GHz CPUs. It did not cost 120 million USD. I accidently added a zero. (Because in my country I have to add/remove a zero when transforming prices back and forth to USD). It would be more correct to say that machine costed 12 million USD. But the correct price on that machine is 17.1 million USD.

http://www.tpc.org/tpcc/results/tpcc_result_detail.asp?id=108061001

Anyway, no DB admin has access to such a pathological 17.1 million USD machine. It is similar for a company to manufacture a car (with a rocket engine and wings etc) that breaks the speed of sound and stating that the company's normal cars are fastest on the market. There is no correlation between pathological machines and the normal every day machines.

If IBM has ranking nr 5 on top500 with 700MHz PowerPC cpus, it doesnt prove anything.

0
0
Badgers

Hang on a second

Now i've had a look at the SAP SD Benchmarks and whats very interesting to note is that the Proliant config delivered pretty much the same level of performance but with 1/4 of the memory of the SUN system.

So is the real winner here actually AMD not SUN or IBM??

T5440 = 4770 Users & 25830 SAP's

vs

DL585 = 4665 Users & 25830 SAP's

So does this highlight a weakness in the Niagara architecture vs AMD in that it requires significantly more memory to achieve similar performance to a Proliant with AMD. It would be interesting to see how the T5440 would bench with just 64GB of RAM.

Now i'm thinking the Proliant will have a significant price/performance advantage here and probably a lower cost of ownership.

0
0
Badgers

Hang on a second

Sorry the Proliant benched at 25,530 SAPs - 300 less han the SUN system - a typo on my part

Still the arguement stacks up

0
0

Thread vs Process

The problem with SAP and other ERP platforms is that in the main they are multi process and not multithreaded. Hence they cannot take much advantage of extra threads available but can take advantage of multiple cores.

Where threading does have an advantage is at the DB level, i.e. oracle 10 is multithreaded. In the past where processors where not multithreaded (but the database was) a process with two threads would take up two cores (if running full tilt) this would not have been available to the application. Now with Niagara/Nahalem it would use two threads on a single core thus leaving the other cores to the application.

However with the Niagara systems (and I know from personal experience, having be forced to migrate from 3Ghz windows machines to Niagara Solaris machnies) single threaded performance is terrible, the users where not happy. Most ERP systems usually run heavy batch load over night, e.g. extracts into BI systems, large interfaces, MRP runs, etc. If these have not been designed to work in parallel performance on slow cores suffers dramatically.

Unfortunatly SAP does not have a Java benchmark suite (e.g. for SAP portal, XI). Here Niagara does give you a massive benefit as Java is inherintly multithreaded and for web based applications single threaded performance is not so much of an issue as network and browser render speeds form the greater portion of the average response time.

0
0

Adam 61

Quite possible. I dont say that Niagara is best. I only say that it is not as slow as liar Matt Bryant claims. And I also say that Niagara is faster than Power6 on some benches, contrary to what FUDer Matt Bryant claims.

0
0
Big Brother

Re:Thread vs Process

Fenton, I think you might have mixed up the threads and processes of the Operating system with the concept of processor core threads.

On for example a niagara based processor each 'processor core thread' is actually capable of running a whole separate OS image, which then again can run many processes and threads.

And I have seen the same problems when migrating to Niagara based servers as you describe.

And don't get me wrong I think Niagara is great, for what it was designed for, workloads that will perform well with many light independent threads. But there is a long way from that to a general purpose Business platform.

// Jesper

Big Brother cause:

1984 was meant as a warning, not a manual.

0
0
This topic is closed for new posts.