Feeds

back to article Sun's UltraSPARC T2 goes flame-retardant

Congratulations go out to Sun Microsystems for doing its best to make a new NEBS box seem exciting. Sun will ship you a NEBS server that supports 64 software threads, don't ya know. Telcos will now find the Sun Netra T5220 server. This system run on Sun's UltraSPARC T2 processor, which has 8 cores and supports 8 threads per core …

COMMENTS

This topic is closed for new posts.
Thumb Down

We already tested the T2 and not impressed

The NEBS applications require heavier threads than the T2 can handle. It's not so much the "castrated" core than it is the dismal amount of cache / core on the chip.

0
0
Silver badge
Pirate

Fight! Fight! Fight!

Hmmm, I thought even Sun admitted that eight cores vertically time-sharing eight threads actually meant eight cores with only one concurrent thread each (and up to seven other threads per chip stalled waiting for data from cache, memory, disk, SAN....)?

/countdown to apoplectic Sunshiners responses started!

0
1
Linux

UltraSPARC T2 info

Just FYI: The T2 has two pipelines per core so it runs two concurrent threads per core with six waiting. You can read all about it at http://www.opensparc.net/

0
0
Coat

Netra T2 is nice step forward

For poster# 1, which applications have you tested that you say that the cores can't handle ? We have a number of applications that we tested on the T1, and our primary limitation was amount of memory, not the processing power. So far the apps we tested on T2 looks good because it doubles the memory capacity/bandwidth and processing and we didn't have to tune the code to minimize the little bit floating point code in our apps.

Matt Bryant, I guess you need to educate yourself little bit before speaking. The T2 has two integer pipelines in each core and runs two concurrent instruction streams. In any case, our apps can only use 4 to 6 threads at the most, we plan to consolidate almost 10 to 15 applications into one chip fits in nicely with our architecture.

0
0
Alert

The good, bad and the ugly

The 20 inch depth is great for those of us that bought our racks more than a half decade ago and the build quality from sun's NEBS machines is what those of us who have been using their hardware for decades expect.

The T1 and T2 are great processors for running poorly optimised code that spends all its time chasing pointers which is typical in modern practice outsourced OOD. The T1 is not really good for much of anything else and the T2 is an ok solution for a larger solution set than the T1 and required for applications that use floating point or SSL crypto.

The real problem with the new beasts is they only runs toy operating systems such as Solaris 10 which seems to be still in beta based on what gets fixed with each patch set. The new system is excessively complex and its code looks like they let open source amateurs rummage through it which has resulted in such wonderful features as the remote telnetd exploit and having init(1M) linked to all sorts of libraries that are not needed but are locally exploitable. Add in their new way of doing things with smf and you end up with a machine that keeps updating critical binary files that you can't checkpoint or audit.

If you're looking at buying one of these beasts, take it for a test drive. You can do that by just looking at http://www.sun.com which was rumoured to be a pair of these. Do you want your corporate web site to be that slow?

0
0
Silver badge
Happy

RE: Netra T2 is nice step forward

"....The T2 has two integer pipelines in each core and runs two concurrent instruction streams. ..." Oh, apologies, sixteen possibly concurrent threads is so much closer to the advertised figure of sixty-four! Do you have any figures on the cache hit ratio for T2 as it was pretty low for T1 (well, compared to any Intel chip anyway)?

0
1
Unhappy

RE: Netra T2 is nice step forward

"sixteen possibly concurrent threads is so much closer to the advertised figure of sixty-four"

Matt, I guess you really could use some education regarding this. The T1/T2 lines are pure throughput processors, this is the only thing these are good at. They suck at pretty much anything that needs fast response time, but fortunately, they can handle very heavy load quite well. The T2 runs 16 concurrent threads out of 64 at any time - which isn't all that bad when you see that Montecito runs 2 out of 4 threads concurrently - but in vastly different ways. The T1/T2 cycles through the runnable threads every cycle, unlike montecito which lets one thread run until it gets stalled. Tukwila will also run only 4 out of 8 threads at the same time.

You can argue many moot technical points like cache hits ratio or such, but the bottom-line is that the T2 can beat any other modern cpu in it's narrow territory of favored applications. I have seen figures where a single T2 beat 4-core power6 and 16-core Clovetowns in some popular server applications, even with these massively underpowered cores. This is because T1/T2 can do zero-cycle penalty thread switching which none of the other cpus can do - that's where these 64 threads are being useful, they are all doing load/store from memory in parallel.

0
0
Silver badge
Happy

RE: Fazzi Auro

"....in some popular server applications...." Let me take a wild guess - webserving, and definately not real enterprise workloads like Oracle.

"....T1/T2 can do zero-cycle penalty thread switching which none of the other cpus can do...." There is no such thing as a penalty-free thread switch, as unless you have a separate cache for each thread you will have to start flushing and loading cache between threads, hence the question on cache hit ratios (which I assume you dodged because you know Intel have developed a massive advantage in the area, and T2 still does not have the same size of cache as its competitors).

"....they are all doing load/store from memory in parallel...." All down the same memory busses? Or are you now telling me each core hase eight memory busses, one for each thread? Not on the Sun CPU slidesets I've seen! And in between that memory and the cores is the same small cache.

"....Montecito runs 2 out of 4 threads concurrently - but in vastly different ways...." Yes, in the manner that customers need for current enterprise apps like Oracle, SAP, Siebel, DB2, even MySQL! In short, Itanium, Xeon, Power and SPARC64 (hey. I'll throw you a bone) do more with their threads than the weenie threads of the T2.

".....Tukwila will also run only 4 out of 8 threads at the same time....." Which means it will be twice as useful, and not vapourware like Rock. Until (or should that be "if ever") Rock hits the streets, it would be better to stick to comparing T2 to low-power Xeon (and hope nobody notices the tenfold price difference!), rather than Tukzilla.

0
1

RE: Netra T2 is nice step forward

"Let me take a wild guess - webserving, and definately not real enterprise workloads like Oracle."

Yes, real enterprise workloads. BTW, web servers, file and prints servers is a very big market. T1/T2 excels in workloads you didn't think. Go check for yourself, I'll give you an idea, start with the SAP-SD scores on T2 here.

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

"Intel have developed a massive advantage in the area"

Yes, can you explain why the Core2 based chips needs only 4MB cache to outperform a Montecito with 24MB cache in pretty much every benchmark with half the number of threads and 1/6th the cache ? Come to the point Matt, Intel cache designs is the best in the Industry - but now Intel can't hide the facts that these dumb IA-64 designs needs shitloads of cache to even come close to competing offerings !! IA-64 inherits it's design philosophy on the 90s workstation computing, it's a wrong product for the wrong market. The mighty chip which was supposed to conquer the world is just an also run, fighting for survival. Today nobody needs this chip except HP for the sole reason that it's enterprise existence depend on it.

"All down the same memory busses? Or are you now telling me each core hase eight memory busses, one for each thread?"

Now you have starting sounding ridiculous. When your background is not in computer architecture, at least you can stop pretending that you know about them and come back to the point, which is how these actually perform.

"Yes, in the manner that customers need for current enterprise apps like Oracle, SAP, Siebel, DB2, even MySQL! In short, Itanium, Xeon, Power and SPARC64 (hey. I'll throw you a bone) do more with their threads than the weenie threads of the T2."

Again, compare actual results. T2 has an entirely design philosophy and it excels in almost all thread-rich workloads, which are most server applications these days. That includes all these apps you point to. Again, I suggest you gather more information instead of sounding like a complete idiot.

"Which means it will be twice as useful, and not vapourware like Rock. Until (or should that be "if ever") Rock hits the streets, it would be better to stick to comparing T2 to low-power Xeon (and hope nobody notices the tenfold price difference!), rather than Tukzilla."

A 2-year old T2 will still outperform 2xTukwila in any of your enterprise apps, even with fraction of space and power. Today SPARC64 fares pretty well with Montecito, and the so called upgrade called montvale. When Tukwila comes, it will face 4-core SPARC64 chips.

0
0
Silver badge
Happy

RE: Fazzi Auro

"....Yes, real enterprise workloads. BTW, web servers, file and prints servers is a very big market...." Yes, a very big market already dominated by Wintel/Lintel, with no sign of T2 or any Sun chip making even the slightest dent. Have you considered the possible market for really heavy paperweights?

"....Now you have starting sounding ridiculous. When your background is not in computer architecture, at least you can stop pretending that you know about them and come back to the point, which is how these actually perform....." Which is a complete evasion of the question. T2 (nor T1, nor Rock) will not have eight individual memory buses per core, but will have to share the memory bus between all the threads, which means they will access the memory in series via timesharing and most definitely not in parallel as you stated, so more time for all those stalled threads to twiddle their thumbs whilst data is fetched and flushed. Would you like to pretend you know what parallel means?

"....Again, compare actual results....." OK, here's a project I worked on not too long ago. A major telco services system running AMDOCS billing software, with Oracle as the database, WebSphere for the load-balanced front-end with customer web access, and a Siebel backend for the inventory (though a supported path to SAP had to be shown for a possible future switch). The design started as an open house, with Veritas chosen as the common clustering software to ensure the design could be implemented across as many platforms as possible, the software stack strictly defined using commodity software, and EMC appointed as storage vendor. A more level playing field would be hard to produce. Proposals were received from FSC, Sun (a mix of SPARC and T1), HP (hp-ux on Itanium), and IBM. The Sun, HP and IBM solutions went through to a shoot-out and Sun were eliminated on poor performance. Sun were granted a second attempt six months down five years old! HP with an hp-ux solution were the eventual winners on performance and price. Is that real-world enough for you, or does your world have an enterprise that is completely made up of webservers?

"....A 2-year old T2 will still outperform 2xTukwila in any of your enterprise apps...." Actually, it won't, and by a large margin. Seeing as I have seen the proof of this first hand I think I can fairly tell you you are talking out of your rectum on that one (and probably in parallel at the same time!).

"....When Tukwila comes, it will face 4-core SPARC64 chips...." That's so funny! I used to have fun explaining why SPARC64 was a better buy than more Sun kit (better performance with Slowaris for less price without having to port anything!), and the Sunshine brigade would quickly poor out a venomous FUD attack on FSC as a "third-tier UNIX vendor" with "no security in their roadmap" and that the FSC servers "lacked the quality inherent in all Sun servers". Lol, hasn't the record changed now that Sun's enterprise bizz is desperately clinging to the SPARC64 life raft! But hey, aren't you supposed to say "4-core Rock chips"? Do I sense a little doubt that Rock will make it out the door?

0
1

RE: Netra T2 is nice step forward

" Yes, a very big market already dominated by Wintel/Lintel, with no sign of T2 or any Sun chip making even the slightest dent. Have you considered the possible market for really heavy paperweights?"

Guess what ? The T1/T2 line is now $285mil/qtr business with profit margin that the x86 vendors can only dream of ! It's good business.

"Which is a complete evasion of the question.

<snip>

...rant

"

What's needed is not eight different memory buses. T1/T2 have four different on-chip memory controllers with ample dedicated memory bandwidth, they don't need to load/store at the same time, even a computer semi-literate can say that.

The design trades latency for bandwidth. Anyway, all arguments can't ring a bell into your tiny brain, public benchmarks substantiate what I mean.

Any why did you evade my question why Itanium needs such massive caches to come not even close to Core2 performance. What mighty chip is slowly dwindling towards extinction. With all might HP had, they managed to grow only 1% in the

HP-UX/PA-RISC/Itanium/Alpha business - it's time for Intel to abandon the loosing chip. The countdown starts now, as the HP/Intel Itanium supply negotiation comes to tragic ends, and Intel slowly clearing the way with CSI aka Quickpath Xeon/Itanium socket compatibility.

" OK, here's a project I worked on not too long ago ....<lies lies lies>"

Again, provide some publicly available benchmarks. With the credentials you have here, no one will believe your lies, specially with your past history of HP drumbeating.

" Actually, it won't, and by a large margin. Seeing as I have seen the proof of this first hand I think I can fairly tell you you are talking out of your rectum on that one (and probably in parallel at the same time!)."

ECHO BACK. Again, point to publicly available results/indications/Intel claims ...whatever - we don't want the noise from the gory holes of your body.

Keep beating the HP drum fanboy,, and sun-bashing for good - someone or the other will surely pay attention.

0
0
Silver badge
Happy

RE: Fazzi Auro

"....<lies, lies, lies>...." Oh dear, denial is not a very good counter-argument, it just sounds childish and petulant. Just ask your Sun salesgrunt what the dominant software stack is in telco billing and on which platform, and after much grimacing and grunting he'll have to admit it's AMDOCS on hp-ux, usually with Oracle. Not Slowaris, not MySQL, and definitely not T2, T1 or UltraSPANKED.

"....The T1/T2 line is now $285mil/qtr business...." In the scale of the enterprise market even just here in the UK, that's peanuts. And that's before you consider that the ProLiant bizz for HP worldwide is probably better than $285m in a fortnight! In Q2 07 alone HP shipped 22.3 times as many x86-64 servers as Sun. That statement is from IDC, but please feel free to call them liars.

"....With all might HP had, they managed to grow only 1%..." Yes, HP saw 1% growth in factory revenue for their enterprise servers. Actually, seeing as the new Integrity systems cost less than the old Alpha and PA-RISC ones, that means they shipped more servers to match the old figures, with more margin per server, which means more customers took advantage of the performance advantage AND price advantage and bought more HP enterprise servers. So HP not only protected it's installed Itanium base, but converted those Alpha and PA-RISC customers that wanted to migrate to Integrity rather than lose them to other vendors, and found time to sell to new customers too, which means they took sales from Sun and IBM. Meanwhile, Sun's enterprise server sales stayed static or declined depending on who you ask, as Sun struggled to hold onto their installed base, and did not expand. Face it, in a $5Bn enterprise segment, 1% growth is a lot better than 0%!

"....Again, provide some publicly available benchmarks...." Last year I put rx2660s up against T2000 for one of our projects. We went to a shoot-out because the Sunshiners disputed the published SPEC results for SPECjapp2004 with Oracle, which was central to the project. After a fortnight of tuning, the Sunshiners couldn't get above 730 jops despite claiming they could hit 730+. First run the rx2660 cracked 855 jops with about twenty minutes of tweaking (HP's SPEC submission states 874 jops). Management didn't bother asking for a second run, the project went ahead with rx2660s. Please, don't look foolish by squealing "lies, lies, lies" as you'll have to go argue with SPEC.

Boom, boom, boom goes the HP drum. Whine, whine, whine goes the Sunshiner.

0
1

RE: Netra T2 is nice step forward

"denial is not a very good counter-argument, it just sounds childish and petulant. Just ask your Sun salesgrunt what the dominant software stack is in telco billing and on which platform, and after much grimacing and grunting he'll have to admit it's AMDOCS on hp-ux, usually with Oracle. Not Slowaris, not MySQL, and definitely not T2, T1 or UltraSPANKED."

Oh yeah, throw us some proofs instead of patching up with another set of lies. HP-UX is dying, the sooner the better.

"In the scale of the enterprise market even just here in the UK, that's peanuts. And that's before you consider that the ProLiant bizz for HP worldwide is probably better than $285m in a fortnight! In Q2 07 alone HP shipped 22.3 times as many x86-64 servers as Sun. That statement is from IDC, but please feel free to call them liars."

Again, don't try to circumvent the argument. It's about shipping the non-x86 boxes and the profit margin on the systems. Show us how the non-x86 business is doing for HP. What a shame that HP can translate their huge (compaq inherited) x86 presence into meaningful non-x86 sales growth !!

" Last year I put rx2660s up against T2000 for one of our projects. We went to a shoot-out ...."

So now we have to rely on Bryan Matty reference benchmark instead of SPEC/TPC/SAP-SD....benchmarks !! Talk about how to keep throwing meaningless numbers without any substantiation.

"Boom, boom, boom goes the HP drum. Whine, whine, whine goes the Sunshiner."

Or rather, "Sink Sink the Itanic with HP fanboys in grim, up goes SPARC and the Sun"

0
0
Silver badge
Happy

RE: Fazzi Auro

Oh dear, more rant than substance yet again!

"....Oh yeah, throw us some proofs...." Erm.... I did.... guess you just had a problem comprehending it. Tell you what, you tell me what your native language is and I'll try and convert the response via Babelfish in very easy (non-technical) words.

"....Show us how the non-x86 business is doing for HP...." Well, like you said, it carried on growing in revenue, whilst at the same time Sun's didn't. Looks like more like Slowaris is dying than hp-ux. Or are you going to tell me a negative growth figure is better than a 1% growth!?!?!? And HP had a very healthy x86 bizz before the Compaq merger, it was called Netserver, and it was bigger then than Sun's tiny slice of x86 is now. But then the merger was a major upset for Sun as Compaq used to be the number one seller of Slowaris x86 (Sun were just as hopeless at selling it then as they are now).

".....Talk about how to keep throwing meaningless numbers without any substantiation....." Actually I stated an example of where real life testing backed up a published SPEC benchmark. You can sit in fantasyland and deny my testing if you like, but you can't deny the published SPEC figures. Just concentrate real hard, try not to let the Sunshine blind you, and type "www.spec.org" into your browser....

"....up goes SPARC and the Sun...." Only a complete idiot would believe that. Sun's SPARC business has been in decline since before they canned UltraSPARCV. It has become such a complete shambles they have had to buy an aging design (SPARC64) from a competitor (FSC) they used to rubbish, and switch to producing a platform (x86) they used to accuse of being a toy - how embarrassing! T1/T2 are so niche it's hard to see how Sun can finance them in the longterm, and Rock is another vapourware product that looks like being canned just like UltraSPARCV. If that's "up goes SPARC" then it must be up in smoke!

Please use the "Joke Alert" icon for your future postings as they have zero technical or commercial content and only serve to show you for the fool you are.

0
1

RE: Netra T2 is nice step forward

"Erm.... I did.... guess you just had a problem comprehending it. Tell you what, you tell me what your native language is and I'll try and convert the response via Babelfish in very easy (non-technical) words."

I think it's you having a problem comprehending, otherwise by this time you would have come up with some reference to substantiate your bullshitting.

"Well, like you said, it carried on growing in revenue, whilst at the same time Sun's didn't."

You call that growth !! After many consecutive years of decline a slight blip does look like growth to you.

"And HP had a very healthy x86 bizz before the Compaq merger, it was called Netserver, and it was bigger then than Sun's tiny slice of x86 is now..."

Very true. Except that the Netserver business was DYING FAST against DELL and Compaq, whereas Sun's tiny x86 business is growing FAST even with all the x86 giants. So HP had to buy Compaq, ditched Netserver and threw up all the towel on Proliant.

"Actually I stated an example of where real life testing backed up a published SPEC benchmark" ...blah blah blah...

That's nice, YOUR "real life testing" !! And no public data to validate !! I think you think everyone is a fool,or stupid as you are.

"Sun's SPARC business has been in decline since before they canned UltraSPARCV. It has become such a complete shambles they have had to buy an aging design..."

What a joke !! So the Itanium design is so very modern that every major server vendor abandons it except the ones who have no option but to cling to it !! The HPC vendor who bets on it goes bankrupt and has to switch to x86 to survive. There is nothing wrong in Sun in embracing x86 since no one can survive without it - it's called being pragmatic. What Intel is doing with Itanic is the last ditch efforts to save the prestige, and HP high end business has nowhere to go without it.

"Please use the "Joke Alert" icon for your future postings as they have zero technical or commercial content and only serve to show you for the fool you are."

Does that really matter ? Specially, when you act stupid and pretend to ignore the inevitable !!

0
0
Silver badge
Happy

RE: Fazzi Auro, live from Hollywoood?

I'd heard the comedy writer strike was off but I didn't realise that extended to the web as well!

"....otherwise by this time you would have come up with some reference...." Check out the SPEC figures. SPEC figures, go check. Running out of ways to say it! Type "www.spec.org" into your browser, select specjappserver2004 public results search.... Oh, what the heck, just forget it, you're obvious not reading before frothing.

"....That's nice, YOUR "real life testing"...." Well, I usually do my own testing, that's how I justify my decisions and recommendations. Sorry if that's a little to real for you, I'm guessing you've never done any enterprise testing but just accepted whatever the Sun salesgrunts told you as gospel.

"....threw up all the towel on Proliant...." You mean HP vomited up a towel???? For a "thrown up towel" ProLiant seems a very healthy product. Maybe Sun should try some of this towel vomiting with Rock!

"....So the Itanium design is so very modern that every major server vendor abandons it except the ones who have no option but to cling to it...." Oh, so in that case we should see Sun announce a range of Itanium servers by next year (probably badged FSC ones). Of course, there's no way any server vendor would make them because customers buy them, then? So, the fact that HP's Itanium server sales are going up faster than the decline in Alpha/PA-RISC sales, that's just fantasy is it? Do you seriously believe that IDC, Gartner, the whole tech press (including El Reg) are just lying to make you feel bad? I think you need to talk to someone about that paranoia of yours.

"....you act stupid and pretend to ignore the inevitable...." Mr Pot, meet Mr Kettle! Try taking a deep breath between rants, it may help with the verbal diarrhea. Then again, probably not. Face the facts - SPARC is dead on its feet, and unless Rock comes out on time with a big advantage Sun will be forced into another embarrassing volte face to Power or Itanium in order to keep some form of enterprise bizz. Seeing as Sun originally did porting work for the first generation of Itanium, it would make more sense for them to complete that than switch to Power. Even if they ditch Slowaris all together and do an even bigger volte face and go with Linux, it would make more sense on Itanium (FSC already do it), and then they could do the biggest volte face of all (which would probably make your toes curl with apoplexy) and do real 64-bit Windoze on Itanium.

0
1
This topic is closed for new posts.