Surely with the 'itanic under discussion they should be abandoning ship (women and children first) before they hit the iceberg?
Come on, where has the Register's style got to?
Microsoft is abandoning Intel's Itanium platform after the current release of its server and tools software. The software giant said Monday that Windows Server 2008 R2 and the forthcoming SQL Server 2008 R2 and Visual Studio 2010 would be the last versions of these server and tools products to run on Intel's giant chip …
Hey Matt B remember heehawing at Red Hat dumping Itanic by saying well its still supported on the most common computer OS? Itanic RIP yet another billion+ dollar chip mistake in Intels annuls. Funny how chipzilla can weather these and still claim world domination (actually more because Hector Ruiz is a short bus window licker but neither here nor there).
..why Intel spends a single cent on this architecture. I guess they get a fat check from HP every year. Still it is a potentially dangerous distraction from the money-making x86 business and the right thing would be to stop it immediately.
They certainly have by now sucked out all interesting knowledge from HP's enterprise business and transferred that into the x86 product line.
"Itanic RIP yet another billion+ dollar chip mistake in Intels annuls. "
I'm not so sure. I mean, Itanic failed -- big time -- but it did succeed in persuading companies to quit developing Sparc, PA-RISC, Alpha, MIPS, and others, most of which were competitive (speed-wise) if not faster than what Intel was producing. They either became dead in the water, or went to squeezing more cores onto a single chip (but without actually advancing the design of each core). Basically, IBM was about the only one to not drop chip development (of the PowerPC/POWER line). Oracle is planning to continue Sparc development now, but this is after years of pretty much stagnation. Instead of having numerous chip designs competing to a greater or lesser extent, Intel pretty much has AMD/ATI in the "normal" range and ARM nipping at the low end.
I don't think Intel really planned that far ahead, I do think they intended Itanium to be a successful line... and they screwed up severely, but this seconday effect worked out great for them.
Well really it was AMD that screwed Intel up by bolting 64bit extensions onto their 32bit x86 offering so that it would run both, then making it work and making it quick. Intel had no choice but to follow suit, something they never ever wanted to do. There was supposed to be a clear dividing line between x86-32 and IA-64 and no cross over. AMD forced them into x86-64 and into playing catch up for a few years. From that moment on the glorious dream of Itanium replacing x86 in the volume server market was doomed to hit an iceberg. It might have helped if the first Itanium chips had shown blistering performance and given people something to be excited about, but Merced turned out to be a real dog.
Itanium's future rests in the hands of HP and HP-UX now that SGI, Redhat and Microsoft have voted with their feet. We will have to see if that's enough. Doesn't look very promising for HP-UX though as it's now caught in a tightening vice with AIX and Linux on IBM's Power7 at the top end, plus Linux and MS Windows on Nehalem-EX and AMD's Magny-Cours chomping deep into its middle ground from underneath.
The next 12 months are the litmus test. HP not only have to make enough sales to make Itanium chip production profitable, but also to fund future development and cover the cost of Fab upgrades to stop it from falling behind. If they can't, then Intel will do to HP and Itanium what they did to Compaq and the Alpha chip. Chip updates and Fab improvements will come later and later and it will fall further and further behind the field until HP are forced to give up the ghost and slap a 10 year tick over retirement plan on it.
"Well really it was AMD that screwed Intel up by bolting 64bit extensions onto their 32bit x86 offering so that it would run both, then making it work and making it quick. Intel had no choice but to follow suit, something they never ever wanted to do."
Hmmm, an interesting, if flawed, perspective you have there.
Firstly, AMD did not "screw" intel. They are a competitor. They saw a market demand and filled it. It is nobodies fault but intels that this happened. If intel were not busy arrogantly trying to force the market in a direction they clearly did not want to go in then perhaps intel wouldn't have had to play catch up later?
As to why intel wanted to force the market to move, well it all comes 'round to dominance and monopoly marketeering. i386 is a relatively open* architecture and intel compete with a couple of other players in that field. Given a choice between competing and monopolising, most corps will choose the latter. Had itanic taken off it would have left intels current competitors behind as also rans and intel would once again own the market. HP and intel bet the house on this happening and were indeed brought undone by AMD.
Too bad, so sad.
* It is licensed to AMD and VIA by court edict and held at a "reasonable level"
>>The next 12 months are the litmus test. HP not only have to make enough sales to make Itanium chip production profitable.
Given that HP see themselves as chief M$ product pushers, this is a serious body blow.
HP may have had a rationale to buy a services company, follow the cloud mob and babble about virtualisation.
T, because it is not the rise of one advanced CPU architecture
Cant find reference, but some-one (Cringely ?, ElReg?) suggested years ago that the Merced/Itanic was a distraction aimed at killing other RISC chips to clear the road for Intel
If so, it has worked admireably, despite the brilliant efforts of AMD to upset the Chipzilla applecart.
Most HPUX developers were canned in 2004 ITIRC, so HP really did not want to keep it alive.
Despite this, HPUX is claimed to have the same virtualization as Solaris 10 and AIX 6.x.
"Basically, IBM was about the only one to not drop chip development (of the PowerPC/POWER line)."
Who's feeling smug for sticking it out and not drinking the Intel kool-aid now then?
What'll be hysterically funny in the not too distant future will be the sight of HP running around like headless chickens when Intel decide to cut their losses.....
I could have sworn that a couple of people well known round here were assuring us in the last few days that Itanic would live on for ages because only Itanic could provide the mainframe-class Reliability Availability and Serviceability that some customers require. (Hello Matt, hello TPM), and that no x86-derived machine could possibly compete.
Personally I see little significant difference in RAS terms between a decent AMD64 Proliant and an entry level Itanium box. Except for the RAS related to the OS, obviously.
Apparently Microsoft don't see that the alleged Itanium-specific hardware RAS is all that interesting to customers (or profitable to Microsoft) either. Fancy that.
The 256 processor limit could be down to a couple of things: an eight bit CPU ID would be an obvious choice.
Possibly, though, there's nothing to be gained by going beyond 256 CPUs because the scheduling and locking overhead is such that there's no performance benefit. Getting Linux to scale to lots and lots of CPUs (I'm not sure what the effective limit is) took a fair amount of effort.
On the other hand, the limit could simply be that MS haven't tested on anything bigger :-)
HP's announce later this month is in deep trouble. The bogus announcement in February without any actual servers is another delay tactic in the multi-year debacle. HP is dropping Linux from its marketing and will only offer VMS on the blades. They are dumping all rack based systems and forcing customers to buy HP bladesystems.
The problem with Tukwila is it only has 4 QPI's and two half QPI;s. This architectural limitation is a product of Intel focusing on chips and not systems and Intel trying to use the same socket as Nehalem years ago. HP is stuck with having to develop a proprietary chipset for this "open" chip to scale past five sockets. The problem is the four wide blade eight socket Tukwila offering does not have the HP 3000 chip so the performance suffers.
For customers who don't have the time to migrate they can continue with Tukwila but the blades will not buy much in space reduction, nothing in performance and the performance/scalability will certainly not be sufficient for databases.
Look at the numbers and you can see why everyone is divesting in Itanium.
65nm when everyone else is 45nm, 4 core when everyone else is 8 core, Same performance per core of prior generation when everyone is increasing performance per core and subsequently no improvement in software licensing.
Don't do status quo, do amazing and move to Power7
So, are the remaining 5% running VMS? I hope so, and also hope that the stern-first slide of Itanic means HP finally porting to x86_84.
I'd like to re-warm my DCL chops on an old AMD box currently gathering dust.
Of course, if someone wants to sell me a desktop Itanium/VMS box, cheap, I'm game.
(What, no Palmer icon?)
I think that competition is good. It forces everyone to be better, and the customers win.
I would not like a scenario where only one chip manufacturer is left. Slowly but surely the price would be increased for each new CPU generation, and development pace would slow down. Us customers would pay more for less performance.
I think the first Intel Pentium cpus costed really much, relatively speaking? Then AMD came along and forced Intel to lower the prices and quicken development.
SP&L huh? Oh joy, feeling quite satified after reading this email....
Spin us a story for old times sake Matty boy, a 1000 word essay should do the trick! See if you can get in an anti-Sun twist (sunshiners, blah, forthing, blah, firesale...) & let us relive the good'ol days.
Hey ho, my turn, P&L, hee hee hee..... :-)
Cheap used itanium? Look online, i think they are in fact VERY cheap. I got several at the surplus i work at, i couldn't sell them. Mh recycler couldn't either. Since Itanium is big & power hungry (and itanium) hobbyists don't want them (except to use with vms). Since they are super high end, those customers don't want used machines they can't get a warranty on.
Sorry, Henry, I missed your post earlier. Hp offer a service to buy up old kit and put it through their Renew program, that reconditions and tests the old kit to factory standards so hp and their resellers can sell it with a full hp warranty, or so it can go into hp spares stores for their engineers. If you want to buy secondhand Itanium gear but not pay out for hp support or hp warranty services then there are third-party companies (such as Pheonix here in the UK) that will offer their own support and services. I hear even IBM will offer support and hosted services for hp Integrity, though I also hear they have to call on hp consultants to do a lot of any subsequent hp-ux work.
Like many geeks, I have squirrel tendencies and I currently have several bits of hobby kit at home rescued from one company's or another's clear outs. Amongst them is a two-CPU zx6000 that dual boots hp-ux and Linux and doesn't make too much noise or heat (less than one of my PCs which has an 125W Athlon x2 6000+ and positively glows!!).
Sorry, looks like some other people beat you to it!
Besides, with hp-ux being the top enterprise UNIX, why would hp need to go back to MPE? I suppose MPE was one of the group of OS that smashed the IBM mainframe monopoly and made a big hole in IBM's profits, but then Non-Stop is doing that today already, so no need for MPE. Oh, I see - you were so busy fantasising about the death of Itanium you lost all touch with reality (again)!
<Yaaaaaaaaaawwwwwwnnnnn!!!!> Aw, not another "Oh-I-really-hope-so-'cos-I-really-want-it-to-be-so" reporting of Itanium's death? How many is that this year alone? And how many times have you all been wrong before?
What, not having Windows (in eight-odd years time) means Itanium is suddenly dead? Puh-lease! By that measure, every one of Itanium's RISC competitor's has been stillborn (well, Rock actually was, but not because of not running Windows). Please remind me how many Power versions have run Windows....? And then go look at how hp-ux on Integrity is dominating the UNIX high-end, and how Non-Stop is taking sales away from IBM's mainframe party. Even the OpenVMS fanbois make IBMers worried - will anyone be as loyal to humdrum AIX after Power gets canned? Unlikely!
If you forgot to read the figures in the Gratner article as reported here on the Reg (http://www.theregister.co.uk/2010/02/24/gartner_q4_2009_servers/), hp seem to be doing quite fine in the high-end. In fact, given the average server shipment value as discussed in the comments off the same article, hp are selling more top-end servers than IBM and much more than Soreacle. I'm betting Ponytail would just love it if Niagara could be as "dead" and make as much of a profit (Sunshiners, you will probably have to take a break here to check the dictionary definition of "profit", you're not going to be at all familiar with it)!
As for Dunked-in-the-brown-stuff Manor, I can quite comfortably say (since the new Integrity blades haven't hit the street yet) you haven't a clue and are talking from your nether regions. Please come back when you have actually had a chance to bench the real kit rather than just fantasised about it. Same goes for your hilarious conclusions about chipsets you definately haven't touched - from the poor qaulity of your FUD I assume IBM marketting is hiring from the Sun carcass?
I know you IBMers like to dream about Power taking over the World, but there's just two big problems with that. Firstly, Power's market is shrinking (it's happening to all the commercial UNIX markets), so IBM is fighting over a smaller pie every year and that smaller pie means less and less money to plough into Power (hence the imminent death of Cell - IBM can't afford to develop both on the shrinking profits). Secondly, Intel has the money from x64 - even if Itanium wasn't making a profit (and it is), they could afford to prop up Itanium for years and not blink. IBM can't say the same about Power, they don't have a partner product of the same scale, which does not bode well for its longterm future (and is the reason for a very short Power roadmap). The story is even bleaker for those Sunshiners still holding out for a rescue from SPARC64. Slowaris's only hope is x64, which is ironic given the years with which the Sunshiners poured venom on anything Intel.
And hp is the leading x64 vendor (and, amusingly, likley to soon be the leading Slowaris vendor with Slowaris on ProLiant sales!), so as the x64 tide sweeps into the RISC space it is hp that benefits most, not IBM, Soreacle or Fudgeitso. So hp also has the profits to keep developing Integrity. Which is the reason the IBMers FUD Itanium so hard, they know it's a two horse race now and the Power horse just doesn't have any more development room. Simply squeezing more and more simplified cores (AKA plan copy Niagara - wierd!) onto a die is not going to defeat x64, let alone stand against Tukzilla. Intel is going to win one way or the other, with AMD happy to ride along on their coattails. The pincer attack of Itanium in the high-end and Xeon from below has already killed Sun, why should Intel or hp let up the pressure on Fudgeitso and IBM when the tactics are working so well?
/Oh yes, I'm certainly SP&L at you clowns!
Bottom line: Microsoft is pulling the plug on Itanium just like RedHat
HP is not dominating the highend market. Power 595 outsells superdome I think 3 to 1.
Not sure why anyone would buy a Stuperdoom when it was supposed to be replaced two years ago and customers are still waiting for the joke which will arrive in August. A 16 socket cell board config that cant compete with the Power 770 let alone the p595 follow-on.
The Tukwila blades will be a clear indicator that HP is divesting in Itanium just like everyone else and is just making two socket blades that fit into the existing HP bladesystems. Ok HP can group 1, 2 or 4 together but the low performance and 4core/65nm chips and lack of sx3000 is a major problem.
Power is taking over the Unix market with HP and Sun being a very distant third place. Funny to see HP becoming #2 again because SPARC is in such a free fall that HP while falling is not as bad. Power7 is the successor for cell and the PS4 is proof of that.
Intel does not like wasting money any more than the anyone else and Itanium is worse than a billion dollar boat. Only Larry can afford those kind of losing boats.
HP only leades in x86 because of Compaq and is losing that install base since they have not been able to have an 8 socket system for intel. The new DL980 does not compare to IBM eX5 and does nothing for the new memory expansion.
Sorry p7 is not simplified cores like Niagara, it is actually more function per core and the eDram allowed for the room to do that with 32MB on chip cache.
x86 on the low end and Power on the highend with the added killer of Oracle pricing is what killed SPARC. Itanium had nothing to do with it.
Matt....what happened? You were starting to move over to the Power camp....if you cant beat them join 'em. Come on over the water is fine :-)
Oh dear, didn't Ms Kebabfart get the memo about trying to raise the industry's perception of women in computing, not blow it with poor quality FUDfests? XD
"Bottom line: Microsoft is pulling the plug on Itanium....." Real bottom line - hp still makes more out of WIndows on x64 than IBM, and will benefit most as x64 continues to eat into the RISC space. And since Power can't run Windows the sum effect is no benefit whatsoever to IBM, just some frothing and FUD from their trolls.
"....HP is not dominating the highend market. Power 595 outsells superdome I think 3 to 1....." I'm beginning to see that thinking is not your strong point. Both Gartner and IDC disagree with your "thinking", and the average unit shipment values discussed in the thread http://forums.theregister.co.uk/forum/1/2010/02/24/gartner_q4_2009_servers/ show that hp is shipping the enterprise high-end boxes, not IBM. That is unless you want to contend that IBM are giving away top-end servers for $60K-odd a pop, which sounds even worse than the old Sun desperation.
"...Not sure why anyone would buy a Stuperdoom...." Well, I'll try and explain in small words so you don't get too confused. In our benchmarks, the Superdomes have been outperforming the pSeries offerings for years. I think you'll find there's quite a few other customers that came to the same conclusion (try here, for example http://h71028.www7.hp.com/integrity/w1/en/high-end/superdome-anniversary-case-studies.html). Please try and tell me in all those Superdome sales there weren't any head-to-heads with pSeries where Power didn't do the job as well as the hp package. If I remember correctly, even AMD chose Superdome over pSeries to run the software for Opteron manufacturing!
"....waiting for the joke which will arrive in August..." Sorry, are we talking about the new top-line Power7 kit now? After all, that won't be available to customers until "August earliest, more likely September", as told to me by our IBM rep. Please feel free to disagree with him, but unless you're in IBM I don't see how you could possibly know more than him.
"....The Tukwila blades will be a clear indicator that HP is divesting in Itanium...." Now that did make me laugh! How can hp taking advantage of its massive lead in the blades market be a bad thing or "divesting"? Everyone knows hp has been leading the blades market for years, even TPM will grudgingly admit it. And I don't see any talk of Power8 blades on the IBM website - has IBM given up on Power blades? They've already cancelled the Cell blades, as reported here on the Reg. Is IBM planning only x64 blades for the next gen BladeCenter? You IBMers really should learn to stay away from the blades topic, it just makes it too easy!
"....Ok HP can group 1, 2 or 4 together ....." Hmmm, so being able to plug modular blades into one larger SMP instance is bad (I'm guessing 'cos IBM can't do it with their blades), but good when IBM do it with their rack servers? Back to troll school for you!
"....but the low performance..." Again, please supply full details of the benchmarking session you have personally run to get any performance comparison (not holding my breath). Oh, was that an admission you haven't even seen the kit, let alone benched it? I suppose the next claim will be that you're a natural blonde, an athletic 5'6'' and with natural 34DDs?
".....and 4core/65nm chips...." How is the chip size a problem if it does the job in real World situations? Can you even spell "feature-sell"? Ask my CEO if he's more worried about our next CRM system performing to requirements, reliably and to a good price, or whether he is concerned that the chips inside are 65nm, 45nm or 25nm? Since I'm also guessing you don't do much of that real business stuff I'm going to say you think he thinks the size of the chips is a really important issue. I'm hoping you're not going to be part of our IBM account team as I'd actually like to get some Power7 kit to play with, not have IBM get laughed out of the building.
"....and lack of sx3000 is a major problem...." What lack of sx3000? You're just making up stuff on the fly now. I can just as validly say that since the new top-end Power7 kit won't be here until August/September that IBM "must have a problem" with the new Power7 chipsets. Well, I would if I was a frothing troll without a clue.
".....Power is taking over the Unix market with HP and Sun being a very distant third place....." Maths is obviously also not one of your hot topics, but a quick check of the Gartner and IDC figures should show even someone as challenged as yourself the truth. Please also try and grasp the concept of what IBM call "leakage" - that's sales that go from being UNIX on EPIC/RISC to being Linux or Windows on x64. The reason this concerns IBM so much is they know the majority of those "leakages" go onto hp ProLiant, which means as the UNIX market shrinks it is hp that benefits most, not IBM.
"....Power7 is the successor for cell...." No, Power is a generic chip that cannot replace the specialised Cell, it is just IBM trying to cut costs as they can't sustain development and production cost of the two lines.
"....and the PS4 is proof of that...." Sony don't have a choice. IBM have dropped them in it, and - like Apple and PowerPC - they're only choice for the next gen PS will be an x64 chip as Power will be far too expensive to compete with other gamestations or gaming PCs. How much money will IBM be losing on each PS4 chip given that they flog the Power CPUs out at many thousands of dollars each? Sorry, I shouldn't really ask you that given your maths issues, it would only give you a headache as you try to reach a conclusion.
"....Intel does not like wasting money...." So who says Itanium is "wasting" Intel's money? Please provide the Intel business report that states that Itanium is losing Intel money. Oh, another surprise, not - you can't. Just more unsubstantiated FUD (I wouldn't accuse a lady of blatent lying). With the MIPS vendors long gone, Sun down on the matt, Fudgeitso on the ropes and IBM spinning like crazy, it looks like Itanium has been a cracking investment for Intel.
"....HP only leades in x86 because of Compaq..." <Yawn> Yes, and hp bought Compaq how many years ago? If hp had done nothing since to grow their x86/64 marketshare then they would have fallen back after all the years, whereas the reality is a constant stream of innovations and developments have not just kept Compaq customers loyal but also meant gains in marketshare at the expense of IBM, Dell, and Fudgeitso. What must really gall IBMers is that even the combined hp and Compaq was still a much smaller company than IBM, but they managed to outperfrom IBM in the market. Maybe that's because, unlike mainframes, x64 is not a monopoly.
".....losing that install base since they have not been able to have an 8 socket system..." Do you do any research before posting your waffle? The gartner report mentioned above says hp again shipped the most server units. And if you had even the slightest inkling about the market then you'd know that the biggest seller is not 8-ways but 2-socket servers (like the hp ProLiant DL380, again the largest selling 2-socket server for Gawd alone knows how many quarters). Your schpiel is all the more amusing given how hard IBM FUDed the 8-socket DL785 G5 servers over the last few years. I'm not surprised you're flipping 180 degrees now you're finally back in the 8-socket x64 market with an Intel-chipped server.
"....p7 is not simplified cores like Niagara..." Power6 saw IBM move ahead away from old-style RISC and out-of-order execution (just like EPIC), because even IBM had to admit RISC has had its day. But power6 had problems and had no headroom for development, after the minor boosts of Power6+. So now IBM have switched back to what is essentially a Power4/5 core with die-shrink and a few extra features. - hardly revolutionary. Scared witless of the growing core count of Nehalem and Magny Cours, IBM have done exactly what Sun did with Niagara - go back to an old core that was easier to manufacture and shrink it so they can get as many cores as possible onto a die. The end result is they get faster clock (by die shrink) and abandon any pretence to real development. The problem for Power is their RISC design has hit the buffers, there is nowhere else to go, and they can only keep doing die-shrinks for so long because they cannot compete with the the economies of scale of the x64 juggernaut.
"....x86 on the low end and Power on the highend with the added killer of Oracle pricing is what killed SPARC...." Power on the high-end? Debateable, especially as hp-ux on Integrity has outsold Power in the high-end for years. But seeing as Oracle's leading partner is hp (more new Oracle installs go on new hp servers than any other vendor) and hp is the leading Wintel/Lintel vendor by a mile, I can actually agree with you in that hp and Intel killed Sun (with a small helping hand from Armonk).
"....Matt....what happened? You were starting to move over to the Power camp...." Just because you confused my willingness to give both platforms a chance in real World tests does not mean I am going to act as an echo chamber for your vacuous FUD. FUD is just a way to turn customer off. And seeing as your FUD is the equivalent of you peeing in your shrinking pool I think I'll stay out here in the open sea, thanks.
/SP&L, though rather sad at the fall in quality of IBM trolls.
--"If you forgot to read the figures in the Gratner article as reported here on the Reg (http://www.theregister.co.uk/2010/02/24/gartner_q4_2009_servers/), hp seem to be doing quite fine in the high-end. In fact, given the average server shipment value as discussed in the comments off the same article, hp are selling more top-end servers than IBM and much more than Soreacle"--
That's kind of bending the numbers to fit your world view isn't it. So HP are getting more value per Itanium server than IBM or Sun/Oracle are from theirs. That's good, but from the figures quoted IBM still made much more money from bare metal P7 sales than HP did from Itanium. Plus Sun/Oracle got nearly three time as much tin on the ground and IBM twice as much. I'd bet that when you factor in support contracts and consultancy fees, and peripheral kit sold along with all those deals, that HP get bumped into a distant 3rd place on total sale value.
You know you can sex this up as much as you like, but Itanium is not delivering what was promised for HP any more. That's not HP's fault, it's Intel's for letting their Itanium road map milestone slip and slip and slip. Remember that in 2005, when all Intel had were single-core Madison Itaniums, Intel was saying that Montecito would be out in 2005, Montvale would follow fast on its heels in 2006, and Tukwila would debut in 2007. As it turned out Montvale came out in 2007 and we've had to wait 3 fecking years for Tukwilla. If we have to wait another 3 years for Poulson to take Itanium to 8 cores, well what's the bloody point. It will be irrelevant by then.
"....That's kind of bending the numbers to fit your world view isn't it...." Well, as someone once said, "there's lies, damn lies, and statistics". Everyone cherrypicks data to back up their arguments, but I didn't need to cherrypick, I just used the one undisputed fact on average server unit value. But are you disagreeing that hp sold their Integrity servers on average for a higher unit price than IBM sold pSeries? Please do, just for a laugh. No? Well, on with driving a bus through the holes in your argument then.
The key fact is that the average server unit sold by hp was approximately 38% more than the average IBM pSeries unit price. I'll pause here whilst you to digest and try mentally to refute that statement. You may want the help of an adult with the calculations. Right, ready for the conclusion?
That 38% indicates the average hp Integrity box sold was indeed a larger beast than IBM's average pSeries sale as both companies offer competitive prices and discount heavily to match each other. I'm sure even someone as obtuse as yourself would not contend that hp would have won deals putting a 38% more expensive solution up against an IBM offer, so it is reasonable to suspect that when they went head-to-head the prices were comparable. That also implies that - since larger boxes usually go into more demanding roles - hp was winning more of the high-end deals (larger boxes go to more demanding roles). Following that logic, it is possible to see that customers prefer hp Integrity (and therefore hp-ux) to IBM pSeries (and AIX) for their high-end requirements. I bet you're wriggling in your seat whilst not trying to admit to that conclusion! Whilst you're busy writhing, let's get to the fun bit - show me the money!
Those high-end deals usually go out with the most expensive support options, more services and lead to more of the same level of margin on opportunities in associated areas such as storage, management software, consultancy services, backup and DR - what is known as "pull-through". It therfore follows quite logically that, even though IBM and Sun sold more low-end kit, hp made much more margin on each server (big iron has higher margins than the low-end), and then made much, much more margin through the pull-through sales. As a rule of thumb, when budgetting for mission critical solutions, we assume the life-time costs of a solution (including planning, upgrades, support, decommissioning and/or re-deployment to other roles), regardless of vendor, will be about five times the original purchase price, and most of that money will go to the hardware vendor. So that means hp not only probably made much more margin than IBM on each server, they also made more on pull-through and will make more during the lifetime of those servers. So your argument that IBM made massively more money than hp simply because IBM sold more units overall is likely to be very wrong due to IBM not making as much money on each server sold, either through margin at the time of purchase or pull-through during its lifetime. This is the prime reason Sun went steeply into decline whilst still shifting a high number of (low-end) units.
"....You know you can sex this up as much as you like, but Itanium is not delivering what was promised for HP any more...." OK, so how is delivering high-end servers that customers prefer to buy over what IBM is offering, for mission critical roles (that's the high-end), "not delivering"? That implies that, as IBM is losing those high-end deals to hp Integrity as the reasoning above shows, then IBM is "delivering" even less! Your argument, not mine, no sexing up required.
OK, so no cherrypicking yet, just logical argument to refute your suggestion. If you'd like some cherrypicking, I can go over how Power6 was supposed to be in for 2006 (so, how late?), or point to a whole number of other slippages in IBM product deliveries. As said before, trolls in glass houses.....
"....If we have to wait another 3 years for Poulson to take Itanium to 8 cores, well what's the bloody point. It will be irrelevant by then....." Strangely enough (well, strangely for those like you that can't see for their blinkers), hp Integrity has won that high-end despite the Itanium delays and despite the latest versions only being dual-core to Power's four-core offerings. Which does imply that Tukzilla, as it offers more than the current Itanium offerings, is not only rellevant but will continue the hp trend for beating IBM in the high-end. And therefore that Poulsen will also, even if it does not arrive for three years, and even if IBM has shrunk it's dated design to squeeze sixteen cores onto a single die. That's if Power survives the x64 onslaught to reach another generation.
/SP&L @ the bus-sized holes in your argument.
Stripping out all the ad homenim and dot to dot Disney Land logic from your argument, what you're basicly trying to sell is that because the unit price of an Itanium sale is higher than the rest (which I never disputed) then this "obviously" means (by magic) that these sales must be to more prestigious customers. What an ego. And that by association (or some dimension bending osmosis) the accompanying sales opportunities (storage, consultancy, etc) are bountiful rivers of flowing gold. And that these lucky customers, with bottomless money wells and orchards of lush fruiting money trees, are falling over themselves to buy the more expensive Itanium solutions because, well, it's Itanium. That's an interesting world you live on. Does your emperor wear any clothes? Meanwhile, where the rest of us live customers are more pragmatic, more discerning, and want the best value for their hard earned and (these days) scarce cash.
--"Whilst you're busy writhing, let's get to the fun bit - show me the money"-- and --"So your argument that IBM made massively more money than hp simply because IBM sold more units overall is likely to be very wrong due to IBM not making as much money on each server sold, either through margin at the time of purchase or pull-through during its lifetime. This is the prime reason Sun went steeply into decline whilst still shifting a high number of (low-end) units."--
What you percieve as wriggling is me dancing all over your arguments. Let me just point out the magnitude of your mathematical folly using your own numbers. If you make 38% more per sale than I do (and lets assume that's for everything) but I make 100% more sales than you (as per the article you quote) who comes out on top? Let's keep this simple and use 100 as a base line: 100+38=138 .vs. 100+100=200. Oh, dear. You also pointed out that you budget 5 x the purchase cost for the lifetime of the servers, so lets do a simple test there too: 138*5=690 .vs. 200*5=1000. Oh dear oh dear, you loose out again too.
--"hp Integrity has won that high-end despite the Itanium delays and despite the latest versions only being dual-core to Power's four-core offerings"--
Dude, Power 7 has 8 cores, not 4. I thought you would have known that. Well lets just wait and see how the benchmark results stack up when Tukwilla eventually makes it into some real systems.
As a developer I absolutely loath having to use Sparc Slowaris machines. About like trying to doing RDC to a windows 2000 machine running on a single core Celeron II. So glad the shop I doing work for moving to x86. As with any of the vendor specific UNIXs (Solaris, AIX, HPUX, etc) though they are all damn near intolerable to work with for any amount of time without installing the GNU tools though (no recursive grep, no bash, wtf is it 1990?).
Well, we can congratulate Intel on swallowing its pride twice, first in adopting AMD's x86-64 as their EM64T, and now in adding the RAS features offered with the Itanium to its latest Xeon 7500 chips.
This lets people have the features they want without giving up their investment in software. So I would rather congratulate Intel on its courage than gloat over the possible demise of the Itanium.
>The pincer attack of Itanium in the high-end and Xeon from below has already killed Sun, why should Intel or hp let up the pressure on Fudgeitso and IBM when the tactics are working so well?
I can buy most of what you say but call bullcrap on Itanium killing Sun. Suns own managerial incompetence killed Sun. And has been stated elsewhere the only thing Itanium ever did right was initially cause FUD in the MIPS crowd about going head to head with deep pockets Intel. Not enough decision makers had the foresight of the engineers and knew Merced would show Itanium for what it really was, nothing but a well covered IP grab by Intel for HP few remaining DEC valuables. Funny how many of the whizbang features are showing up in Xeons years before IA64. Besides Sun was never that big of threat to HP. HP's biggest threat is probably other ink makers, as after all HP is the world's highest tech ink seller.
Man, you take your responsibility as HP-Evangelist-In-Residence really serious.... You should relax a little.
a) SPARC is crap. Solaris/x86 is tolerable. You are right on that.
b) Power is the fastest CPU around (of *all* CPUs)
c) Itanium is dead in the water and SPARC's sinking doesn't change that.
d) I am not working for IBM or any other hw or sw vendor.
e) x86 Methusalem and Power 7 will duke it out.
According to spec.org, a single core of Itanium can achieve a rate of 17 CINT2006(base) at max. Power does 25 CINT2006(base) per Core. Which is completely logical, considering the clock difference of 1.8GHz vs 4.7 GHz.
Actually the EPIC compiler seems to be pretty good, as I would expect a difference of about 2.6 from the clock rates. According to spec, it is just 1.4. Still a tough sell to explain 40 % better single-thread performance to customers.
"....You should relax a little...." You obviously can't hear me laughing from over there. I enjoy a good game of whack-a-troll, and you IBMers seem to have taken over from teh Sunshiners in the rabidity stakes. At least I can salute your fervour, if not your nouse.
"....Facts are...." Why do I sense an imminent barrage of poorly though-out opinions presented as "facts"?
"....a) SPARC is crap. Solaris/x86 is tolerable. You are right on that...." Actually, I don't think much of Slowaris on x64 compared to RHEL or SLES. Try reading my posts before assuming you now what my prefs are.
"....b) Power is the fastest CPU around (of *all* CPUs)...." What, even including lab specials? Even those nanotube jobs? And how are you measuring fastest? Try and set the right boundaries for such a statement, such as saying "in my opinion, Power is the fastest CPU for commercial UNIX when running an application such as DB2". Please also try and give some supporting evidence (if you have any), otherwise you sound like you're evangelising rather than discussing, and the House of Sunshine will get upset at you blaspheming their Faith with your own! By the way, the fastest I've ever seen DB2 run was on a Superdome, but not as fast as it ran Oracle.
"....c) Itanium is dead in the water...." <Yawn> Another IBM troll that needs to read some marketshare reports. Please share the reading with Allison, she probably needs a hand.
".....and SPARC's sinking doesn't change that...." A pretty myopic view of what represents the largest migration opportunity available to both IBM and hp. Or could it be that IBmers just don't want to consider it since the EDS acquisition puts hp in the heart of the majority of those ex-Sun accounts?
"....e) x86 Methusalem and Power 7 will duke it out...." Careful of the religeous overtones, you'll be mistaken for a Sunshiner! Besides, I thought you IBMers thought BillG was the horned dude?
"....d) I am not working for IBM or any other hw or sw vendor...." Actually I can accept that fact given that it is hard to believe even a recession-hit IBM's standards have fallen that far!
".....According to spec.org....." Shouldn't that read as "According to my one-point feature sell based on a benchmark that tests only one tiny portion of the overall solution...."? Face it, I don't know any business that runs SPECINT for a living, and I'm betting you don't either. Until I test both in my environment, with my stack and apps, my LAN and my SAN, you're just blowing smoke. I would advise any customer thinking of buyiing Power7 or Tukzilla kit to do the same - test it at the vendors' expense. If they won't put up then it tells you all you need to know.
"....Still a tough sell to explain 40 % better single-thread performance to customers." Oh, so now single-threaded performance (on what test, using what data and appication?) is suddenly the key criteria for ALL computing requirements? Strangely, despite IBM's desperate performance feature sells on Power5, Power6 and Power6+, customers just didn't believe them - I wonder why? Instead, many of them tested the options and chose Itanium in hp Integrity. If what you claimed was true then hp would never have sold a single Itanium system (and we all know even IBM shifted 10,000+ Itanium units, even when trying thier hardest not to). Now, please tell me all those custoemrs were stupid, or just hated IBM, etc, etc, etc. Just to give us all a bigger laugh.
Everyone here can agree that the loss of Windows and the decline of HP/UX leave no growth markets for Itanium. It could take a decade, but the sunset of Itanium is certain. Now we have to see if the PC/Power discussion is another sinking ship. Desktop PCs may be on the wane, smart handhelds (and various thin-clients) are clearly encroaching on their market.. If you believe in cloud computing, user-configured servers are also on the wane.
We can only guess where this goes over time, but expecting x86-x64 PCs and Power servers to dominate to market would be ignoring these leading indicators.
A new poster? Someone a bit vague and wolly in their "arguments"? The trolls sense fresh meat!
"....Everyone here can agree that the loss of Windows and the decline of HP/UX leave no growth markets for Itanium..." Oh dear, you really didn't think that statement through, did you? Agreement in a trollathon!?! Ignoring the existing hp-ux, OpenVMS and Linux on IA64 installed base that will consider the new Tukzilla kit ('cos it's always easier to upgrade the hardware than change to a new hardware, change the OS or app stack), you're forgetting all those SPARC customers that have been left little choice. Whilst Slowaris on x64 will do for some (and, going on x64 marketshare, more likley on hp ProLiant than any other x64 vendor's offerings), those needing large SMP instances really only have a choice of pSeries or Integrity. And EDS puts hp in the front running to convert those accounts to hp Inetgrity, hence a nice possibility for Itanium growth.
"....It could take a decade, but the sunset of Itanium is certain...." That's in the same logic class as saying "we're all going to die someday even if we get really old in the meantime". All chips will be retired eventually (yes, IBMers, even Power will go the way of the dodo some day). If Itanium does get retired in ten years time then it will be over twenty years old - not bad for a commercial CPU. But your reasoning as to it dying in ten years is a bit weak, would you like to actually supply some form of argument to support your opinion, or can we safely say it is just another unsubstatiated musing?
"....Desktop PCs may be on the wane, smart handhelds (and various thin-clients) are clearly encroaching on their market..." How does the future of desktop PCs impact large commercial server solutions? A bit off-target, aren't we? Is the presumption that low-end devices will alter IBM's and Intel's plans? Not really - after all, IBM can fab the new low-end chips, and Intel actually design and fab their own, so not really a major impact for either.
"....If you believe in cloud computing, user-configured servers are also on the wane....." Good thing I didn't swallow the cloud hype, then, isn't it? Cloud is just ousourcing, and there are always bits of the business (for large corporates at least) you just won't ever outsource.
/SP&L, just not meat in that post to chew on.
>Intel is going to win one way or the other, with AMD happy to ride along on their coattails. The pincer attack of Itanium in the high-end and Xeon from below has already killed Sun, why should Intel or hp let up the pressure on Fudgeitso and IBM when the tactics are working so well?
For the record the amount of pressure HP with the Itanium is actually putting on IBM is near zero. x86 perhaps more and probably most in services. My impression is that IBM doesn't really want to talk to you on service side unless your talking multimillion dollar. HP is cheaper and most of their consultants I have worked with tend to be very knowledgeable but a bit shifty characters. Nobody ever got fired for selecting IBM but count on the IBM parasite being around awhile. HP more fly by night and have to be careful around them but less likely to have to pay IT tax to them every year.
Just to show I am not shill for any of these companys I will cut them all down. Intel will own Power over the long term unfortunately. HP actually did help this not by putting out the disasterous distraction that is the Itanium but due to giving Intel their IP on designing enterprise chips. Intel doesn't exacly always have its pulse on customers needs and trends (see power hungry P4 disaster when oil went over $100 a barrel) but it does have more money than God and that can overcome many problems (see Core microprocessor savior in record time). It will simply utilize its scale of economy as each generation gets more expensive to push IBM out. IBM has deep pockets too but has a lot less interest in risking it trying to keep up with the Jones (IBM happy to join you if can sell their precious services). In fact I dare say the only reason Power is keeping up so far has to do not with super high end business machines (where margin great and all but volume also super low) but by being able to get a monopoly on this generation of gaming consoles. As far as I can see after losing Apple etc IBMs only hope of keeping up is to continue getting most of the gaming console contracts. Unfortunately after the minor debacle that was the Cell this might not be realistic in a few years time.
"...Oh dear, didn't Ms Kebabfart get the memo about trying to raise the industry's perception of women in computing, not blow it with poor quality FUDfests? XD..." What do you mean with this? I dont get it. Please explain.
BTW, when you talk about "Slowaris" do you have any proof of that? Here are SAP official benchmarks showing that Solaris is faster than Linux. Solaris used a slower cpu, and used slower RAM sticks and still was faster on SAP benches with 48 cores. So I dont get it when you say "Slowaris"? Are you just FUDing, just like all the rest of the IBMers here?
Argh, web site is down. But here is a link confirming Solaris being faster than Linux. If someone wishes, I can post links to official SAP benchmarks showing that Solaris is faster using slower hardware (on 48 cores)
Regarding POWER7, yes it is fastest, maybe 30% faster than Intel Nehalem-EX. And POWER7 is much more expensive. So I dont see why you should pay, maybe 2-3x or more, for as little as 30% extra performance? Sure, POWER7 has better RAS than x86, but I hope HP ports OpenVMS to x86 to take advantage of cheap fast CPUs. OpenVMS is arguably the most stable OS out there, almost on par with Mainframes. I bet OpenVMS on x86 will have better uptime and less problems than AIX on POWER7.
OpenVMS cluster tech is legendary. Even during attacks 9/11, no single transaction was missed on the OpenVMS systems running financial systems in the twin towers.
I have heard that HP-UX is one of the most stable Unix out there (but barebone) but I dont have any experience of HP-UX.
"Regarding POWER7, yes it is fastest, maybe 30% faster than Intel Nehalem-EX. And POWER7 is much more expensive. "
As I wrote to you in another thread, as the socket count goes up so does the difference between POWER7 and Nehalem EX. So yes at 1 socket it is pretty close, 2 sockets it is still pretty close, at 4 sockets POWER7 takes a more serious lead.. and at 8 Socket POWER7 annihilates Nehalem.
Have a look at the newest SAP 2 Tier benchmarks.
Fujitsu PRIMEQUEST 1800E, 8 proc / 64 cores Nehalem-EX@2.26GHz does 87550 SAPS.
(10943 SAPS/ Socket)
IBM Power System 780, 8 processors / 64 cores POWER7@3,86GHz does 202180 SAPS.
Now that is not 30% difference.. that is a 131% difference.
Haven't you got a hat (hopefully chocolate) somewhere you have to eat :)=
"....What do you mean with this? I dont get it. Please explain....." Sorry, Kebby, but after Allison's last FUDfest I took the liberty of christening her Ms Kebbabfart. Apologies if you thought the reply was directed at you, or if you thouhgt I was implying some sort of real-World relationship between yourself and the deluded Ms Park. I apologise unreservedly for any implication that your tastes would have such a low threshold of satisfaction as to be sated by the vacuous Ms Park.
"....BTW, when you talk about "Slowaris" do you have any proof of that?..." Been there, proved that. Do you really want reminding of SAP bench number 2008064? Where the hp ProLiant DL785 G5 8-socket posted 35400 SAPS, compared to bench session 2008061, where the similar Sun Fire X4600M2 8-socket only managed 29670 SAPS. Yes, I know the Sunshiners will start whining about how the hp server had the 2.7GHz 8384 4-core Opterons, whereas Sun's server had to make do with the 2.5GHz 8360 4-core Opterons, but it's strange that 8% jump in CPU clock gave a 20% jump in performance. The differentiator - the Sun box was struggling with Slowaris, whereas the DL was humming along with SuSE Linux.
"....Even during attacks 9/11, no single transaction was missed on the OpenVMS systems running financial systems in the twin towers...." One of the reasons many companies saw a DR failure during the 9/11 attack was they had their production systems in one Tower and their DR in the other. Anyway, I'm not sure it's really fitting to be bragging about DR successes from such an awful event. It's much more fun to go watch the video on the hp webby where they blow up a datacenter and time the failovers for OpenVMS, hp-ux and Windows! http://h71028.www7.hp.com/enterprise/us/en/solutions/storage-disaster-proof-solutions.html
As regards the comparison of Power vs Nehalem, you forgot to include the eye-watering expense of having IBM Global Scewups coming onsite to fix the "solution" that IBM claimed woul perform to requirments "out-of-the-box". My advice is always POC and bench at the vendor's expense! Getting IBM GS to do it afterwards is just a whole world of expensive pain.
/SP&L @ Ms Kebabfart.
Ouch I just heard that the Superdome2 will not expand beyond the one chassis this year which means it is only a 16 socket box. Two cell board based chassis' can be hooked together next year for 32 sockets and the 64 socket box does not even have a date.
Looks like HP is getting out of the high end space
".....I just heard...." You just heard from where exactly? The voices in your head again? Unless you can post at least a source if not a confirmable hp document I'm going to assume that all you heard was your own wishful thinking.
"....Two cell board based chassis' can be hooked together next year for 32 sockets...." Hmmm, so that would be a 128-core SMP server, then, with real enterprise (and not weiner) cores linked in scaleable and high-memory-bandwidth solution. Ooh, sounds awful - not! Seeing as the current Superdome offering scales to 128 cores what your saying is it is terrible news that hp will be able to offer a solution of the same size that uses faster cores, more and faster memory, all linked by faster interconnects, and in a much smaller footprint, and that they will be able to double it later? I can't wait to see what the equivalent pSeries looks like if that disappoints you! Will it come with a built-in coffee machine and poptart toaster?
By the way, when will 128-core pSeries be hitting the streets? Doesn't the current Power595 stop at 64 cores? I hear (from the Reg article http://www.theregister.co.uk/2010/02/08/ibm_power7_systems_launch/ <= please note the checkable source, not wishful thinking) thet the IBM big iron is not arriving until 2010 too. So even if what you "hear"/imagine does turn out to be even remotely true, the reality is IBM will have no advantage. Can I claim IBM are "getting out of the hihg end space" too seeing as they can't even match today's Superdome for socket scaling?
Please, can the IBM trolls try posting something that can't be shown up as sheer bunk in less than five minutes? It's hardly worth the effort of opening you to ridicule if you're not going to at least make a proper effort. I'm almost missing the days of the good old Sun FUD, which was at least a bit more inventive.
Ok, then I know. Please be clearer when you are, or not adressing me, because I am not a native english speaker. I sometimes miss the subtleties. It is not always clear to me. Jokes and sarcasm are the easiest things to miss for a foreigner.
Regarding your SAP benchmarks . You claim that Solaris is slower than Linux.
1. The Solaris SAP setup used Unicode.
"According to SAP a Unicode system needs 10-30% more resources to support the same number of users."
2. The Solaris machine used a previous generation AMD Barcelona cpu, whereas the Linux machine used a newer generation AMD Shanghai cpu - which is faster. That explains the 20% jump, for only 8% clock increase.
"The Barcelona-based system was able to yield 29,670 SAPS, the Shanghai-based system was measured with 39,270 SAPS. At 108% the clock rate the new processor generation was able reach a 132,3 percent higher benchmarking result."
3. Solaris had better CPU utilization than Linux on 32 cores. 93% vs 90%.
Here are new SAP benchmarks from last year showing that Solaris is faster, despite using slower clocked cpu, and slower RAM sticks, than the Linux machine. No Solaris unicode in this setup. On a higher core count, 48 cores, Solaris has 99% cpu utilization and Linux has 87% cpu utilization. Some say this is because Linux doesnt scale as good on many cores. (Sure, Linux scales good in a cluster, but that is basically a network with a bunch of PCs - a different thing). It seems that Solaris is faster, and scales better than Linux to me if we look at these new SAP benchmarks? So I must protest when you name it "Slowaris". Solaris is not slower, but it scales better and is faster than Linux. But who doesnt?
I agree that POWER7 scales better than Intel Nehalem-EX. We have discussed this earlier. But you also admitted that an 64 socket POWER7 would probably not be faster than an 256 socket Nehalem-EX machine. And maybe there would only be a 32 socket POWER7 as highest?
Just throw more and more cheaper Intel Nehalem-EX, and sooner or later POWER7 will not catch up. Anyway, such a big 256 socket Nehalem-EX will not be common. But the main trend is that x86 is catching up on POWER and the rest of the RISC cpus. Wait until Nehalem-EX arrives in 32nm and higher speeds. And then the next generation x86 server CPUs will arrive much earlier than POWER8. Do you agree that POWER will lag behind at some point? x86 develops much faster than any other CPU architecture right now. The curve is steeper. I forecast that IBM must power AIX to x86 sooner or later. Solaris is running on x86, on Niagara and on SPARC. No matter who wins, Solaris is there. Future proof. All OSes must be ported to x86. Sun ported their big Unix first. The rest must follow.
First to your comments to Matt.
(I haven't verified your data and links but... )
Damn, what has happened to you ? Serious argumentation and links to build your argument. I tip my hat, and welcome to serious discussions :)= I am so proud.
And now for my part.
Big 256 socket Nehalem-EX boxes will most likely only be made by SGI. They won't be made for commercial performance. Those will be HPC systems. And sure there will be specint_rate2006 measurements for those, and perhaps also some SPECJBB benchmarks, but there won't really be many customers for such systems. They will be SSI.
They will also scale quite terrible, if you use them for workloads that have significant cross node communication, like for example a large Oracle database. These systems are not made for such workloads.
I have still only read about the 775 having Infiniband with the capability of doing something similar to an SGI, hence creating a SSI machine with an enormous amount of sockets, here on TheRegister. So I guess we will wait to see on that one.
If you read my comments carefully you will see that I have never claimed that a 64 socket POWER7 system will beat a 256 socket Nehalem-EX system :)=
But as we see that a 8 socket POWER7 system already is beating a 8 socket Nehalem-EX system with a factor of 2.3, on commercial workloads, then who knows what will happen at the higher socket counts. if then for example at 16 sockets POWER7 is 4 times faster then.....
For commercial workloads I think the biggest POWER system will be a 795? or whatever it is going to be called with 32 sockets, so you are quite right that there will be no 64 way commercial system.
As for Nehalem-EX on 32nm.. or rather Westmere-EX as I imagine it will be called. That will take some time.. I wouldn't be surprised if it would only beat POWER7+ with a few month. And sure more cores are nice.. but you also have to have the infrastructure in the server to match.
The x86 sweetspot is still 2 socket servers, with a dash of 4 socket. This leaves the 4+ to RISC and Itanium. And of those POWER7 is king.
And you whole rant about x86 taking over the word.. well we have heard that now for 10 years.. and it still hasn't happened. And when I retire in what 15-25 years depending on how much the stock marked goes up and down. (up and down is good when you have funds to invest), then I'll still be able to do architecture on POWER systems... and as there have become fewer people who can do that.. I'll make more money !
Jesper, you should be very careful of offering vendor benchmarks as some sort of proof of superior design. We all remember the laughable IBM Power6 record TPC benchmark where the real work was being done by $4m (10992 disks!!!) of very-short-stroked storage running a completely unrealistic and highly-partitioned database layout. After all, Kebby gave you a kicking before when you IBMers were getting uppitty about the Soreacle TPC record benchmark which dethroned IBM's Power6 one. For those that need reminding, Kebby posted the following link to an analysis of the record TPC results at http://www.c0t0d0s0.org/archives/5941-The-new-TPC-H-benchmark-from-IBM.html.
Most vendor benchmarks sessions bear no resemblence whatsoever to what customers will actually be doing with the kit in the real World. It made it all the more embarrassing when Power6 didn't perform to even half IBM's advertised figures in our production environment.
And for those that don't remember, when Power6 came out, IBM was making mucho, mucho noise about its amazing performance on a select few and carefully crafted IBM bench sessions, but at the same time IBM's own best rPerf figures were showing that P6 had only gained 41% on P5 despite the core frequency having doubled. We're seeing the same again - IBM has boosted core frequency and wants to talk about nothing else, but hasn't done the necessary to make the overall solution anywhere near as much of an advance as they want to make out. Don't be fooled - if IBM say they can make something go twice as fast on P7 then make them prove it before you place an order! Remember that none of IBM's carefully crafted bench sessions are anything like what you will be doing in reality.
/SP&L, especially at the irony of using Kebby's Soreacle defence to beat up on Jesper!
"when Power6 came out, IBM was making mucho, mucho noise about its amazing performance on a select few and carefully crafted IBM bench sessions"
What are you talking about? Power6 systems have a complete set of benchmarks on just about all industry standard benchmarks - in fact, it has released more benchmarks than your beloved HP Itanium machines. You can review all the benchmarks at your leisure. If you do, you will see that the Power6 performs much better than the HP Itanium boxes. It doesn't matter how they did it. i.e. higher frequency - the fact remains that they are much faster.
"Don't be fooled - if IBM say they can make something go twice as fast on P7 then make them prove it before you place an order! Remember that none of IBM's carefully crafted bench sessions are anything like what you will be doing in reality."
Well proving performance is applicable to any vendor. All vendors will optimize until they get the best benchmark score they can to publish it. IBM has already released a wide set of industry standard benchmarks for their recnelty released Power7 boxes. Where are the benchmark of your beloved HP Tukwila boxes? The only one release so far (specint_rate/fp_rate) shows that POWER7 is >4x faster for both int and fp per socket than Tukwila. Now that is embarrassing for HP. I'm sure all future benchmark release will also be embarrassing as Tukwila is a dog of a chip in comparison.
"Jesper, you should be very careful of offering vendor benchmarks as some sort of proof of superior design. "
Well Now Benchmarks are nice, but they are no substitute for a RL benchmark, where you get to actually test the solution and see if it holds water.
But when that is said you have to remember that almost all of those nice sizing tools and metrics that are used to size actual customer workloads, actually comes from these same benchmark. rPerf for POWER for one example, SAPS or server vendor neutral RPE2 from Ideas International is also based upon Benchmark data. So perhaps you think about that next time you use a sizing tool. If you do that in your line of job.
"We all remember the laughable IBM Power6 record TPC benchmark where the real work was being done by $4m (10992 disks!!!) of very-short-stroked storage running a completely unrealistic and highly-partitioned
"We all remember the laughable IBM Power6 record TPC benchmark where the real work was being done by $4m (10992 disks!!!) of very-short-stroked storage running a completely unrealistic and highly-partitioned database layout. "
Well, this is what it takes to run these large benchmarks. The number of disks is surely redicolous. But hey the POWER6 595 benchmark isn't the worst, sure it used 10992 disks to get to 6M TPMC, but hey the HP SuperDome used 7762 disks to get to 4M TPMC. Now that is 1897 Disks per million TPMC for the SD versus 1806 for the POWER 595. Sure in Real life you would be using a Disk system with a lot of cache etc. But the sheer fact that these big TIN boxes can use that many disks is and sustain these IO levels, is partly due to these benchmarks. They are pushing the limits for what these systems can do. Now if you want to take a shoot at the nature of the tpc-c Benchmark that is fine for me, but taking a cheap shoot, at the POWER 595, for just being the fastest single system the doesn't give you credit. Sorry.
As for the "highly-partitioned database layout", have you ever read the tuning whitepaper where the HP SD team describes how they got to 4M TPMC, and the techniques used ? Unfortunately I can't find the link right now, or perhaps it is no longer on their website. It shows how important partitioning is for a machine with such bad cross board communication as the SD has. Naah.. HP should have keept Alpha, the Alpha 1280 still has more than double the STREAM memory bandwidth than the current Montvale based SD. But again the benchmark did give us a good whitepaper on how to tune large Database on a SD (we have one that runs on 50+ Montecito cores here) so that is how I know that the paper could actually be used in RL.
"After all, Kebby gave you a kicking before when you IBMers were getting uppitty about the Soreacle TPC record benchmark which dethroned IBM's Power6 one. For those that need reminding, Kebby posted the following link to an analysis of the record TPC results at http://www.c0t0d0s0.org/archives/5941-The-new-TPC-H-benchmark-from-IBM.html."
Well now I have the luxery to have actually worked with POWER for many years. And I find the guy at c0t0d0s0 quite a knowledgebale guy. But he's knowledge of POWER systems is very limited, and mostly seem to come from SUN. This is ofcause a problem with people who work as consultants or whatever, and only touch one platform. Now it's a bit different for us who actually handle multi vendor solutions every day and has to know both how to size Oracle on POWER as well as Websphere on Itanium, or whatever. But I had a good discussion with him around POWER7, where he started out with quite a few misunderstandings. And our good friend Keb was quick to say that I got trashed, but as the discussion ended. KEB did conseed that I was the one who did the trashing, in a polite and civilized manner.
I haven't read his TPC-H analysis, cause to be honest.. after my last discussion with him, I would consider it a waste of time. I know FUD when I see it.
"Most vendor benchmarks sessions bear no resemblence whatsoever to what customers will actually be doing with the kit in the real World. It made it all the more embarrassing when Power6 didn't perform to even half IBM's advertised figures in our production environment."
Sure the systems don't deliver what the same number of transactions in RL. Your transactions are normally a fair bit heavier, you don't have the same IO bottleneck, you don't have people with the right skills etc etc. If you don't know that.. then well.. At least POWER6 has the frequency to run RSC (real stupid code) fairly well as the sheer frequency and enomours memory bandwidth makes the system pretty forgiving. Not like some other architectures.
"And for those that don't remember, when Power6 came out, IBM was making mucho, mucho noise about its amazing performance on a select few and carefully crafted IBM bench sessions, but at the same time IBM's own best rPerf figures were showing that P6 had only gained 41% on P5 despite the core frequency having doubled. "
As for frequency you do know the difference between POWER5 and POWER6 right ? You also know that rperf numbers change over time, as the benchmark numbers change ? Now for example something like DFP, wasn't really supported on many software stacks as POWER6 came out, so that is one of the reasons why for example SPECJBB and SAP improved as this was supported by the SW stacks. This is also true for most other platforms. For example compilers for Itanium keeps getting better and better, hence if you update your software stack your applications will run faster. Well, if it wasn't for the fact that more and more crap gets put into those stacks... :)
"We're seeing the same again - IBM has boosted core frequency and wants to talk about nothing else, but hasn't done the necessary to make the overall solution anywhere near as much of an advance as they want to make out. Don't be fooled - if IBM say they can make something go twice as fast on P7 then make them prove it before you place an order! Remember that none of IBM's carefully crafted bench sessions are anything like what you will be doing in reality."
Well I have ordered our first POWER7 machines, so that we can get to play a bit with them and get to know the flight envelope of these babies. And to be honest I am not nervous, IMHO POWER normally delivers what it promises, they also tend to deliver their platforms somewhat on time. Which makes it much much easier for me to do upgrade planning, for our customers. That can't really be said about neither SUN or HP's UNIX platforms. Not that that I haven't been disapointed, I did some software development tests on AIX 5.0 on Merced back in what.. late 2000 early 2001, now there was crap for you, the damn worst platform I've ever tried, bloody compling Hello world tok forever.
And again, carefully crafted benchmarks sessions is not something specific to IBM. Empty marketing drums will always be empty marketing drums. Be it IBM, HP or SUN/Oracle. I normally find HP and IBM more clueless than mischivious. But I must admit I absolutely detest SUN/Oracle, perhaps it is mostly due to their religious fanatism, which in many cases go beyond reason. If you've ever read BMSeer blog then you know how.
/SP&L, especially at the irony of using Kebby's Soreacle defence to beat up on Jesper!
Well there is beating up and then there is taking cheap shots that doesn't really hold water. Sorry, if I did expect more from you, as You do look like a guy that at least partly know what he is talking about.
"...Damn, what has happened to you ? Serious argumentation and links to build your argument. I tip my hat, and welcome to serious discussions :)= I am so proud...." Jesus, Jesper, this leaves a bad taste in my mouth. :o(
I have ALWAYS asked for links, and I have ALWAYS shown links that supports my claims. A mathematician always asks for, and gives proofs (white papers, benchmarks, etc). And I have repeatedly stated that I accept benchmarks, I have no problems to accept a proof, even if it is unfavourable to me.
You, on the other hand, have repeatedly rejected MY benchmarks and white papers as "no, they are cherry picked by Sun", "they are crafted by Sun", and "no, even though you need four POWER6 to match two Intel Nehalem, the POWER6 is faster" and similar weird non serious argumentation, like "because POWER6 core is faster than Niagara core, then the entire POWER6 cpu must be faster - therefore I reject your benchmarks", etc etc etc. I would definitely call this to be not serious argumentation from you.
I, on the other hand, have NO PROBLEM accepting POWER7 benchmarks, nor do I reject them and call them "cherry picked by IBM", etc etc. In fact, I said that earlier: "I accept benchmarks even if they are unfavourable to me".
Now it seems that you try to make me look like I at long last, finally accept your benchmarks that you show??? I mean, you always rejected every benchmark of mine, and I always accepted your benchmarks! So, no, I dont like this. You always had problems with my links, I never had problems with your links. I also had problems with your weird claims such as "POWER6 is faster than Nehalem" and such. Such reasoning is highly non professional, highly non serious.
So do you claim that I used to reject your benchmarks, and I used non serious argumentation and weird logic when discussing earlier??? Prove it. Show me links where I do that. Otherwise, please dont try to make this look like something else. It wasnt I who used weird logic and non serious argumenation like "no, POWER6 are faster than Nehalem". It was you. I can prove my it.
At the place where I currently work we have a pretty big OpenVMS installation. It's doing a rather critical part of financail transactions here. Maybe the biggest installation in Europe, actually. The problem is that the cluster can't be made bigger any more.
Is HP not interested in *really* moving OpenVMS into the future ?
(the institution is currently in the process of migrating to Linux..)
jlocke wrote: "Is HP not interested in *really* moving OpenVMS into the future ?"
Not really - they got that from DEC/Compaq - and seem to only see it as a chance to "upgrade" folks to HP/UX.
From a biz perspective I'd love to see OpenVMS on Xeon boxes, from a personal perspective then open source it as well.
Biting the hand that feeds IT © 1998–2021