IA64 and emulator performance
|
From: Rick Jones on 18 Oct 2006 14:42 Ketil Malde <ketil+news(a)ii.uib.no> wrote: > Rick Jones <rick.jones2(a)hp.com> writes: >> *) "x86" under Linux > (And under Windows, surely?) I would think so, but even via netperf, my exposure to Windows is limited. > Anyway, I apologize for not making myself clear enough. The problem > is that I have a program compiled for x86 (on a P4), and a potential > user of this program with an IA64 box. One particular test runs in > about ten seconds on my machine, but a similar test takes about > thirty minutes on his. "Similar" or "identical?" And is this the old HW-based emulation, or the newer SW-based? What sort of work does this program do? > I don't have access to IA64 hardware Ah, but you can :) http://www.testdrive.hp.com/ rick jones -- a wide gulf separates "what if" from "if only" these opinions are mine, all mine; HP might not want them anyway... :) feel free to post, OR email to rick.jones2 in hp.com but NOT BOTH...
From: John Dallman on 18 Oct 2006 18:22 In article <1161157626.465073.266220(a)b28g2000cwb.googlegroups.com>, robertwessel2(a)yahoo.com () wrote: > A 180-fold penalty is *not* common. Self-modifying code, dynamically > generated code, and/or code and data intermixed on a page might lead > to that. I'd agree that 180-fold is extreme, but lots of I/O or other system calls can make performance rather worse than the best cases. At least, that was what we found when we started running the IA-32 to IA-64 cross-compilers on an IA-64 machine. It may have got better since 2003, when we decided that we were going to phase out IA-64 support and hence stopped trying to improve it. --- John Dallman jgd(a)cix.co.uk "Any sufficiently advanced technology is indistinguishable from a well-rigged demo"
From: jsavard on 20 Oct 2006 18:40 Rick Jones wrote: > IIRC there are at least three architectures for which there are > emulators on IA64: > > *) "x86" under Linux > *) PA-RISC under HP-UX > *) SPARC under something from Fujitsu I found out about a *fourth* one thanks to a sponsored link from Intel on Google Groups. A company emulates the z/Architecture - the 64-bit version of the venerable IBM 360 - on the Itanium and gets respectable performance. This is supposed to mean that the era of proprietary architectures is over! Although I thought Intel had patented EPIC, so I am puzzled at the claim that the IA-64 architecture is not proprietary. Although many people seem to shy bricks at the Itanium's current market success (or lack thereof) it seems to me that the Itanium is the closest thing to a mass-market supercomputer chip there is. Which may not be saying much, of course... but it definitely has at least a few features the Pentium lacks, and it must be justifying its price with some degree of increased performance over common garden Pentium IV chips? John Savard
From: Bill Todd on 20 Oct 2006 21:19 jsavard(a)ecn.ab.ca wrote: > Rick Jones wrote: >> IIRC there are at least three architectures for which there are >> emulators on IA64: >> >> *) "x86" under Linux >> *) PA-RISC under HP-UX >> *) SPARC under something from Fujitsu > > I found out about a *fourth* one thanks to a sponsored link from Intel > on Google Groups. > > A company emulates the z/Architecture - the 64-bit version of the > venerable IBM 360 - on the Itanium and gets respectable performance. > > This is supposed to mean that the era of proprietary architectures is > over! Although I thought Intel had patented EPIC, so I am puzzled at > the claim that the IA-64 architecture is not proprietary. Although many > people seem to shy bricks at the Itanium's current market success (or > lack thereof) it seems to me that the Itanium is the closest thing to a > mass-market supercomputer chip there is. Which may not be saying much, > of course... but it definitely has at least a few features the Pentium > lacks, and it must be justifying its price with some degree of > increased performance over common garden Pentium IV chips? It enjoys some performance advantage over PIV in some benchmarks (especially SPECfp), though not much in most commercial ones when compared with IBM's (up to 32-socket) and/or Unisys's (up to 16-socket?) large Xeon systems (save where Itanic system's higher socket counts let it pull ahead in absolute throughput). But 'Core2Duo' is miles ahead of it in SPECint and not far behind in SPECfp (AFAIK no high-socket-count Core2Duo systems yet exist). - bill
From: John Dallman on 21 Oct 2006 10:55
In article <1161384017.296321.183600(a)k70g2000cwa.googlegroups.com>, jsavard(a)ecn.ab.ca () wrote: > This is supposed to mean that the era of proprietary architectures is > over! Although I thought Intel had patented EPIC, so I am puzzled at > the claim that the IA-64 architecture is not proprietary. You aren't the only one. The claim may be based in the fact that IA-64 processors are on general sale to anyone who wants them. However, POWER4 was formerly used by Apple, as the "G5", and the SPARC architecture definitions were generally available, as were chips, if you wanted them. I tend to read the "IA-64 is not proprietary" claim as that special category of truth known as "marketing". > it seems to me that the Itanium is the closest thing to a mass-market > supercomputer chip there is. Absolutely. It's serving as a demonstration of the "supercomputing fallacy", the belief that almost everyone has problems that are amenable to the techniques of the category of machines that have been known as "supercomputers". Such problems have historically been large-scale simulations, where the market for solutions was willing to go to great lengths to coerce their problems into forms suitable for special-purpose hardware, because they needed solutions really badly. Unfortunately for IA-64, while there may be more people with such problems than have access to "supercomputers", there is a large set of people who have problems that they can solve on general-purpose computers, slowly, and don't feel it worth ("find it extremely hard") coercing into a form highly suitable for IA-64. I work for one of them, a large CAD/CAM/CAE ISV. We support the HP-UX Itanium servers as large-scale general-purpose servers, for data management applications. They work, but they don't seem to have any amazing virtues for such jobs. However, Itanium proved to have minimal benefits for desktop CAD, as demonstrated by HP's withdrawal of all their Itanium workstations. > Which may not be saying much, of course... but it definitely has > at least a few features the Pentium lacks, and it must be justifying > its price with some degree of increased performance over common > garden Pentium IV chips? Oh, yes. But the justifications are mainly psychological, not technical. The people who are buying Itanium systems fall into several distinct groups: * People with supercomputing problems that fit really well onto Itanium, who are prepared for the idea that Itanium may go away in a few years. That seems to be OK for "real supercomputing" people, who are used to reworking their codes for different hardware. * People who have taken on HP's line that it's the natural successor to PA-RISC, and feel that it will be better for them than switching vendor. * People who simply need, for political or psychological reasons, to be using the large, expensive, and very "corporate" computer. They would have bought System/Z in the nineties, rather than large Suns. The last two groups don't tend to believe that there is a real risk of Itanium disappearing, largely because they believe the statements to that effect from HP and Intel executives. The problem is that (a) Intel is now in complete control, because HP sold their IA-64 design teams and intellectual property to Intel, so HP's controls are purely contractual (b) Intel will, in the end, do whatever it takes to survive, and some point, they'll get fed up of loosing money on Itanium. Going back on things executives have said is comparatively easy for corporations; they just have to fire the executives. Shareholder lawsuits might be a different matter, but they're a question of management of perceptions and expectations. Intel claimed when they introduced EM64T that Itanium would be back, and that EM64T was just a short-term diversion while development happened. Nearly three years later, we have Core 2 Duo, and the Itanium development programme seems to be being stretched out, with nothing radical seeming to be in the offing. --- John Dallman jgd(a)cix.co.uk "Any sufficiently advanced technology is indistinguishable from a well-rigged demo" |