curious performance enhancement or punishment?
6 Oct 2008 2:35:23 UTC
Topic 193966
(moderation:
I reached a point where I needed to add 200K+ points(currently only 180K left to get) to my Einstein account, so I attached all my hosts (in a share with atleast one other project). After many days I noticed something. I really hadn't paid attention during the early days of this run of einstein.
I have:
2- AMD 9950 BE Quads
1- AMD 9600 BE Quad
1- Intel Q6600 Quad
1- AMD64 X2 6000 dual
1- AMD64 X2 5200 dual
1- AMD64 X2 4800 dual
1- AMD64 3700 single
1- Mobile AMD64 3700 single (32b windows only)
1- AMD64 2800 single
All running 64b linux and 64b boinc (except the one as annotated)
Anyway, On the other projects (most anyway), the AMD Quads seem to perform nearly as well as the Intel(sometimes better). But with Einstein, the Intel Q6600 seems to get 33% speed bonus, or the AMD's get a 50% penalty. I understand about the curves of the different freqs and haven't done in "in depth" analysis. It was sufficient to see that the Intel was taking in the 6 hour range, and the AMD's in the 9-10 hour range.
The result was that I couldn't justify leaving the AMD's getting work and have let them do other projects, while I have just the Intel running Einstein (at 100% share to offset the loss of the AMD's to some extent). I'd really like to be able to unleash the AMD's here to get my einstein credit up to my Simap credit and have both of them be "tier 1" at the highest credit projects I have. At the present rate with only the Intel, I won't reach the simap level before they have work again.
I've set up tiers based upon my perception of project worthiness, and am running projects up to the level I feel they should be at. (see signature and pie chart).
Anyway, is there something in the works to even up this??
tony
curious performance enhancement or punishment?
)
I suspect much of the Intel vs AMD variations depends on how memory-hungry an app is, AMD is very superior in RAM efficiency so in some projects they will perform equal to an Intel, even if the Intels are better processors.
Various benchmarks from hardware review sites leaves no doubt that an Intel SHOULD perform noticeable better than an AMD, and on the other hand no one can deny the fact that Intels memory system is antiquated and painfully slow.
Team Philippines
I have an AMD Opteron 1210
)
I have an AMD Opteron 1210 running Linux at 1.8 GHz and in Einstein it is about 10% slower than most Intel processors, while in SETI, running the optimized app it is faster than most Intel processors (quads excluded). In QMC the quorum is 1, in CPDN and CPDN beta the deadlines are faraway and LHC@home rarely has work. MY RAM is 2 GB, my L2 cache is 2 x 1024 MB.
Tullio
Hi Tony, The Einstein way to
)
Hi Tony,
The Einstein way to get the most out of your computers is to use Linux, which I know you have played with. It is reported to be 20-25% faster than the Windows app.
Andy
Just some quick
)
Just some quick observations:
1) Your Intel Quad shows a stunning performance based on WU runtime, comparable to the top hosts for this project (e.g. >= 3GHz Xeons). It appears it is either overclocked considerably or the runtime readings are somehow messed up and more optimistic than in reality.
2) Your AMD 9950 quads are in fact listed as running under Windows, not Linux as you indicated in your posting. The Windows app is known to be slower than the Linux app. at the moment. Work on improving the Windows app is under way.
3) Your AMD 9600 quad is terribly slow, even if compared to your other AMD CPUs. Does it have the TLA bug fix enabled which slows down the CPU?
This is really surprising to me, as the consensus is that AMD CPUs at the moment (with the possible exception of systems > 8 cores) is unable to compete with current Intel CPUs when compared core for core and clock cycle for clock cycle. That's why AMD loses a lot of money atm and Intel is earning money like crazy. Which are the projects where AMD can still stand its ground against Intel?
CU
Bikeman
RE: This is really
)
In SETI I have six pending results. Five of my wingmen have Intel CPUs, one an AMD CPU. So it is 5 to one.
Tullio
Hi, just crunshed my first
)
Hi,
just crunshed my first Einstein WU for a long time with my old single core K8 @2,6 GHz:
http://einsteinathome.org/workunit/43595962
It really needed ages ... but well it was Windows ...
Is there a quick explanation why Linux is faster ? I though you guys were already doing SSE handcoding, thus it cant be Complier related, can it ?
Edit:
The AMD problem is probably Cache related ... I just noticed that the progress bar jumped really fast up to ~13% and now it is going slooooow. Maybe the L2 Cache is used completely ?
thx
Opteron
RE: Hi, just crunshed my
)
Yes, but the handcoded SSE stuff was improved again near the end of S5R3, but that latest improvement didn't yet make it into the Windows built. This is work in progress. Also the Linux version has an executable that is built for SSE2, while the Windows app has only variants for SSE and non-SSE.
Let's not forget that K8 is really quite old an architecture by now, it competed (and very well indeed!!) against Pentium III and Pentium 4, but is no match for recent (say) Intel Core2 chips.
CU
Bikeman
The q6600 is on an Asus
)
The q6600 is on an Asus P5K-VM microATX board and is overclocked to 2.997 Ghz. The 9950's are OCed to 2.8, the 9600 to 2.63, and the rest of the pea patch to some modest OC. I found out Mandriva 2007 spring free wouldn't install on the 9950's so I installed a temp version of winxp pro 64 until I could get a linux version which recognized the Asus M3A78-T ATX board. Finally got Ubuntu 8.04.1 installed on them.
The 9600be and 6600 were purchased together, and have the exact same memory (ddr2 1066, corsair xms2), the same case, the same PSU (Enermax ELT 620 80+ eff), and the same cooler (Artic Cooling Freezerpro 64 and 7 respectively), and the same SATA Hitachi hard drives. The 9950's are identical in every way, same case, same PSU (enermax ELT620 80+ eff), same Ram (Corsair Dominator ddr2 1066), same mobo (Asus M3A78-T), same cooler (Artic Cooler Freezerpro 64), and the same Hitachi hard drives, same OC.
All the above run headless (no keyboard, mouse, monitor). The 9600 and 6600 are using Mandriva 2007 spring free x86-64, and the 9950's are using Ubuntu 8.04.1 lts amd64.
I have just gone through them all and allowed work for Einstein, so hopefully they all get work from similar freqs.
EEEK, the 9950's downloaded work and they went instantly to comp error one after another. message log says "output file absent". I tried resetting Einstein, it resent lost wus, re-downloaded the executables and skygrid data, and again, they all went instantly to comp error. I suspended Einstein on these 9950's and went back to continuing Docking and Primegrid PSP Sieve wus. If memory serves, 3 weeks ago when I tried to do einstein with the 9950's under windows, I experienced the same issues. perhaps an OC issue?.
If I get the time, perhaps, I'll scale back the OC on them and see if that's causing the comp errors. OK, I reset the bios to auto on everything (on 9950two), now it's telling me "no work" I've reached my daily quota of 4....grrrr. OK, back to re-establish the OC and allow Docking and Primegrid to carry on until later. grrrr
OK, I reset everything on 9950one to auto, now it's resending lost work to this machine and the executables.
OK, the only resend also ended in immediate computation error. Says output file absent.
The rest of the machines are happily crunching away. if you're thinking of why the computation errors, remember those are the only two machines with ubuntu installed on them. I installed boinc via the usual sh method, not some package installer, so boinc is on the Desktop
So. with the exception of the 9950's, all have work and are 100% Einstein ATM. (note: the AMD643700 and Mobile AMD643700 are winxp 32bit)
Unless someone has an idea of
)
Unless someone has an idea of why the Einstein wus are going to comp error immediately, even at stock speeds on the 9950's, I guess we should just wait a day or so and let the clients work and return some samples. I have a 1.5 day cache, so it shouldn't take long. Also, I've re allowed docking and Primegrid on the Q6600 and suspended Einstein, so samples can be gathered.
[edit, reread some posts]NO the TLA bug fix is not in effect.
RE: Unless someone has an
)
You have to install the 32 bit compatibility libs on your 64 bit Linux, as the Einstein app currently is 32 bit only on both Windows and Linux. This can be done via the package management mechanism of your Linux distribution, I think it's "sudo apt-get install ia32-libs" for ubuntu.
Good Luck
Bikeman