Significant performance disparity (Linux on 64bit Xeon -VS- Mac on Core 2 Duo)
6 Feb 2008 11:35:11 UTC
Topic 193492
(moderation:
Hello everyone, Im an old SETI@home user joining the Einstein@home crunching (which frankly fascinates me even more than SETI).
I've noticed a big difference between WU completion times on two machines:
Machine A: 3.4 GHz Xeon (64 bit, dual core), running Debian GNU/Linux 4.0 (sci app 4.20)
Machine B: 2.33 GHz Core 2 Duo (32 bit, dual core), running Mac OS X 10.4.11 (sci app 4.04)
Interestingly "Machine B" takes 10-12 hours to complete a WU, while "Machine A" takes three times as long: 30-35 hours! (Both are running Hierarchical all-sky pulsar search, not the Faster version)
On "Machine A" I tried both ways: Install 64 bit OS and run 686 science app, as well as installing 32 bit OS and running 686 science app - both ways yield the same 30+ hours results... :-(
As a comparison, SETI@home WU on "Machine A" took 1-2 hours, and on "Machine B" about 2-4 hours.
So, can someone enlighten me on what is the deal here? What is slowing things down here, is it the linux app in general, 64 bit CPU's in general, or the particular linux/amd64 combination? Or perhaps even a problem with the particular app version(2)?
I don't have any 32 bit machines of comparable clock speed to test the 32 bit linux scenario, but VMware gets pretty close to native speeds, so will give it another run in VMware on "Machine B".
P.S. I did search the forums, and did read about the 0 benefit of 64 bit CPUs, but this isn't just a lack of benefit, this is 3x slower... :-(
Significant performance disparity (Linux on 64bit Xeon -VS- Mac
)
The default Linux app that you're using doesn't contain any optimized code. A new "power users app" is available that will speed things up considerably. See this thread for more details:
Power users app
RE: Hello everyone, Im an
)
Welcome to Einstein@Home!
One additional note:
It's not just that the E@H app is slow on "Machine A", look at the BOINC benchmark result: it's terribly slow. Just from the benchmark results and the E@H runtime, I would doubt that the CPU is really a double core Xeon at 3+ GHz, I'd suspect it's an older Pentium 4 generation single core, with "Hyperthreading" enabled (one physical core which acts as two virtual CPUs). This would be about consistent with the observed performance.
The only other explanation that I can think of would be that the CPU is not running at its maximum speed (maybe throttled because of overheating ??).
CU
Bikeman
RE: RE: Hello everyone,
)
Actually, I guess that you would be correct about that. None of the current generation Xeons have a default 3.4 GHz clock speed as yet, so it would have to be an old single-core P-4 type Xeon with Hyperthreading. Still though, the new app would make a difference.
I was thinking Dempsey for a
)
I was thinking Dempsey for a while, they were 65nm dual core Xeons at up to 3.73GHz but iirc they all had HT enabled, so a single Dempsey should show up as 4 CPUs in Boinc. In addition Dempsey were mostly 1066MHz bus, 3.4 is not dividable by 266...
Anyway, OP, you should follow Donalds advice, install the Power Users app, you will see a nice speedup. Performance still cant compare to your Mac, Core 2 is a superior CPU and MacOS X Intel is still the fastest OS for E@H.
Team Philippines
RE: The default Linux app
)
Awesome! Just swapped the app on the host, will see how it goes.. I'm also running 4.20 on VMwared Debian on "Machine B" right now, so far 4.5 hours and 42% into the WU, so it may be comparable to Mac's results. After that one is done I'll swap the app to 4.27 and let you guys know how that changed things!
Thanks for the tip!
RE: RE: RE: Hello
)
Ahhhh, yes, of course! That is exactly what that is. I recently picked up these used Dell PowerEdge servers, and I guess got a lil too excited when I saw 2 and 4 CPUs in htop :-)
Great news for the 64bit
)
Great news for the 64bit users, BM compiled a 64bit GNU/Linux power users app:
http://einsteinathome.org/node/193498
You were one day too early to install the power users app then =D
I must confess i had expected it would not speed up things, but as BM said, maybe on some procs its different, testing will show. Im still wondering where is AMD when it comes to 64bit, they havent exactly done much, if anything, to support and improve 64bit computing compared to the marketing hype they made about it over the years. "Lots of screams but little wool".
Team Philippines
I have bought a 64 bit AMD
)
I have bought a 64 bit AMD Opteron 1210 but I installed 32 bit SuSE Linux 10.3 because there were few 64 bit Linux apps. But I have a copy of 64 bit SuSE Linux 10.3 if I wanted to try 64 bit computing. I am running 4.27 Einstein app and have almost 2000 pending credits, most due to CPUs clocked higher than my 1.8 GHz chip, the cheapest Opteron box I could find.
Tullio
RE: RE: The default
)
4.20 - 10.46 hours
4.27 - 7.65 hours
looks good, now time to try that native amd64 4.28! :-)
P.S. The irony is that now it actually makes more sense for me to run Einstein in a virtual Linux guest, rather than on the native Mac host... ahh, reminds me of back in the day when the fastest SETI@home app was a win32 native ran in wine on a Linux host :-) (at that time it was because Intel compiler's code was faster than gcc's, and Linux version was not yet available)
RE: I have bought a 64 bit
)
If software availability is your concern, give Debian a try. Has a very nice and stable amd64 (among many others) port, and there's a crapload of packages in repository. Apt sorts-out all the dependencies, so no need for rpm hunting ;-)
BTW, a very nice feature of installing boinc-client from atp, is that it is by default configured to run as daemon. Personally, I kept having problems with lockfile lingering after client's shutdown (and preventing its automatic restart after reboot) when I tried to set-up boinc daemon in RHEL/CentOS by hand, but Debian package is already sorted out and works flawlessly. I hear RPM is in the works, but not here yet...