TR Forums

Hi koinkoin,
I've just moved house last week so don't have too much time to look at this at the moment.

It looks like your issue is with awk in building the C libraries. I build on Debian, although unstable rather than Etch, but I don't think awk should be too different. For reference

Alternatively, you could skip the building from source and just try taking apart the benchmark CD image and then rebuilding it from the binaries and libraries that are in the image. You may have to rebuild busybox and set CONFIG_FEATURE_WGET_LONG_OPTIONS=y in busybox-1.2.0/.config to get it to honour the proxy, I don't know - I've not played with proxies and busybox executables.

Perfect it was version of awk,
with gawk it work, not with mawk or awk-orginal package
Give the tips in .txt file in source to use gawk
Bbl i have to bench some machines

I have some trouble on rescent server with core 2 duo xeon cpu:
Process always die on Bonus Gromacs WU

Processor Detection
Processor 0 is an Intel(R) Xeon(R) CPU 5140 @ 2.33GHz
Processor 1 is an Intel(R) Xeon(R) CPU 5140 @ 2.33GHz
Found 2 processors

Progress
Starting benchmark of Tinker WU
Starting benchmark of Amber WU
Finished benchmark of Tinker WU
Starting benchmark of Bonus Gromacs WU
Benchmarking error - some processes died. Please reboot ...

Results
Core Points Per Hour Points Per Day
Tinker 5.39552 129.493

notfred, any plans to update the CD with the Linux SMP client? Folks are asking about it.

Good idea, I'll look into that.

Took a look in to this and at the moment I am going to hold off on the SMP client. There are 2 problems:
1) It's still in beta - which means they are likely to change it before it goes production and are likely to give wildly different points for it than what they have at the moment.
2) It's 64 bit only. The folding benchmark CD is 32bit. Whilst I could move to 64bit that instantly removes all 32bit only processors from the ability to run the benchmark and I think there are too many 32bit processors around folding at the moment for this.

1) Make sense. I'd wait.

2) Both Intel and AMD have been selling 64-bit desktop processors for quite a while. AMD since late 2003, and Intel for nearly two years now...

Hello,

I recently installed Ubuntu 64bit on my Athlon 64 X2 4200+ for the sole purpose of running the SMP Linux Folding@Home client. It is fast -- very fast! This machine went from 200-250 PPD (running 2X 32bit F@H) up to 800-1000 PPD (running the SMP 64bit F@H).

Here's what I learned: the client itself is 32bit, and the cores are 64bit, so you have to install some 32bit libraries. I would have been lost w/o help from the F@H forums.

There are specific work units just for the SMP clients -- I don't think that you could run the same WU on different cores.

Hello,

A question on the 64bit SMP Folding@Home: I assume that this would use all the 16 registers in the Athlon 64 cores? Does the Core 2 Duo have 16 registers, or just the 8 "normal" x86 registers? What are the differences between the EMT64 on the Core 2 Duo and the x86-64/AMD64 on the Athlon 64 X2?

It may well be that the 64bit F@H cores "like" larger L2 caches as much (or more) than the 16 registers; since my brother's 17" iMac Core Duo (1.8gHz IIANM) runs the same 1,760 point work units in virtually the same time it takes my Athlon 64 X2 4200+ to complete them.

From what I understand, AMD's CPU advantages are not enough to match the advantage of the super fast SSE of the C2D.
The next gen AMD CPU (later this year?) should significantly improve AMD's SSE.
Also with a built in memory controller, AMD CPUs lack of L2 cache does not make that big of a difference in most cases.

NeilBlanchard wrote:

Here's what I learned: the client itself is 32bit, and the cores are 64bit, so you have to install some 32bit libraries. I would have been lost w/o help from the F@H forums.

Yup, I've been slowly working on building an SMP diskless folding CD, network, etc image but that is what has been killing me in that I have to build both 64bit and cross compile to 32bit. Had a big break through last week, not had a chance to do much more on it and see if I am close on this yet.

Hmm.. I've been trying to test a couple of systems with the bench CD, and they haven't been returning WUs for Gromacs 3.3, leading them to hang and not produce any totals, either. Anybody else see this problem?

Sorry it has taken me a little while to follow up on this but I know what the problem is. I have to download the folding software from Stanford (due to the licensing) and they just made an alteration to the Gromacs3.3 core on 17 April where it no longer prints "(1%)" in the log file, but now prints "(1 percent)". Stanford up revs the cores periodically.

I think I've fixed this for all WU types so it should work regardless of whether there is "%" or " percent" (well tinker reports frames, so I just added space/tab checking) and it doesn't matter if there is space or tabs in there either (well tinker reports frames, so I just added space/tab checking). I'm leaving the new version building overnight, will start a test run tomorrow AM and if everything goes well, upload a new version either tomorrow or Sunday sometime. Of course, if it doesn't work, it may be a little longer....

OK new version uploaded with the fix for this. It will still benchmark with exactly the same values as the previous one as the only change is in parsing the log files.

Works perfectly for me on a couple of test rigs, one 4-core and one 8-core. Thanks! How does 1565 PPD from a single system sound?

Hmm.. Looks like the changes to the Gromacs 3.3 core include a PPD performance decrease vs. the older revision. That throws off my results some, but I guess it can't be helped.

Damage wrote:

How does 1565 PPD from a single system sound?

depends - what's the system's power consumption & initial cost?

eitje wrote:

Damage wrote:

How does 1565 PPD from a single system sound?

depends - what's the system's power consumption & initial cost?

Breathtaking.

Stay tuned.

Damage wrote:

Hmm.. Looks like the changes to the Gromacs 3.3 core include a PPD performance decrease vs. the older revision. That throws off my results some, but I guess it can't be helped.

Unfortunately that's the Stanford license, redistribution of the folding software is prohibited, everyone has to download it from Stanford.
I guess that makes comparing benchmark runs a little dodgy unless you know exactly what you are doing. I think it might be a good idea to print out the core versions along with the results, should make it a little easier to spot when they change.

notfred wrote:

Damage wrote:

Hmm.. Looks like the changes to the Gromacs 3.3 core include a PPD performance decrease vs. the older revision. That throws off my results some, but I guess it can't be helped.

Unfortunately that's the Stanford license, redistribution of the folding software is prohibited, everyone has to download it from Stanford.
I guess that makes comparing benchmark runs a little dodgy unless you know exactly what you are doing. I think it might be a good idea to print out the core versions along with the results, should make it a little easier to spot when they change.

Sounds like a good idea to me. Should at least help us track changes over time.

Damage wrote:

Works perfectly for me on a couple of test rigs, one 4-core and one 8-core. Thanks! How does 1565 PPD from a single system sound?

Considering my 2 year old dual-core AMD, with half-slower-than-Intel's SSE, can deliver in the neighborhood of 1170 PPD even while I'm using it for other work... (though I'm only using Linux SMP on it and no other types of WU's [though that is running through VMware], and that's its peak speed), I suppose it's good for an average of all of the types.

What does it report for the GRO-SMP WU's?

dang, what kind of OC do you do on that, RD?

Damage wrote:

eitje wrote:

Damage wrote:

How does 1565 PPD from a single system sound?

depends - what's the system's power consumption & initial cost?

Breathtaking.

Stay tuned.

ahh, fooey! i was sure you were gonna answer my non-chalant, simple question.

eitje wrote:

dang, what kind of OC do you do on that, RD?

An amount over 700 MHz. Not all that much speed considering it's the slowest Opteron available (a 165 @ 1.8 GHz stock), and they make the Athlon FX as fast as 3.0 GHz, now.

ah! cache. that's the difference.

I'd rather have DDR2 RAM so the larger WU's went faster. Of course, now that my frigging VMware is not working, and as of a few hours ago I can neither upload nor download any WU's, it's meaningless anyway.

notfred, I seem to have run into a problem. The benchmark CD (25 May 07 edition) doesn't seem to work on either of the Intel P35-based motherboards I have here, the Gigabyte P35T-DQ6 and the Asus P5K Deluxe.

Looks like some kind of odd networking problem where it thinks networking is working, but it's not. The CD proceeds through booting, then fails to download anything. It throws some errors:

bunzip2: Decompression failed
bunzip2: Decompression failed
chmod: core_*fah: No such file or directory

Then it acts like it's kicking off tests, until it croaks with:

killall: FAH502-Linux.exe: no process killed
Benchmarking error - some processes died. Please reboot ...

When I first ran into this problem with the Asus board, I tried some PCI NICs, but they didn't help. Now the Gigabyte has the same problem.

FWIW, I can't hit the http server on the IP address of that system's DHCP lease. Obviously, networking isn't working right.

Help? I'll provide any more info I can.

Does the CD still work on other boards? Just trying to work out if it is particular to the P35-based motherboards or if something else has broken.

I can probably up-rev the kernel fairly easily - I've been rewriting the diskless CD and I have a later version working on that so I can move it across - and see if that helps fix it as it should have more modern network drivers in it.

Yep, the CD works fine on other boards still.

OK just rebuilt the Benchmark CD with linux 2.6.22.1 and looking around on various message boards, I think it is new enough to support P35 boards.
Grab the latest version from http://reilly.homeip.net/folding/benchmark.html

Seems odd to me that a PCI card wouldn't make it work, though. Unless our host forgot to disable the onboard NIC, I'd think it would work perfectly well with the old hardware.