SMP client going to breach deadline - please help

Chris Beard · 30 May 2007 at 08:55

Hello again,

So I installed the windows SMP client and am only running one of them on my C2D E4300 @ 3.2GHz. It's SOOOOOO slow. I get about a percent an hourish and sometimes a percent takes a few hours.

Anyway, my latest work unit according to FAHMON will breach it's deadline before it's complete. Is there a way for me to send the work unit back and say I can't finish it in time or should I just abandon it and go back to using two normal clients?

Thanks in advance,
Chris.

Whitestar · 30 May 2007 at 09:11

An SMP unit should'nt take that long on a C2D, unless the overclock is unstable.
If it is and the client crashed it won't use SSE when it's next run. In the log, check for 'Extra SSE Boost OK'. If it isn't there, use Ctrl+c to close the client properly and then start it up again. If it is there, then you have a problem that i'm not familiar with. You'll have to wait till SiriusB pops in. He knows more about the SMP client that I do.

verbal · 30 May 2007 at 09:14

How many hours in each day is the PC on for? The SMP client is really only suitable for people running 24 hours a day it seems.

However if yours is on 24/7 then there is something wrong because you should be able to easily meet the deadlines with a day or more to spare. I'm temporarily running 2 core2duo E6320's at stock speed (1.86ghz) and they will easily make the deadlines.

Chris Beard · 30 May 2007 at 11:10

It does indeed use the SSE - I can see it when it starts a new unit (was interested as it was the first SMP unit I'd completed).

My overclock was 8 hours Orthos stable 3 weeks ago - not changed any settings.

My computer is on 24/7 and nothing else is running. The only processes using processor time from taskmanager are the 4 SMP threads and the one thread that manages them all - not infront of my PC right now.

Has anyone else run the SMP windows client on a E4300?

Whitestar · 30 May 2007 at 11:21

My 840EE will complete the natiest of SMP WU's with an about an hour to spare from a 3 day deadline.

An E4300 should easily do one of those in about a day. Did you set the client to use 100% cpu time?

Chris Beard · 30 May 2007 at 11:25

Whitestar said:
My 840EE will complete the natiest of SMP WU's with an about an hour to spare from a 3 day deadline.

An E4300 should easily do one of those in about a day. Did you set the client to use 100% cpu time?

Yes, I left that option on default. The current estimate from FAHMON was 7 days!

SiriusB · 30 May 2007 at 11:48

Open FahMon and check the log for the current Work Unit. Go to where the WU started and see if says something along the lines of "Using Standard Loops".

If that's there your client has crashed at some point and reverted to the old fashioned way [slow] of crunching.

If that isn't there then all I can assume is something is wrong with the WU itself.

booyaka · 30 May 2007 at 12:08

started the windows SMP client yesterday with a E4300 @ 3ghz. first unit was showing as 1.5 days to complete. Think it was a 2613 gromac (not totally sure)

seems to be ok - had to shut it down today as electric meters getting changed over. will fire it up again later.

Also was gaming online etc for about 2 hours last night and it was still ticking away in the background ok.

rich99million · 30 May 2007 at 12:10

SiriusB said:
Open FahMon and check the log for the current Work Unit. Go to where the WU started and see if says something along the lines of "Using Standard Loops".

If that's there your client has crashed at some point and reverted to the old fashioned way [slow] of crunching.

If that isn't there then all I can assume is something is wrong with the WU itself.

He said above already that it's using SSE

SiriusB · 30 May 2007 at 14:05

rich99million said:
He said above already that it's using SSE

Sorry, I assumed he meant he checked his first ever WU. Re-read and he means he looked at the end of his first WU for the next one.

My bad.

Chris Beard · 30 May 2007 at 14:39

hmm, had some crazy process using the CPU that was nothing to do with FAH - I thought it was as a quick google showed it was a thread manager or something.

Killed it and started again, will report when I get home from work (was at lunch)

shadowscotland · 30 May 2007 at 16:03

was the process svchost.exe? - we had a PC here at work freeze due to this process taking all the resources - It's a problems with some of the MS critical updates/auto update - according to our IT suppliers it's affected about 1 in 20 pc's in the UK! mostly P4's but some PD and C2D

Mattus · 30 May 2007 at 16:44

I used to have svchost hogging an entire core of my Core Duo laptop for hours, usually coinciding with Windows updates. The only way I could stop it happening was to disable the automatic updates service. Didn't know it was that common!

shadowscotland · 30 May 2007 at 17:03

try this link to MS support - works for ~20% of people

But if that doesn't work switch back to windows update (not MS update)

From the Microsoft Update site, go to the Change Settings and choose the disable Microsoft Update option.

It will give you all kinds of warnings about not getting Automatic updates from the Microsoft Update site, but once you go through the uninstall, you get another message that you can now go to the Windows Update site and do your regular updates, and turn Automatic Updates on etc..

You should notice that your CPU problem will go away, and your updates will go back to only taking a minute or so, instead of the 5 or more minutes for the Microsoft Update page (if you actually waited long enough before killing the svchost service running the updater).

uncle_fungus · 30 May 2007 at 17:34

If you tried to install the SMP client as a service, you'll find that windows throws a fit at this. It will cause csrss.exe and either smpd.exe or mpiexec.exe to start sucking up CPU time for no reason at all.

If this has happened, you need to reconfigure the client and disable the service before running the client from a shortcut again.

Chris Beard · 30 May 2007 at 18:54

uncle_fungus said:
If you tried to install the SMP client as a service, you'll find that windows throws a fit at this. It will cause csrss.exe and either smpd.exe or mpiexec.exe to start sucking up CPU time for no reason at all.

If this has happened, you need to reconfigure the client and disable the service before running the client from a shortcut again.

Why uncle, however did you guess. Just got home and it was the mpiexec.exe that sucked up CPU time. Will uninstall the service and try again.

Thanks ever so much for your help, will report back shortly.

WoZZeR · 30 May 2007 at 19:05

This rather belatedly explains the problem I had when I first started FaH.

Hearing about the huge points to be had from Win SMP, I installed it and after settling in FaHmon told me it would take 6 days to finish it............way past the deadline.

Needless to say, I just assumed everyone else had super computers and that my feeble OCed E6700 wasn't man enough.

I should have asked.

shadowscotland · 30 May 2007 at 20:50

always worth asking - even if you feel a bit of a wally - not only will you get an answer but the thread could take on a whole new life of its own.

Chris Beard · 30 May 2007 at 21:08

Right, bit better lads, we're down to 2days and 21hours. Whitestar was right though, as I've restarted the computer and killed the console window a few times it does say it's using standard loops. Then it tells me that the SEE boost is ok when it starts the first percent. Actual output looks like this:

[19:55:04] Preparing to commence simulation.
[19:55:21] - Ensuring status. Please wait.
[19:55:21] - Looking at optimizations....
[19:55:21] - Working with standard loops on this execution.
[19:55:21] Examination of work files indicated 8 consecutive improper terminations of core.
[19:55:35] - Expanded 3924792 -> 21619596 (decompressed 550.8 percent)
[19:55:35]
[19:55:37] Project: 2610 (Run 0, Clone 846, Gen 9)
[19:55:37]
[19:55:40] Entering M.D.
[19:55:46] Calling FAH init
[19:55:47] Writing local files
[19:55:47] ing from checkpoint)
[19:55:47] Read checkpoint
[19:55:47] (23 percent)
[19:55:47] in
[19:55:47] Completed 115000 out of 5000000 steps (23 percent)
[19:55:48] Extra SSE boost OK.

I had to type that all out as you can't get it to mark on the console window - might be a logfile somewhere but as before - lazy. So - given I've double checked my overclock and I am at 3.2Ghz (well 3214Mhz) and my RAM is in 1:1 at 4-4-4-12 (2GB), nothing else is taking processor time - what could be slowing my progress?

EDIT: Oh, it'd Version 1.74 Gromacs SMP core

WoZZeR · 30 May 2007 at 21:12

The average PPD will be dragged down by the earlier mishap.

How long is each frame now taking?

EDIT: In FaHmon, Ctrl B or click on Benchmark.