Just Installed The SMP Client - Question

[Anderu]

It looks like the SMP client is working OK. It should turn in its first WU sometime today or tomorrow. I was wondering if this beta client works as well with other applications as the more refined single CPU clients? Does it give up cycles when other applications need them, or do you pretty much have to shut it down when you want to do anything else?

 

[Xilikon]

It looks like the SMP client is working OK. It should turn in its first WU sometime today or tomorrow. I was wondering if this beta client works as well with other applications as the more refined single CPU clients? Does it give up cycles when other applications need them, or do you pretty much have to shut it down when you want to do anything else?

The SMP client is as good as the regular client when it come to giving cycles to apps when needed. I neved had to shut off the client when I want to play, encode DVD or do intensive apps.

What is beta isn't giving the cycles but how the mpiexec program work with many threads and sending/getting WU. Those 2 aspects is a bit problematic for some and fine for others. However, as long as you watch it pretty often, it will work fine and in case issues arise, a reboot and/or a reinstall will fix.

 

[charpman]

I've had the windows SMP client running just fine on my wifes Vista Business pc, no problems and it behaves well. I have the Linux SMP on a few other boxen and all is well with them too.

However I have failed utterly to get it running on my main work pc (Vista Ultimate). Nothing but random EUE's. After trying everything I could think of, reinstalls, different user accounts etc. I threw in the towel and figured it just didn't like something about my system. Waiting for the next beta release or final release to try again.

I am going to try to get the windows SMP client to run on my game pc which is currently running Linux SMP via VM Ware. I think the slightly less performance of windows smp vs. linux smp is made up for by the vm ware hit. also it's easier to run natively

 

[Xilikon]

I've had the windows SMP client running just fine on my wifes Vista Business pc, no problems and it behaves well. I have the Linux SMP on a few other boxen and all is well with them too.

However I have failed utterly to get it running on my main work pc (Vista Ultimate). Nothing but random EUE's. After trying everything I could think of, reinstalls, different user accounts etc. I threw in the towel and figured it just didn't like something about my system. Waiting for the next beta release or final release to try again.

I am going to try to get the windows SMP client to run on my game pc which is currently running Linux SMP via VM Ware. I think the slightly less performance of windows smp vs. linux smp is made up for by the vm ware hit. also it's easier to run natively

Yes, on normal computers you use daily, unless you like working with Linux, it's easier to just run the windows SMP client since the difference between them is just 30 seconds or 1 minute per frame, which is meaningless in the grand scheme of things. The only way you will consider is only if you are pinching for 10-20 extra ppd

 

[APOLLO]

Yes, on normal computers you use daily, unless you like working with Linux, it's easier to just run the windows SMP client since the difference between them is just 30 seconds or 1 minute per frame, which is meaningless in the grand scheme of things. The only way you will consider is only if you are pinching for 10-20 extra ppd

That's 10-20ppd per SMP client, correct? IOW, 20-40ppd on a machine running 2 SMP clients (8 cores) which translates up to a hypothetical potential of 280ppw. That's quite significant on a dual quad machine, if I have my calculations right.

 

[Xilikon]

That's 10-20ppd per SMP client, correct? IOW, 20-40ppd on a machine running 2 SMP clients (8 cores) which translates up to a hypothetical potential of 280ppw. That's quite significant on a dual quad machine, if I have my calculations right.

Yes, it may be useful for those who only have 1-2 machines. For those who have 12-14 SMP machines running like me, a few points doesn't truly matter It's more about tradeoffs between usability and ppd output. For a dedicated folding machine, it's more efficient to put a linux 64 bits OS and run (that's what I will do with my file server which will be 99% idle, only serving files less than 1 hour per week).

 

[BillR]

I've had the windows SMP client running just fine on my wifes Vista Business pc, no problems and it behaves well. I have the Linux SMP on a few other boxen and all is well with them too.

However I have failed utterly to get it running on my main work pc (Vista Ultimate). Nothing but random EUE's. After trying everything I could think of, reinstalls, different user accounts etc. I threw in the towel and figured it just didn't like something about my system. Waiting for the next beta release or final release to try again.

I am going to try to get the windows SMP client to run on my game pc which is currently running Linux SMP via VM Ware. I think the slightly less performance of windows smp vs. linux smp is made up for by the vm ware hit. also it's easier to run natively

Actually you will gain about 12% running VM. I was running VM is Vista ultimate and it works very well despite all the warnings. I also like the way Vista handles memory and resources. Having run the client both ways I’ve never had a performance issue.

As for your Vista Ultimate Install, have you checked for heat problems? Next obvious question is, will it run the standard client with no issues?

Luck and fold on

 

[charpman]

Actually you will gain about 12% running VM. I was running VM is Vista ultimate and it works very well despite all the warnings. I also like the way Vista handles memory and resources. Having run the client both ways I’ve never had a performance issue.

As for your Vista Ultimate Install, have you checked for heat problems? Next obvious question is, will it run the standard client with no issues?

Luck and fold on

Heat could be an issue, I have 3 10K rpm drives in there, but the CPU is running stock. The standard client was running fine with 2 instances. I haven't tried the registry cleaning trick yet and will try that today. With 33 pc's running in my office and utility/sever room heat is alwasy one of the first thigns I look at Luckily our power is cheap so the AC runs alot! heheh

I was looking at FahSpy and comparing 2 of my X2 4200's that had the same work unit. One was windows SMP and one was running VMWare and the VMWare one had a lower frame time and thus higher PPD. Interesting. So I'm leaning towards leaving the VMWare setups alone for now. Easy enough to pause VMWare to play games

I am now supecting I have some driver issues with the Vista Ultimate box. Last night I tried to run an offline boot defrag with Perfect Disk and it couldn't lock the drive for excluseive access...at boot time..... so something funky is going on.

 

[Anderu]

[13:49:57] - Machine ID: 1
[13:49:57]
[13:49:57] Loaded queue successfully.
[13:49:57]
[13:49:57] + Processing work unit
[13:49:57] Core required: FahCore_a1.exe
[13:49:57] Core found.
[13:49:57] Working on Unit 01 [May 23 13:49:57]
[13:49:57] + Working ...
[13:49:58]
[13:49:58] *------------------------------*
[13:49:58] Folding@Home Gromacs SMP Core
[13:49:58] Version 1.74 (March 10, 2007)
[13:49:58]
[13:49:58] Preparing to commence simulation
[13:49:58] - Ensuring status. Please wait.
[13:50:15] - Looking at optimizations...
[13:50:15] - Working with standard loops on this execution.
[13:50:15] Examination of work files indicates 8 consecutive improper terminatio
ns of core.
[13:50:34] - Expanded 3968076 -> 21618912 (decompressed 544.8 percent)
[13:50:36]
[13:50:36] Project: 2608 (Run 0, Clone 131, Gen 6)
[13:50:36]
[13:50:43] Entering M.D.
[13:50:50] Calling FAH init
[13:50:57] Read topology
[13:50:58] (Starting from checkpoint)
[13:50:58] Read checkpoint
[13:50:58] Protein: Protein
[13:50:59] Writing local files
[13:50:59] Completed 64310 out of 500000 steps (12 percent)
[13:51:02] Extra SSE boost OK.
[14:02:34] Writing local files
[14:02:35] Completed 65000 out of 500000 steps (13 percent)
[15:24:43] Writing local files
[15:24:45] Completed 70000 out of 500000 steps (14 percent)
[16:47:25] Writing local files
[16:47:26] Completed 75000 out of 500000 steps (15 percent)
[18:10:13] Writing local files
[18:10:14] Completed 80000 out of 500000 steps (16 percent)
[19:32:34] Writing local files
[19:32:35] Completed 85000 out of 500000 steps (17 percent)
[20:55:00] Writing local files
[20:55:02] Completed 90000 out of 500000 steps (18 percent)
[22:18:19] Writing local files
[22:18:21] Completed 95000 out of 500000 steps (19 percent)
[23:40:36] Writing local files
[23:40:37] Completed 100000 out of 500000 steps (20 percent)
[01:02:54] Writing local files
[01:02:55] Completed 105000 out of 500000 steps (21 percent)

These frame times seem a little long compared to what I have seen posted. I am on a Core2 Duo E6400, with 2gb of memory.

Any thoughts/suggestions?

 

[fenderltd]

Compare your times here

http://members.cox.net/foldinginfo/smp2.htm

http://members.cox.net/foldinginfo/allsmp.htm

 

[APOLLO]

Those first several frame times do appear quite long, and from the looks of things, you might not be able to complete that particular WU in time before the deadline, unless the quickened pace of the last few frames is maintained.

I tried the SMP client on two machines, a dual Opteron@2.3GHz (2 cores) and a dual Athlon (Barton@2.2GHz). Both of them were completiing frames between 49-56 minutes. They both failed to make it before the deadline. Your times even with 'only' 2 cores shouldn't be much over 30 minutes a frame with a C2D unless the WUs have become a lot more complex recently. If that's the case, Stanford should increase the deadline. BTW, what clockspeed is that machine running at?

Quick question: How important is the amount of system memory when running the SMP client, anyone know?

 

[Anderu]

The CPU is running at stock speeds. I would push it a little higher, but it is an HP (off the shelf) DC7700. There is no obvious way of making it run any faster than it does.

I guess I was suffering from the delusion that the Core2 Duo was a crunching monster, and that running the SMP client, even on an un-overclocked unit, would reap more points. From what I gather here, I will be lucky to get this turned in before the time limit expires. I seemed to have been doing better using the graphical client. Might not be using both cores, but at least it was completing results every day or so.

This is my home PC, and it runs for about 18 hours a day. That is probably hindering my ability to complete the large WU in sufficient time.

 

[Killer[MoB]]

Apollo: The most ram hungry I have noticed for WinSMP is the P2610. It's using 340 megs of ram for the one I am folding right now. It also seems to be one of the units that is most affected by cache. I currently have a P2609 on a linux box that is using about 320 megs. Any less that that plus what it takes to run the OS on a dedicated box would cause lots of swapping to take place which would definatly increase frame times.

Anderu: Not a lot of data on the P2608 yet, but that's pretty slow. Something is not right with that for sure. Thermal throttling maybe? I'm getting 33:48 per frame on a X2 @ 2.65 ghz. A E6400 @ stock should be at least around 40m/frame I would think.

 

[BillR]

Heat could be an issue, I have 3 10K rpm drives in there, but the CPU is running stock. The standard client was running fine with 2 instances. I haven't tried the registry cleaning trick yet and will try that today. With 33 pc's running in my office and utility/sever room heat is alwasy one of the first thigns I look at Luckily our power is cheap so the AC runs alot! heheh

I was looking at FahSpy and comparing 2 of my X2 4200's that had the same work unit. One was windows SMP and one was running VMWare and the VMWare one had a lower frame time and thus higher PPD. Interesting. So I'm leaning towards leaving the VMWare setups alone for now. Easy enough to pause VMWare to play games

I am now supecting I have some driver issues with the Vista Ultimate box. Last night I tried to run an offline boot defrag with Perfect Disk and it couldn't lock the drive for excluseive access...at boot time..... so something funky is going on.

I run perfect disk on this machine and the msg you are getting is normal, I double checked with Raxco. Funny thing is, it will defrag anyway. That only started happening with one of the last updates (windows) as well.

KillerMob and I were talking about this earlier maybe this will help. In many cases the SMP client is very sensitive to whether or not you have an active network connection. Along with that info add the fact that Vista and XP have power savings turned ON by default for your network. SMP doesn’t like disks going to sleep or the system going to sleep either.

Just a few things to check

 

[Xilikon]

From what I've gathered around, the 2608, 2609 and 2610 are the WU who love 4 Mb of cache and bigger RAM so for those with less than this, it will take longer to crunch them compared to others.

On my box, any WU who can fit the 2 Mb of cache usually take ~11 mins per frame and the bigger ones will take 15-16 mins per frame. What help my speed is the 2 Gb of RAM at tight latencies and a 975X chipset. This time can be compared to the E6600 from what I saw.

To get the best results, the easiest way would be to configure the RAM so it provide plenty of bandwidth at tight timings.

 

[APOLLO]

From what I've gathered around, the 2608, 2609 and 2610 are the WU who love 4 Mb of cache and bigger RAM so for those with less than this, it will take longer to crunch them compared to others.

On my box, any WU who can fit the 2 Mb of cache usually take ~11 mins per frame and the bigger ones will take 15-16 mins per frame. What help my speed is the 2 Gb of RAM at tight latencies and a 975X chipset. This time can be compared to the E6600 from what I saw.

To get the best results, the easiest way would be to configure the RAM so it provide plenty of bandwidth at tight timings.

What about enabling NUMA in Opteron systems? I realize that F@H SMP client probably doesn't natively support it, but would this help or hurt?

 

[Xilikon]

What about enabling NUMA in Opteron systems? I realize that F@H SMP client probably doesn't natively support it, but would this help or hurt?

I have absolutely no experience with this system so I cannot comment. Just do 1-2 WU with NUMA disabled then redo it after enabling it to see if the performance will improve.

EDIT : After googling for NUMA on Opteron system, I understand a bit better how it works so my guess would be to disable it so both cpu will share the same memory pool since a SMP client will spawn 4 threads and the dual dual-core opterons will have 4 cores total. However, if you have a dual quad-core Opteron system, NUMA would be better enabled since each SMP client will have his own memory poll with no need to exchange data between the 2 cpu but you will need to have at least 1-2 Gb per cpu to be efficient.

However, I could be wrong since a shared memory bus will have limited bandwidth. That's why the best way to know if it is good or not is to experiment both settings.

 

[charpman]

I run perfect disk on this machine and the msg you are getting is normal, I double checked with Raxco. Funny thing is, it will defrag anyway. That only started happening with one of the last updates (windows) as well.

KillerMob and I were talking about this earlier maybe this will help. In many cases the SMP client is very sensitive to whether or not you have an active network connection. Along with that info add the fact that Vista and XP have power savings turned ON by default for your network. SMP doesn’t like disks going to sleep or the system going to sleep either.

Just a few things to check

Odd about Perfect Disk... heh but good to know. I turn all my settings to "consume as much power as possible" right after installs, so that shouldn't be it. I used to mess with that stuff but got annoyed waiting for things to spool up, seemed to defeat the point of leaving my PC on all the time. I've also noticed that my failures are less (on hardware) stuff only seems to break when it gets shut off! heh

Work got busy so after doing the registry cleaning I didn't have time to do the install of SMP client yet. Hopefully today sometime.

 

[charpman]

From what I've gathered around, the 2608, 2609 and 2610 are the WU who love 4 Mb of cache and bigger RAM so for those with less than this, it will take longer to crunch them compared to others.

On my box, any WU who can fit the 2 Mb of cache usually take ~11 mins per frame and the bigger ones will take 15-16 mins per frame. What help my speed is the 2 Gb of RAM at tight latencies and a 975X chipset. This time can be compared to the E6600 from what I saw.

To get the best results, the easiest way would be to configure the RAM so it provide plenty of bandwidth at tight timings.

I run 2gigs ram in any PC I have that is actually used (started doing that few years ago with XP mostly becuase of EQ2, it really sped up zone times! ), be it Windows or Linux, and 1Gig for dedicated folders with Linux and they seem to do fine. My C2D's (E6600's OC'd to 3.4) are in the 10-12min range depedning on the WU. That cache does make a huge difference. Probably more than ram it seems. I'm still getting used to the high latency in DDR2 compared to DDR. All my AMD systems have 2-2-2-4 ram in them and the DDR2 latencies are just poor by comparision, though I undertand the bandwidth is greater, which lots of stuff seems to like. Vista's performance tests really like the DDR2 bandwidth as wel it seems. I haven't got any high end DDR2 yet and tried to mess with the timings.

 

[APOLLO]

EDIT : After googling for NUMA on Opteron system, I understand a bit better how it works so my guess would be to disable it so both cpu will share the same memory pool since a SMP client will spawn 4 threads and the dual dual-core opterons will have 4 cores total. However, if you have a dual quad-core Opteron system, NUMA would be better enabled since each SMP client will have his own memory poll with no need to exchange data between the 2 cpu but you will need to have at least 1-2 Gb per cpu to be efficient.

ATM, I have a dual uni-core Opteron system with 2GB of RAM. Dual quad-core Opterons (Barcelona) aren't released yet, and probably won't be until well into Q3, from the looks of things.

However, I could be wrong since a shared memory bus will have limited bandwidth. That's why the best way to know if it is good or not is to experiment both settings.

I have already tried the SMP client with NUMA enabled several times. It kept crashing and turned in abysmal frame times that couldn't make the deadline. I have yet to attempt it without NUMA, but I don't believe that alone will make a difference in processing the WUs. Maybe it will make a difference in stability though, and worth a try to see if it helps in that department alone. Thanks for the informative reply.

 

[Xilikon]

ATM, I have a dual uni-core Opteron system with 2GB of RAM. Dual quad-core Opterons (Barcelona) aren't released yet, and probably won't be until well into Q3, from the looks of things.

I have already tried the SMP client with NUMA enabled several times. It kept crashing and turned in abysmal frame times that couldn't make the deadline. I have yet to attempt it without NUMA, but I don't believe that alone will make a difference in processing the WUs. Maybe it will make a difference in stability though, and worth a try to see if it helps in that department alone. Thanks for the informative reply.

I read a topic in the F@H community forum about NUMA. Vijay and others said that the client is written using the most conservative coding so NUMA support isn't integrated in their clients. The only way to use NUMA is indirectly via OS level or driver level.Since you said it's crashing left and right, i believe running the SMP client without NUMA is your best bet, based on the comment from Vijay Pande

 

[BillR]

Odd about Perfect Disk... heh but good to know. I turn all my settings to "consume as much power as possible" right after installs, so that shouldn't be it. I used to mess with that stuff but got annoyed waiting for things to spool up, seemed to defeat the point of leaving my PC on all the time. I've also noticed that my failures are less (on hardware) stuff only seems to break when it gets shut off! heh

Work got busy so after doing the registry cleaning I didn't have time to do the install of SMP client yet. Hopefully today sometime.

Consume as much power as possible???? Who do you think you are, Al Gore?

Kidding aside I do the same thing as you, never allow disks to shut down, no sleep mode and I don’t use a screen saver because all my monitors are lcd and the power switch is handy.

I did down load a new client the other day and configured it with user settings turned off as well as set permissions in the FAH client to all users. I rebooted back into admin mode (turned user settings back on) and now I’m folding 2651s with no early ends as well as no left over junk file warnings. Go figure.