slow SMP performance

magicmedicine

I'm currently averaging 35 minutes a frame on project 2653. running a q8200 with 4gb of ddr2 1066. i am also running a GPU client.

another question.. when i try to instal FahMon, i get an error message saying language not supported. anyone know the fix to this?

Leonardo

Hey, I just saw your thread over at Foldingforum.org.

Concerning slow performance: check Processes under Task Manager. There should be four FahCore_a1.exe running. (might be off on the exact file name - not on a Folding machine ATM) Yes/no? What percentage of CPU utilization are they at. Let us know, and we'll move on from there.

magicmedicine

okay. four instances of fahcorea1 and 1 instance of fahcore11 for the GPU. i used process lasso to assign the four fahcorea1s to cores 0,1,2 and fahcore11 to core 3.

CPU usage is 100 percent. fahcore11 is at about 25% avg and the 4 fahcorea1s are at 23,21,10 and 12 percent respectively.

Leonardo

Is your GPU2 Folding with an Nvidia card? If so, you should uninstall your current video drivers and install drivers 180.60 or 180.84. Those drivers put almost 100% of the processing requirements on the GPU core, leaving the CPU cores dedicated for SMP. That will decrease GPU productivity slightly (approx 150ppd, IIRC) but will increase the SMP Folding productivity significantly. Another advantage of the 180.60/84 drivers is that should not need to assign affinities.

i used process lasso to assign the four fahcorea1s to cores 0,1,2 and fahcore11 to core 3.

That is the best method for Nvidia + SMP if you are running drivers older than 180.60.

I'm currently running overclocked Q6600 systems. Each system also has one or two Nvidia GPU Folding clients running. The Qs are clocked at 3.2 - 3.6GHz. If I remember correctly, the SMP clients with the 2653 work units are returning on average 2200 - 2700 points per day.

magicmedicine

im using an ATI HD4850. looking back at the logfiles.. i noticed the SMP client started to do 35 mins perframe only when i turned the GPU client on. would it be better to just run the GPU client alone? or run 3x uni processor clients + GPU?

Leonardo

magicmedicine, I don't remember processing times for frame. I usually think in productivity terms of PPD (points per day). With different processes - GPU vs CPU - and different clocked systems, for me it's just easier to compare in terms of processor (core) output, rather than fram times. Just the way my mind works.

To track PPD, you can use the free tracking program called FahMon - Folding at Home Monitor.

Leonardo

Concerning ATI GPU + SMP combined production, you might be better served to get advice from an ATI GPU Folding person. If you could give a rough estimate of PPD for your system, I could be of more help.

magicmedicine

couldnt install fahmon, using fahspy instead. currently it states ppd for my CPU is 884 and 2357 for GPU. i dont think my cpu should be that low... should it?

Leonardo

For a 2653 work unit, it should be 2000+ PPD.

Leonardo

OH, sorry, I didn't read your posts thoroughly enough. You are Folding with an AMD 8200. Sorry, I do not know what production should be expected from one of those, but it certainly should be more 884PPD.

magicmedicine

confirmed, its the GPU client slowing it down. after i closed the gpu client, it finished a frame in 12 minutes. now how to work around this??? not using an AMD, leo. its a intel c2q q8200 @ 2.33 ghz

Leonardo

Oh, your last step is join Team 93...if not already a member.

Snarkasm

Leo - Intel C2Q 8200 processor, ATI HD4850 graphics.

I'd check and make sure you've properly assigned your cores (or that when you stop SMP, only one core stays at 100%). Stopping GPU shouldn't jump your SMP frames by 8 minutes unless it's taking some of their cycles.

magicmedicine

can you please give me a screenshot of your process lasso? maybe im not doing it right. now my fahspy says 2018 ppd on the SMP client, after closing the GPU. before this was 884 ppd.

Snarkasm

I'm not at my folding farm at the moment, so I'm afraid I can't. If you end SMP and run the GPU client, does only one core jump?

magicmedicine

one core will be around 80% the others at 20% or so. does that mean anything?

Snarkasm

It sounds to me like, as that's showing, GPU is sucking some processor cycles from your SMP processes. If you're only running GPU right now, make sure EVERYTHING related to FAH that's currently in your processes is assigned to the same single core in Process Lasso. Then run SMP only and make sure EVERYTHING related to FAH at that time is then assigned to the other three cores. If it's not obeying Process Lasso, I'm not sure where to go from there. Others may have ideas.

magicmedicine

after assigining the actual folding@home.exes to the specific cores, my smp is now reading 1718 ppd. the gpu is reading 2358ppd. i think that gives me a nice 4k ppd. but id still like the cpu client to be faster though. considering buyin a new cooler to OC.

Snarkasm

Much better. I thought that might be the case.

If you can give it an OC, that'll improve. My Q66, when it's actually running, can get about 2200ppd on SMP at 3.6GHz.

magicmedicine

after running for about 2 hours, the SMP client has slowed back down to about 1.3k ppd. whats going on guys?!! argh.

Leonardo

Update: alright, I'm looking at FahMon now, and can report accurately.

System 1: Q6600 @ 3.6GHz, 2 X GPU2 clients running, 1 X SMP. Work unit 2653 at 2987PPD.

System 2: Q6600 @ 3.2GHz, 1 X GPU2 client running, 1 X SMP. Work unit 2653 at 2800 PPD.

System 5: Q6600 @ 3.5GHz, 2 X GPU2 clients running, 1 X SMP. Work unit 2653 at 2999 PPD.

Leonardo

the SMP client has slowed back down to about 1.3k ppd.

It sounds like you have some type of power saving/CPU throttling service kicking in. In the BIOS, have you turned off power saving settings, such as EIST?

magicmedicine

eist was turned off all the while and so was c1e.. i read somewhere that vista does speed stepping on balanced power profile.. so switched to high performance just to be sure. rebooted and now running both clients again

_k

Um it might have just been as it was stepping through the iterations of the WU the average worked out to be more accurate. WU ppd production will change of the course of some units, mas0n and I have had long conversations about this...nothing to worry about. Of course if it keeps dropping there is something not being controlled correctly.

I have seen Fahmon start pulling ~20% or more off of one core and it greatly hindered ppd, duh. Double check Process Lasso to make sure there are not .exes' that are pulling CPU usage. If there is nothing currently just sort it by average usage to double check. That 1.3k should be about right. On WU 2653, 1760 point, I get 3.1k ppd with my q9450 at 3.55GHz. Any OC will help a lot because you are on a 1066 FSB currently.

magicmedicine

i figured it might be something like that. my fsb is 1333. 333x7. no one seems to know about this q8200. maybe it was a mistake buying it. i hope it ocs well. and after the restart and everything the first frame took 18 minutes to do. sigh. i feel like just running 3 uni proc clients.

btw, could it have something to do with the heat? ambient temp where i am is 30c and my core temps are 73,66,69,67 currently.

mas0n

For reference, my rig with Q6600 @ 3.2GHz + HD4870 under Vista 64 runs 13-14 min/frame and produces ~1800PPD on WU 2653. This is with cores 0,1,2 assigned to SMP and core 3 assigned to GPU2.

core temps are 73,66,69,67 currently

Those cores are hot hot hot. Are you sure you're making good contact? I think the Q8200 should load 50-55C @ stock speeds w/ stock cooler.

_k

Those temps are as hot when I was running 3.7 24/7 at 1.36 vCore with my OCZ V2. Thats not right.

Leonardo

Mind you I only use Windows XP, but I've never seen a CPU client change its rate of production on a work unit once the WU is fully engaged. Either there's some type of application/service kicking in or...something is using CPU computing power other than Folding.

i figured it might be something like that. my fsb is 1333. 333x7. no one seems to know about this q8200

FSB is mainly irrelevant for Folding production. The most important factors for a core's performance are core frequency and CPU cache, in that order.

73,66,69,67

Yeah, that is hot for a 45nm core architecture. Those temps wouldn't make much of any difference with a 65nm Q6600 (especially the hot B3 steppings). Maybe the CPU clock is lowered as your temperatures rise? Some type of power/heat management. (sorry, I know I keep harking on this. please excuse my ignorance about the newer CPUs and Windows Vista)

my rig with Q6600 @ 3.2GHz + HD4870 under Vista 64 runs 13-14 min/frame and produces ~1800PPD on WU 2653

I'd be interested to see what production you get with all four cores' affinity on SMP, with GPU2 turned off. Q6600-3.2GHz-SMP-WU 2653 should be about 2800PPD.

If you already posted about, sorry, I missed it. Do you have cores 0,1,and 2 assigned to SMP and only core 3 assigned to GPU? That would be the best affinity assignment for your setup.

mas0n

Leonardo wrote:

I'd be interested to see what production you get with all four cores' affinity on SMP, with GPU2 turned off. Q6600-3.2GHz-SMP-WU 2653 should be about 2800PPD.

It's been awhile since I ran only the SMP client but I think I was getting 2200-2400PPD under Vista. I know it's lower than I was getting under XP, but never had the time to look into it and see what is normal.

magicmedicine

i got a bit lazy on this build and didnt use any thermal paste. so no im probably not making any good contact. keep in mind though, i live in a tropical country. haha. going to test a bit more and im almost set on a complete reinstall of both clients. since 6.23 for SMP came out. in fact, if this last bit of testing proves my thoughts to be true, i might skip the GPU altogether and run dual SMP with a new cooler.

magicmedicine

well guys i found something pretty interesting. after OCing the chip to 2.81ghz from 2.33, it goes back down to about 13 minutes a frame. possible oc gem, this q8200. i did it on the stock cooler and temps are only about 2 or 3 degrees higher than stock speeds. im just wondering how a simple 500mhz overclock could yield such a huge difference in performance.