Forceware 180.70 + SMP = Win

[geokilla]

The new Forceware 180.70 does NOT require lots of CPU usage to fold anymore. I have my CPU usage set to 100% for GPU folding, and FahCore_11.exe uses around 5% of my CPU. This lets me run SMP on a dual core machine! Now as others said, the reason to make the switch to Vista is now gone.

It took a while to get SMP right, but I got it set up. It's running as a service as well, which is a major plus for me. FahMon says I should get 5300PPD if I fold 24/7. SMP takes just under 15 min to complete one percent.

Quick question though. I noticed in Task Manager that there's 4 processes of FahCore_a1.exe. Is this correct? I have the -smp 2 flag setup last night.

[06:23:09] - Ask before connecting: No
[06:23:09] - User name: geokilla (Team 38296)
[06:23:09] - User ID: 4ECB63996A2BD1C9
[06:23:09] - Machine ID: 1
[06:23:09]
[06:23:10] Loaded queue successfully.
[06:23:10]
[06:23:10] + Processing work unit
[Nov23:10] Work type a1 not eligible for variable processors
[06:23:10] - Autosending finished units... [Nov23:10]23 06:23:10 UTC]
[06:23:10] Core required: FahCore_a1.exe
[06:23:10] Trying to send all finished work units
[06:23:10] + No unsent completed units remaining.
[06:23:10] - Autosend completed
[06:23:10] Core found.
[06:23:10] Using generic mpiexec calls
[06:23:10] Working on queue slot 02 [November 23 06:23:10 UTC]
[06:23:10] + Working ...
[06:23:10] - Calling 'mpiexec -np 4 -channel auto -host 127.0.0.1 FahCore_a1.exe -dir work/ -suffix 02 -cpu 95 -checkpoint 15 -service -forceasm -verbose -lifeline 736 -version 623'

[06:23:10]
[06:23:10] *------------------------------*
[06:23:10] Folding@Home Gromacs SMP Core
[06:23:10] Version 1.74 (March 10, 2007)
[06:23:10]
[06:23:10] Preparing to commence simulation
[06:23:10] - Ensuring status. Please wait.
[06:23:27] - Assembly optimizations manually forced on.
[06:23:27] - Not checking prior termination.
[06:23:27] - Expanded 608256 -> 3255645 (decompressed 535.2 percent)
[06:23:27]
[06:23:27] Project: 3062 (Run 0, Clone 127, Gen 68)
[06:23:27]
[06:23:28] Assembly optimizations on if available.
[06:23:28] Entering M.D.
[06:23:34] Protein: p3062_lambda5_99sb
[06:23:34] Writing local files
[06:23:34] Extra SSE boost OK.
[06:23:34] point
[06:23:34] Protein: p3062_lambda5_99sb
[06:23:34] Writing local files
[06:23:34] Extra SSE boost OK.
[06:23:34] Writing local files
[06:23:34] Completed 0 out of 5000000 steps (0 percent)
[06:38:34] Timered checkpoint triggered.
[06:38:40] Writing local files
[06:38:40] Completed 50000 out of 5000000 steps (1 percent)

I know I'm folding for another team, but I got my other computes folding for AT as well. In the end, to me, it's all for a good cause.

 

[Insidious]

Four processes is right. (SMP was originally designed for quad cores)

Glad you're folding!

-Sid

edit: I never heard of the -smp flag with a 2 after it.... what is that?

 

[Insidious]

OK, so I put the 180.70 drivers on one of my PCs a couple hours ago.... curious me

This PC is a quad with 2 GPU clients and 1 SMP client running.

Yes, I did see a small increase in the SMP PPD (~200), BUT I also see a small decrease in each of the GPU clients' PPD (~200 each)

So, buyer beware. Now mine is the second post I've read that indicated a small PPD DEcrease when switching to the 180.70 drivers.

-Sid

 

[Gravity]

Hmm...well, my practice of letting Sid be the guinie pig, I mean, well, er....pioneer, pays off again! Unless there's some added benefit I don't see here, I'll stick with the current driver.

Keep folding,

Gravity

 

[Insidious]

Originally posted by: Gravity
Hmm...well, my practice of letting Sid be the guinie pig, I mean, well, er....pioneer, pays off again! Unless there's some added benefit I don't see here, I'll stick with the current driver.

Keep folding,

Gravity

I think with one GPU, and one SMP on a dual core PC, it might work out well (WinXP). Most Dual core machines don't have enough CPU left for an SMP if a GPU is running. I have to run a CLI with the SMP on my Opty (X2). I am going to try the 180.70 drivers on it and see.

(I did roll the drivers on the quad video(s) back to the 178.24)

-Sid

 

[Markfw]

I tried just a regular client (not even smp) and one gpu, and I lost 300 ppd, and didn;t get that much with the client, so I gave up

 

[Foxery]

I take it you mean WinXP? Vista is completely different, and was already pretty low on CPU use.

-SMP # flag: For future use. Setting it to anything but 4 will not function with A1 core WUs.

 

[geokilla]

Yea, I meant Windows XP. Sorry for that mess up. But the little to no CPU usage for GPU folding means a lot to me.

I ALMOST finished a WU last night. Turned off my computer when it completed 95%. Turned it back on when I got home an hour ago, and the WU was gone. Turns out that in order to run MPICH SMP, my computer needs to be on 24/7, which I can't do. I now made the switch to the DEINO SMP client, and everything seems to be all good.

Oh a dual core Windows XP machine, SMP + GPU is the way to go. And the Forceware 180.60 does it better than the 180.70 imo. By better, I mean that there's less lag when I'm switching between my MSN convos, Firefox, etc.

@Fox. What do you mean it won't function with A1 core WUs? I have it set to -smp 2 and according to Task Manager, the four FahCore_a1.exe use a total of around 93% of my CPU.

 

[VirtualLarry]

So does that mean, that with this new 180.70 forceware driver, I could run FOUR 9600GSOs, with a GPU client for each one, on a dual-core CPU, under XP? That would save me having to build the VGA dummy loads in order to run this on Vista. I would much prefer NOT to install Vista; I hate it.

 

[Insidious]

I think you can. You will be the guinea pig for this one.

be sure to create and set a windows system environment variable

NV_FAH_CPU_AFFINITY

set the value equal to 0

-Sid

(PM me if you are not familiar with how to set a system environment variable in windows XP)

 

[VirtualLarry]

Ok, I got my rig built. See http://forums.anandtech.com/me...ht_key=y&keyword1=quad

I didn't set the NV_FAH_CPU_AFFINITY env. variable, but I did select "Do NOT lock CPU to core" for each GPU shortcut. So I don't know, perhaps that does the same thing.

CPU usage is literally non-existant! 0% nearly all of the time, even with the GPUs running at what seems to be full-speed. Amazing.

 

[Insidious]

Originally posted by: VirtualLarry
.
.
.
CPU usage is literally non-existant! 0% nearly all of the time, even with the GPUs running at what seems to be full-speed. Amazing.

Sounds like you are using one of the 180.xx driver versions which do really help keep the GPU off the CPU cycles. I think it will make the affinity variable needless so it sounds like you are good to go. (I am experimenting with the 180.60 driver so see if I can run an SMP with the GPU on my Opteron 170 dual core to get more PPD than I was getting with a CLI and an SMP.)

:thumbsup:

-Sid

 

[geokilla]

Is it just me or did PPD drop a lot? Some people are reporting a small decrease with the 180.60 drivers, but I'm losing about 300PPD. Gonna try tinkering with my OC settings and see what I get. I think because of the low CPU usage, the speed of the GPU core became a bit more important now.

Project : 5506
Core : GPUv2 Gromacs
Frames : 100
Credit : 480


-- GPU --

Min. Time / Frame : 1mn 38s - 4231.84 ppd
Avg. Time / Frame : 1mn 40s - 4147.20 ppd
Cur. Time / Frame : 1mn 54s - 3637.89 ppd
R3F. Time / Frame : 1mn 54s - 3637.89 ppd
Eff. Time / Frame : 11mn 29s - 601.92 ppd

 

[Insidious]

The drop is hardware dependent. The less shaders a card has, and the narrower memory bus it has will cause a greater decrease. 300PPD on a 9600GT is about right.

-Sid

unless you are trying to free up your CPU for a client on a dual core (WinXP) PC, you want to use 178.24 drivers for max PPD.

I upped the drivers on my Opty 170 to 180.60 hoping to be able to run an SMP client alongside the GPU (8800GTS 512) instead of the CLI that I have been running. The increase in PPD due to SMP vs. CLI was almost totally offset (edit: make that about 1/2)by the decrease in PPD from the GPU client.