PPD Results & Suggestions

[Eveofwar]

AMD Athlon 6000+ 64 X2 @ 3.00GHz (Stock)
nVidia GeForce 9800GT @ 650MHz GPU/950MHz Memory/1620MHz Shader (Stock)
2GB DDR2 RAM

Currently I'm getting ~3246 PPD according to FahMon, using just the GPU2 core set to allow assignments greater then 10MB, lowest possible core priority, 100% CPU usage, 12 minute checkpoints, NOT locking CPU to specific cores, and using the -advmethods -forceasm parameters.

I'm using the 6.24 GPU2 System tray client with special viewer for NVIDIA GPU's and only see my GPU temperatures at 60 celcius on average with fan @ 75%, below are my current benchmarks according to FahMon.

Project : 5751
Core : GPUv2 Gromacs
Frames : 100
Credit : 511


-- MAIN --

Min. Time / Frame : 2mn 13s - 3319.58 ppd
Avg. Time / Frame : 2mn 13s - 3319.58 ppd
Cur. Time / Frame : 2mn 13s - 3319.58 ppd
R3F. Time / Frame : 2mn 13s - 3319.58 ppd

I was wondering if there's any safe overclocking I can do with the nVidia Control Panel to increase my PPD ? I've tried running my current GPU2 client along with the SMP version via VMware, but I notice the SMP does barely any work, and greatly reduces my PPD on the GPU2 client unless I set the VMware application to the alternate affinity. So I don't think the multi-client method would work well for me, unless I run the SMP client on a single CPU core without using VMware and the GPU2 client on a single CPU core.

Would I notice more PPD output by using the GPU2 console version as opposed to the GPU2 GUI version ? Either one is possible for me, I'm just looking for maximum PPD output.

So fellow folders, what's the best possible PPD output for me with my current setup ?

[DanGe]

lowest possible core priority, 100% CPU usage

For one thing, set your GPU client's priority to "Slightly higher". The GPU client needs some, but very little, CPU resources to feed the GPU. Keep in mind, though, this setting only affects the CPU side, but if the SMP client takes over all the CPU resources, the GPU client will not be able to do much as it can't give data to the GPU.
By the way, -forceasm only works for CPU clients.

with special viewer for NVIDIA GPU's

Would I notice more PPD output by using the GPU2 console version as opposed to the GPU2 GUI version

FYI: Since many people do this, I will tell you in advance: although you can use the systray client, do not use the viewer since it takes up extra GPU and CPU resources. Also, using a console version will not necessarily speed things up unless you keep the viewer on the systray client open all the time.

I was wondering if there's any safe overclocking I can do with the nVidia Control Panel to increase my PPD

I don't think the NVIDIA control panel does any overclocking. I suggest you use RivaTuner for that. Remember to check your log file frequently for UNSTABLE_MACHINE or EARLY_UNIT_END when you're trying out a new OC setting. Windows/gaming-OK overclock settings are not necessarily FAH-OK settings. When you try out new OC settings, remove -advmethods. When you find a stable setting, you can add it again.

NOT locking CPU to specific cores

Someone correct me if needed, but I think you should keep the default setting (not checked).

[Eveofwar]

For one thing, set your GPU client's priority to "Slightly higher". The GPU client needs some, but very little, CPU resources to feed the GPU. Keep in mind, though, this setting only affects the CPU side, but if the SMP client takes over all the CPU resources, the GPU client will not be able to do much as it can't give data to the GPU.

Alright, it's set to "slightly higher" now. Also, you're saying with that then that running the GPU + SMP clients on the same machine is pointless then ? Even with the GPU set to "slightly higher", meaning that it should take priority over the SMP client.

FYI: Since many people do this, I will tell you in advance: although you can use the systray client, do not use the viewer since it takes up extra GPU and CPU resources. Also, using a console version will not necessarily speed things up unless you keep the viewer on the systray client open all the time.

I don't run the viewer, I just leave the icon alone in the systray, I don't need visuals to know it works, I just like big numbers =]

I don't think the NVIDIA control panel does any overclocking. I suggest you use RivaTuner for that. Remember to check your log file frequently for UNSTABLE_MACHINE or EARLY_UNIT_END when you're trying out a new OC setting. Windows/gaming-OK overclock settings are not necessarily FAH-OK settings. When you try out new OC settings, remove -advmethods. When you find a stable setting, you can add it again.

The NVIDIA control panel gives me the ability to change the core and memory clock speeds, but not the shader clock. I downloaded the RivaTuner, and was unsure of stable/good settings to use it at, so I left everything at default. Would you happen to have some suggestions as far as stable OC settings ?

[Eveofwar]

Also, is it resonable for the SMP client to take ~25 minutes to complete 1% ?

[07:29:49] (0 percent)
[07:44:50] Timered checkpoint triggered.
[07:53:57] Writing local files
[07:53:58] Completed 5000 out of 500000 steps (1 percent)

[bruce]

Alright, it's set to "slightly higher" now. Also, you're saying with that then that running the GPU + SMP clients on the same machine is pointless then ? Even with the GPU set to "slightly higher", meaning that it should take priority over the SMP client.

Not at all. The GPU client gets a lot of work done with the resources of the GPU but it needs a very small amout of CPU resources to keep it going. If the SMP client runs successfully on your CPUs without the GPU client running, it will run very nearly as fast after the GPU client steals close to zero processing from your CPUs.

[bruce]

Also, is it resonable for the SMP client to take ~25 minutes to complete 1% ?

[07:29:49] (0 percent)
[07:44:50] Timered checkpoint triggered.
[07:53:57] Writing local files
[07:53:58] Completed 5000 out of 500000 steps (1 percent)

Maybe ... maybe not. It's impossible to say unless we know what WU was assigned to you and the capbilites of your CPUs.

25 minutes per percent means that the WU will finish in less than 2 days. Many of the SMP WUs have deadlines of 3 or 4 days so you're probably OK but many are shorter. Look for your project number here: http://fah-web.stanford.edu/psummary.html

[Eveofwar]

Not at all. The GPU client gets a lot of work done with the resources of the GPU but it needs a very small amout of CPU resources to keep it going. If the SMP client runs successfully on your CPUs without the GPU client running, it will run very nearly as fast after the GPU client steals close to zero processing from your CPUs.

Awesome !

The GPU2 console version is installed and running, and just got the SMP version running as well. The SMP seems fairly slower, but also has a greater amount of steps, so I assume it's evened out. I'm currently running them both simultaneously. CPU temperature is at 138 degrees @ 100% load, and the GPU stays at around 122 degrees. The CPU temperature is a bit high I'm sure, but I'm using the stock AMD heatsink with it, until I find a good one that fits my ASUS M2A-VM motherboard.

bruce, would you happen to have good/stable OC settings for my GeForce 9800GT to be used in RivaTuner ?

[Eveofwar]

Maybe ... maybe not. It's impossible to say unless we know what WU was assigned to you and the capbilites of your CPUs.

25 minutes per percent means that the WU will finish in less than 2 days. Many of the SMP WUs have deadlines of 3 or 4 days so you're probably OK. Look for your project number here: http://fah-web.stanford.edu/psummary.html

Project 2653, 3 days preferred, with 4 days deadline. Roughly 41 hours @ 25mins/percent...puts me at about 1.7 days if I leave it running 24/7, I doubt I can make that deadline hmm...

Do you get partial credit for the data submitted come deadline time ?

[bruce]

bruce, would you happen to have good/stable OC settings for my GeForce 9800GT to be used in RivaTuner ?

There are lots of other people that can give you better advice on that than I can.

If a WU passes the final deadline you get zero credit. If it is completed after the Preferred deadline, you'll get full credit but it will be considered "lost" and sent to someone else so the work gets duplicate, which is a bit of a waste. If you can't get work done by the preferred deadline, you probably don't have enough CPU (or enough hours per day) to do useful scientific work with that client.

[Eveofwar]

There are lots of other people that can give you better advice on that than I can.

If a WU passes the final deadline you get zero credit. If it is completed after the Preferred deadline, you'll get full credit but it will be considered "lost" and sent to someone else so the work gets duplicate, which is a bit of a waste. If you can't get work done by the preferred deadline, you probably don't have enough CPU (or enough hours per day) to do useful scientific work with that client.

Well then, would it be better to set the data file size to "small" instead of "normal" then ?

[bruce]

Well then, would it be better to set the data file size to "small" instead of "normal" then ?

It's not clear whether there are any "small" SMP WUs, but even if there are, the size is related to upload and download sizes and memory demands. The relationship between the processing time and the deadline may be very similar. SMP is designed to run 24x7 on fairly robust computers.

I don't see where you described your hardware, but instead of running SMP with the GPU, you can install one classic CPU client per CPU. The deadlines will not be a problem, but the total points will be somewhat lower.

EDIT: Never mind. I see AMD Athlon 6000+ 64 X2 @ 3.00GHz (Stock) in your first post.

[Eveofwar]

It's not clear whether there are any "small" SMP WUs, but even if there are, the size is related to upload and download sizes and memory demands. The relationship between the processing time and the deadline may be very similar. SMP is designed to run 24x7 on fairly robust computers.

I don't see where you described your hardware, but instead of running SMP with the GPU, you can install one classic CPU client per CPU. The deadlines will not be a problem, but the total points will be somewhat lower.

EDIT: Never mind. I see AMD Athlon 6000+ 64 X2 @ 3.00GHz (Stock) in your first post.

Appreciate all your help.

I run the PC maybe 12-16 hours per day, and I can usually get a GPU WU done in about 4-6 hours, so it's nice.

So, you think 2x UNI-processor console clients and set them to each CPU core ? That would probably be best. Even just one would work too I assume.

Thank you, I'll try that out. Is there a switch to set the CPU core to use (whether it's 0 or 1) ?

[bruce]

One or two UNI-processor clients will be fine. It (or they) will use either processor and that's not something that's going to matter anyway. Assigning a specific CPU is called setting "affinity" and I recommend you just ignore the whole concept.

Each client needs a unique value for MachineID, but MachineID is not related to processor number.

By the way, I should have said something sooner. Welcome to the FoldingForum.

[Eveofwar]

One or two UNI-processor clients will be fine. It (or they) will use either processor and that's not something that's going to matter anyway. Assigning a specific CPU is called setting "affinity" and I recommend you just ignore the whole concept.

Each client needs a unique value for MachineID, but MachineID is not related to processor number.

By the way, I should have said something sooner. Welcome to the FoldingForum.

Thank you for all the help and the welcome =]

I'm very computer literate, programming in VB6/C++ so I know what I'm doing for the most part.

Currently I have the GPU2 console + 2x UNI running with WinAFC spreading them between the two CPU cores...hopefully this will provide maximum ouput for my current setup @ 12-16 hours a day.

[Eveofwar]

Sidenote - I don't notice much difference in processing time between the 2x UNI and the SMP...the UNI's are doing 1% per 25 minutes EACH...and at about half the points per WU. My conclusion is to stick to the GPU2 + SMP.