I was wondering if the BOINC project has to be GPU enabled to use the GPU procesing. The reason I ask is because I participate in the "World Community Grid", and it would be nice to harness the power of the GPUs for that project.

Additionally, is it possible to run one BOINC project on the GPU and one on the CPU(s). If this is the case I might start generating maximum heat and put gpugrid.net on the GPU and WCG on the CPU.

Additionally, is it possible to run one BOINC project on the GPU and one on the CPU(s). If this is the case I might start generating maximum heat and put gpugrid.net on the GPU and WCG on the CPU.

In the moment (Boinc 6.3.10) the boinc-client assigns one WU to each CPU, regardless if this WU is a GPU-WU or a CPU-WU.
To meet your target, you have to fool the boinc-client by announcing one additional CPU, e.g. Host 8166 with 5 CPU's.
On this host i am crunching 4 CPU-WU's (PrimeGrid) and one PS3-WU; the PrimeGrid-WU's with 99% per CPU and PS3 with about 5%.

To meet your target, you have to fool the boinc-client by announcing one additional CPU, e.g. Host 8166 with 5 CPU's.
On this host i am crunching 4 CPU-WU's (PrimeGrid) and one PS3-WU; the PrimeGrid-WU's with 99% per CPU and PS3 with about 5%.

???? How did you manage to do that??

???? How did you manage to do that??

1. check out the sources.
2. file boinc/client/hostinfo_unix.C in line 764: p_ncpus = p_ncpus + 1; (+=1)
3. compile and use it (linux version)

for a windows version #2 should be done in ...hostinfo_win.C

Additionally, is it possible to run one BOINC project on the GPU and one on the CPU(s). If this is the case I might start generating maximum heat and put gpugrid.net on the GPU and WCG on the CPU.

In the moment (Boinc 6.3.10) the boinc-client assigns one WU to each CPU, regardless if this WU is a GPU-WU or a CPU-WU.
To meet your target, you have to fool the boinc-client by announcing one additional CPU, e.g. Host 8166 with 5 CPU's.
On this host i am crunching 4 CPU-WU's (PrimeGrid) and one PS3-WU; the PrimeGrid-WU's with 99% per CPU and PS3 with about 5%.

Hi,
is the GPU application any slower doing this?

gdf

Hi,
is the GPU application any slower doing this?

gdf

In 21 hours and 26 minutes my 8800-Card has made 1 WU and 36,5% of the next WU. That matches exactly the average time of 54200 seconds per WU with the 6.42 application.
Perhaps (?) the CPU-usage of 5%-6% is one percent higher than before, but this is not significant. It is a great success that you can use (nearly) one more CPU for other projects.

1. check out the sources.
2. file boinc/client/hostinfo_unix.C in line 764: p_ncpus = p_ncpus + 1; (+=1)
3. compile and use it (linux version)

for a windows version #2 should be done in ...hostinfo_win.C

I searched for hostinfo_win.C in the boinc/client path, but was not able to locate it. Any ideas? I installed as protected so BOINC is running as a user account. Could this have any impact on me being able to locate this file?

1. check out the sources.

svn co http://boinc.berkeley.edu/svn/trunk/boinc

Have you done this? Or what do you mean with "installed protected mode"? "Check out" is not an installation!

He's talking about downloading the BOINC source files, editing a line and compiling it yourself. I don't want to be rude, but if you didn't get this it's probably not meant for you ;)

MrS

---

Scanning for our furry friends since Jan 2002

He's talking about downloading the BOINC source files, editing a line and compiling it yourself. I don't want to be rude, but if you didn't get this it's probably not meant for you ;)

MrS

Please do not change the BOINC code if possible. It will be difficult to debug issues for us if there are different policies in place.

If this user test confirms that the GPU application runs at the same speed, we will change the server policy to release the processor.

If the gpu application is slower with the current boinc client and all cpu taken, you will be losing credits instead of increasing it, because you are slowing down the fastest device on the computer. Just wait a couple of days of testing with this user machine.

gdf

Hi gdf,

the only modification is: p_ncpus = p_ncpus + 1;

And this works (for me) very well...

I think GDF is right. We don't want people to start messing with compilers and learning all the tricks and glitches the hard way by causing errors in a beta project. The only realistic option would be to distribute the modded BOINC, but since 6.3.11 is not that far away this probably isn't worth the hassle.

MrS

---

Scanning for our furry friends since Jan 2002

Hi gdf,

the only modification is: p_ncpus = p_ncpus + 1;

And this works (for me) very well...

That's good. If we see that works, we will release the cpu. The new boinc client is designed to make sure that the gpu application is not suspended while time-slicing the cpu between the other tasks. This can potentially making the gpu application slower (it was so when we tested months ago, but we were also using 100% of cpu).

gdf

Hi gdf,

the only modification is: p_ncpus = p_ncpus + 1;

And this works (for me) very well...

That's good. If we see that works, we will release the cpu. The new boinc client is designed to make sure that the gpu application is not suspended while time-slicing the cpu between the other tasks. This can potentially making the gpu application slower (it was so when we tested months ago, but we were also using 100% of cpu).

gdf

Also, now that the cpu is free, we may end up using it anyway for doing other computing tasks while GPU is crunching. This will give additional credits of course (work in progress).

gdf

There are other, more easier ways to let BOINC use more processors than the system actually has. Don't ask... Figure it out, its not difficult or far away!

I configured my C2D to use 3 cores and my Quad to use 5 and right now each crunches n+1 processes. So far this looks good, times seem not to drop.

My Quad with the 9800GT used to take 58600 seconds for one unit with 6.42. With 6.44 I can say that it takes somewhere in the range of 16-17 hours, but its a little difficult to say, as wall time for that unit can not be found under Linux. Time per step was 69.4ms.
Yesterday I switched to 5 tasks on that hosts and the unit that was just reported took no longer than 16,5 to max 17 hours at 70.754ms per step.

So it was a little slower, but not much. The 2% slowdown I'm willing to take, when I can have that core free.

On the C2D with 8800GTS the time per step increased from 52.8ms to 53.8ms.

In our forum other users (windows) have reported that this increases their crunching time for one unit a lot. So for Windows users forcing one extra core / thread is not expedient.
Might be some scheduler issue in the operating system, I don't know...

But for me and Stoffel it seems to run fine so far.
Lets monitor this for some days. I set my client to return results immediately, so that I can use the upload time as run time indicator. Maybe Linux is already ready for running the GPU thread in parrallel without (or just very small) impact to the run times... :)

He's talking about downloading the BOINC source files, editing a line and compiling it yourself. I don't want to be rude, but if you didn't get this it's probably not meant for you ;)

MrS

Over my head. :-) Just running BOINC to help out science. Don't want to get too involved or create issues by tweaking something that possibly will be a feature in a future version of BOINC. I'll keep crunching along -- thanks for the response.

???? How did you manage to do that??

1. check out the sources.
2. file boinc/client/hostinfo_unix.C in line 764: p_ncpus = p_ncpus + 1; (+=1)
3. compile and use it (linux version)

for a windows version #2 should be done in ...hostinfo_win.C

can be done both in linux and windows in cc_config.xml with attribute
<ncpus>

see here for documentaiton of cc_config
http://boinc.berkeley.edu/wiki/Client_configuration

But it slows down processing ALOT (maybe only on windows, problem with the scheduler?) 31k seconds before and 55k seconds after the modification.

We tracked down the issue (it was not the firewall causing the increase in computation time).

???? How did you manage to do that??

1. check out the sources.
2. file boinc/client/hostinfo_unix.C in line 764: p_ncpus = p_ncpus + 1; (+=1)
3. compile and use it (linux version)

for a windows version #2 should be done in ...hostinfo_win.C

can be done both in linux and windows in cc_config.xml with attribute
<ncpus>

see here for documentaiton of cc_config
http://boinc.berkeley.edu/wiki/Client_configuration

But it slows down processing ALOT (maybe only on windows, problem with the scheduler?) 31k seconds before and 55k seconds after the modification.

We tracked down the issue (it was not the firewall causing the increase in computation time).

This <ncpus> is meant for diagnostics only and not meant to be used as a normal way to run boinc.

I would advise to wait for an official release.

This <ncpus> is meant for diagnostics only and not meant to be used as a normal way to run boinc.

I would advise to wait for an official release.

Well, this is what I use currently under Windows after Stoffels hint at our team forum.
It works pretty well for me, already crunching the second WU with it, while the 2 main cpus doing PrimeGrid.

---

Member of BOINC@Heidelberg and ATA!

things go bad when a project goes into "High Priority Mode" when you have it configured that way. And it doesn't stop BOINC from running 3 WUs from primgrid or any other project.