You might lose up 100% of performance of 256 GPU cores to save 1 CPU core!

Comments?

gdf

The reason the CPU performance is important to many of is, well, look down at my sig. We run other applications, most of which don't run on GPUs. Many, many problems simply don't lend themselves to parallelization very well.

While helping with medical research is fine and noble, so are other tasks such as preventing the spread of Malaria or keeping an asteroid from falling on my head. One of the criteria which I use to decide whether or not to run a project is whether it will interfere with other projects I run. A GPU-based project is already consuming a highly valuable resource. I don't want it to also gobble up the a CPU core that would otherwise be used by projects that can't run on the GPU. Likewise, I won't run something like Einstein@Home's APB1/2 GPU task because it wastes the GPU, which could be put to better use by projects such as GPUGRID.

---

Want to find one of the largest known primes? Try PrimeGrid. Or help cure disease at WCG.

Running MW on a HD4870 and a C2Q. If I set BOINC 6.10.29 to use 75% CPU is launches 3 CPU tasks and one MW@ATI. This way performance is much better than at 4 CPU + 1 MW, even though MW itself uses little CPU. The catch here is that it needs cpu support often and in precise intervals. So effectively you have to dedicate one core here as well.. or live with a slower GPU.

This hasn't been my experience at all. I'm running 6 HD4770 ATI cards on 4 machines. MW tasks reliably take 210 seconds CPU time and 212 seconds elapsed time no matter what project is running on the 4 CPU cores (Athlon II 620, Win7-64, ATI v9.12). So thought I'd give it a test to check out your theory. The machine I tested is running 4 instances of SIMAP on the CPU (10 WUs run each way). Times:

MW + 4 instances SIMAP: 3:30 CPU time --- 3:32 elapsed time
MW + 0 instances SIMAP: 3:29 CPU time --- 3:31 elapsed time

There you have it. The dedicated quad with nothing running but the OS saves 1 second or about .47%, not much at all. As a trade off each CPU core pumps out a SIMAP WU every 28 minutes.

If your results differ much from these you might try adding -b to your MW app_info.xml commandline parameters:-)

Good news. We have found the problem of the hanging.
Ready for 6.07.

gdf

Edit. Uploaded.

Should we abort the v6.06 WUs that haven't started yet?

There was a problem 7.07. So, we removed for now.

gdf

Beyond, thanks for that post. My test were done a few months back, can't remember if it was 0.20 or 0.19. Now I'm running a C2Q, Win7 64, Cat 9.11 and MW 0.21 without app_info.

MW + 3 instances SIMAP: 170s CPU time --- 171.7 +/- 0.3 elapsed time
MW + 4 instances SIMAP: 173s CPU time --- 175 - 176 elapsed time

That's much better than last time I checked and just about 2% slower (-> 2k credits/day). Would you mind sending me your app_info via PM?

MrS

---

Scanning for our furry friends since Jan 2002

I have been trying to track the various beta WUs
but the results seem to be all over the place, meaning that sometimes with the CPOU fully utilized by other BOINC projects (WCG, Einstein, Climate Prediction, Seti) the betas seems to run about the same as 6.71. Other times they process very quickly with a big decrease in runtime (-55%) , and finally I have had a couple instances where they ran worse than usual (+55%). I have seen all three scenarios on both my GTX285 and on my GTX295. I'll keep trying to find the magic combination.
I have one running right now that is displaying the "elapsed time", "progressbar", and the "estimated time to completion" accurately (boicmanager 6.2.28, WCG edition).
ps If you added the ability to sort the results page by Sent Time and also by Returned Time it would make tracking this stuff much easier because your taskIDs are no longer in sequence.

---

Thanks - Steve

The timing of 6.05 and 6.05 could be wrong as there was a problem in the restarting. Only the fast times are correct.

Just uploaded acemdbeta6.08.

Enjoy.

gdf
EDIT. Abort any job with acemdbeta6.06 still running.

Beyond, thanks for that post. My test were done a few months back, can't remember if it was 0.20 or 0.19. Now I'm running a C2Q, Win7 64, Cat 9.11 and MW 0.21 without app_info.

MW + 3 instances SIMAP: 170s CPU time --- 171.7 +/- 0.3 elapsed time
MW + 4 instances SIMAP: 173s CPU time --- 175 - 176 elapsed time

That's much better than last time I checked and just about 2% slower (-> 2k credits/day). Would you mind sending me your app_info via PM?

MrS

You should have a PM...

Also sorry, made a typo in the previous post. The commandline should be:

<cmdline>b-1</cmdline>

not -b

That should get you a bit more GPU usage :-)

A problem with an ACEMD beta version 6.06 (cuda) workunit:

At 13.791% progress, this error message was shown on the
screen; progress then stopped, but the elapsed time kept
rising.

acemd_6.06_windows.intelx86__cuda has
stopped working

A problem caused the program to stop working correctly.
Windows will close the program and notify you if a solution is
available.

(A button labelled) Close program

The workunit was L2-TONI_TEST0202-0-RND6466_2.

http://www.gpugrid.net/result.php?resultid=1805197

After clicking on the button, the error message window
disappeared, with no apparant immediate change in the way
the workunit was running; but within a minute, a computation
error was reported and the GPU was left idle for a few minutes
it took to get another GPUGRID workunit.

<core_client_version>6.10.18</core_client_version>
<![CDATA[
<message>
The extended attributes are inconsistent. (0xff) - exit code 255 (0xff)
</message>
<stderr_txt>
# There is 1 device supporting CUDA
# Device 0: "GeForce 9800 GT"
# Clock rate: 1.50 GHz
# Total amount of global memory: 1073741824 bytes
# Number of multiprocessors: 14
# Number of cores: 112
MDIO ERROR: cannot open file "restart.coor"
SWAN : FATAL : Failure executing kernel sync [frc_sum_kernel] [999]
Assertion failed: 0, file ../swan/swanlib_nv.cpp, line 203

This application has requested the Runtime to terminate it in an unusual way.
Please contact the application's support team for more information.

</stderr_txt>

BOINC 6.10.18
64-bit Windows Vista SP2

6.08 is looking very nice!!!
Windows XP Pro 32 bit, Boinc Manager 6.2.28 (WCG edition)
Core i7-920, GTX285 w/ 191.07 driver

Constant CPU load with 8 CPU WUs for WCG (all different types, HFCC, HCMD2, HPF2, NRFW, FAAH) and also a couple of Einstein GW6 WUs.

GPUGrid 6.08
Elapsed Time, Progressbar, To Completion all working properly.
The "Result" column says .29 CPU but Task Manager barely registers any CPU usage for this WU at all. After 3 hours 50 minutes elapsed time the WU "Properties" reports CPU usage at 10 minutes 36 seconds.
Currently this WU is 92.416% complete and looks to be done in another 18 minutes. This GPU typically proceses a GIANNI_BIND in 6 hours.

Let me know if there is anything else I can do to help.

Best regards,
Steve

<edit>
My first 6.08 completion
runtime = 15,021.08
cputime = 692.70

It looks like we still get a nice improvement (30%) even when all CPU cores are loaded.

If I am around and see another 6.08 show up on thios machine I will shut down a couple of CPU tasks and see if there is any further improvement unless GDF can say that test would be a waste of time.

</edit>

Tom Philippart wrote:

I'm not talking about credits, in that perspective the new solution would outweight the old, but I'm talking in terms of output.

You're measuring with two different gauges here. A reduction of your WCG output by a factor of 1.33 does count, but a GPU-Grid output increase by a factor of 1.66 does not count? You must not count it purely in terms of "cores", as that can be quite misleading. Or is one core of a Celeron 266 MHz worth as much as one of an i7? Or a GTX260 or a GTX380?
MrS

As I said, as nearly no one is running gpugrid as their only project, I want to run it on top of the regular production: stay with the normal output of my primary project and only take minor losses to contribute with the gpu. This is a higher output than running nothing on the gpu. My point is that I want to run it "on top" of my primary project. I don't want to touch the output of my primary focus.

That's also the reason why I didn't run the folding ati client until they build the workaround to reduce cpu usage

Anyways I think aiming to reduce cpu usage, by for instance using buffering techniques (as far as it's possible) is the best solution. Einstein@home is also working on reducing the cpu usage of their app.

Had two 6.08 Beta's fail within 3 seconds: Beta WU 1 and Beta WU 2

Both had the message "SWAN: FATAL : swanMalloc failed"

-----------------EDIT--------------------

May have been just something weird with BOINC switching tasks or something. I rebooted and two new 6.08 Beta's are crunching along fine now.

Please go ahead with the test.
You have too look at the elapsed time or in the result output ms/step to compare WUs of the same name. The unit we are sending around is not a GIANNI_BIND, but they are all the same for the beta app.

gdf

6.08 is looking very nice!!!
Windows XP Pro 32 bit, Boinc Manager 6.2.28 (WCG edition)
Core i7-920, GTX285 w/ 191.07 driver

Constant CPU load with 8 CPU WUs for WCG (all different types, HFCC, HCMD2, HPF2, NRFW, FAAH) and also a couple of Einstein GW6 WUs.

GPUGrid 6.08
Elapsed Time, Progressbar, To Completion all working properly.
The "Result" column says .29 CPU but Task Manager barely registers any CPU usage for this WU at all. After 3 hours 50 minutes elapsed time the WU "Properties" reports CPU usage at 10 minutes 36 seconds.
Currently this WU is 92.416% complete and looks to be done in another 18 minutes. This GPU typically proceses a GIANNI_BIND in 6 hours.

Let me know if there is anything else I can do to help.

Best regards,
Steve


My first 6.08 completion
runtime = 15,021.08
cputime = 692.70

It looks like we still get a nice improvement (30%) even when all CPU cores are loaded.

If I am around and see another 6.08 show up on thios machine I will shut down a couple of CPU tasks and see if there is any further improvement unless GDF can say that test would be a waste of time.

GTX260 (754 MHz, Shader: 1568 MHz, Speicher: 1211 MHz) (WinXP_32, Kentsfield, 4xSpinhenge@home)


ACEMD beta version v6.08:

Run time 16629.203125
CPU time 608.4688
Time per step: 26.600 ms


ACEMD beta version v6.06:

Run time 17283.195
CPU time 2136.792
Time per step: 27.649 ms

I got a 6.08 result which, unfortunately, was consuming a significant portion of a C2Q core (7% on task manager, so about 28% of the CPU time of one core.) Typical usage is 0 to 1 percent in Task Manager.

I aborted the WU, and also turned off the beta WUs for now.

CPU usage is not important anymore. Just look at the time/step with and without load. So far, it seems very good.

gdf

Running three 6.08 units and eight WCG units together on one machine using 275's and looks fine. CPU load for 6.08 units is 17-18% and WCG units average 89%. GPUgrid tasks showing 60% performance increase.
Job well done GDF and team.

I got a 6.08 result which, unfortunately, was consuming a significant portion of a C2Q core (7% on task manager, so about 28% of the CPU time of one core.) Typical usage is 0 to 1 percent in Task Manager.

I need to correct my own statement.

Either I'm misremebering how each project behaves, or something is different on this box today. Both are fairly likely. :)

The production (6.71) application is showing 5 to 6 percent CPU utilization in Task Manager, so the difference between the CPU usage between the two is insignificant.

Fully loaded with 8 CPU WUs, one 6.71 GPU and one 6.08 GPU.

While watching task manager on my system with an old style GTX295 (2 physical cards) I had one 6.71 WU and one 6.08 WU. I bump the Priority up to Normal on the 6.71 and it starts using about 5%-10% CPU. After watching that for a few minutes I thinking about what might be going on so I bump the beta 6.08 priority up to Normal which I noticed is at "Below Normal" (most BOINC WUs are set to Low) and the 6.71 returns to using 2%-3% (sometimes up to 5%) like it was before I bumped it to Normal. So I then push 6.71 down to Below Normal and the 6.08 starts to take up more CPU, 5%-8% and occasionally 10%-12%.

Perhaps because it is a GTX295 it can only use 1 core for both cards and not only is it competing against other processes on that core but the two WUs are competing against each other. I'll see if the behavior is any different when I have two 6.08 WUs running.
The above behavior only is apparent when running fully loaded with 8 CPU WUs running. If I reduce to 7 CPU WUs then the CPU usage stays fairly constatnt between 2-3 for each GPU WU no matter what I set the Priority to. I will see what the CPU usage is long term, but wonder if it will run slower overall because it has to wait for CPU cycles to be available which I am assuming is what the spikes to 8%-12% I am observing are.

<edit>and if you change the Priority of the WUs enough you WILL crash them. I just lost two betas this way, I will just let them run instead of changing them.</edit>

---

Thanks - Steve