When will ACEMD be used for all WU's given that (at least for Windows) it is much more efficient?

---

Radio Caroline, the world's most famous offshore pirate radio station.
Great music since April 1964. Support Radio Caroline Team -
Radio Caroline

ACEMDv2 is already used by half of the jobs. All the new workunits are going to the new application and the old ones will finish up in a month or so progressively.

gdf

But how about high CPU consumption under Linux? Is it in developers plans to reduce it or it OK in their opinioin and will remain at present level?

---

From Siberia with love!

So far, app v.604 is running fine. It has completed a GIANNI WU in 9.5 hrs vs 14.6 (about 35% faster) and 3 TONI_CAPBIND WU's in about the same time. Drawback is CPU time doubles meaning at completion of the WU, CPU time is almost exactly equal to GPU time. As questioned above, will there be a fix for this in the near future for Linux users? The increased CPU demand does negatively impact my throughput for WCG WU processing.

Platform: Phenom II x4 4GB, Fedora 11, GTS-250 1GB, 6.10.17

LINUX only:
The low-cpu version was hanging for some linux user, so we used this version which should use CPU only if it is available. The are only these two ways of synchronizing the driver. What we could do is to provide an environment variable to choose the synchronization.

gdf

Understood. I should have checked this earlier but I noticed while processing a 6.70 app, ACEMD is nice'd at 10 whereas WCG WU's are all nice'd at 19. I guess that explains the CPU availability for ACEMD WU's over the WGC counterparts.

Understood. I should have checked this earlier but I noticed while processing a 6.70 app, ACEMD is nice'd at 10 whereas WCG WU's are all nice'd at 19. I guess that explains the CPU availability for ACEMD WU's over the WGC counterparts.

Are your WCG WUs CPU or GPU? I know on Windows, BOINC runs the CPU thread of GPU tasks at a higher priority ("below normal") than it does CPU WUs ("Low"). I have never run a CUDA GPU under Linux, so I don't know if BOINC does the same thing there, but it wouldn't surprise me if it did.

I'm guessing the process priority is set by the main BOINC client, not the individual app.

The WCG only has CPU Work Units.

Boinc task priority is set by Boinc.

Are your WCG WUs CPU or GPU? I know on Windows, BOINC runs the CPU thread of GPU tasks at a higher priority ("below normal") than it does CPU WUs ("Low"). I have never run a CUDA GPU under Linux, so I don't know if BOINC does the same thing there, but it wouldn't surprise me if it did.

I'm guessing the process priority is set by the main BOINC client, not the individual app.

As SKGiven mentions above, WCG only offers CPU WU's. In Linux, nice values roughly equate to CPU priorities (-20 Highest, 19 Lowest). Since WCG is nice'd at 19, about equivalent to Windows Idle 1 (lowest priority) and the ACEMD CPU thread nice'd at 10, ACEMD has a little more CPU priority than the WCG WU's. With ACEMD1, the GPUGrid thread used about 40-45% of the total 400% CPU time (4 cores) leaving me with essentially 3.5 cores for WCG work running concurrently with GPUGrid. Now with ACEMD2 and the 35% reduction in GPU WU process time, the GPUGrid thread uses 100% leaving me with only 3 cores for WCG WU's.

Although I still run 5 jobs simultaneously and successfully, the WCG WU's time to completion are increased roughly 33%. This is not a complaint but merely an observation for the GPUGrid scientists and thanks for the posts as I wasn't sure if the app or BOINC assigned the priorities.