FGRP4-SSE2 hogging scheduler vs SETI
5 Apr 2015 4:33:58 UTC
Topic 198035
(moderation:
I have noticed that my Einstein FGRP4-SSE2 are getting all my scheduler time (running 6 CPUs flat out), and in particular with my SETI v7 7.00 not receiving any time, not even starting. My SETI Astropulse v7 7.05 are getting time but they are running on my nvidia GPU. I have been observing this over about a week.
Earlier in the week I did have some SETI v7 7.00 run after being excluded for several days; it was not clear why those suddenly started; while they were active they shared CPU use normally with Einstein. It was not a case of EDF getting activated: if it was then the SETI would have taken over all of the CPUs.
The FGRP4-SSE2 tasks that are currently active for me are all downloaded later than the oldest SETI tasks that are not getting scheduled. They do have earlier report deadlines than any of those tasks, but report deadlines is not the active factor: if it was then at least one of the S6BucketFU1 would be running as those are due before any of the active FGRP4-SSE2.
The S6BucketFU1 were downloaded after after the active FGRP4-SSE2 tasks, and are not due for a few more days, so the fact that they have not run yet is in accordance with my understanding of scheduling (but I will still be keeping my eye on them to ensure they run in time.)
I have more credits achieved with Einstein than I do with SETI, and my Average Work Done is about 3 times higher on Einstein than SETI at the moment. So Einstein can't be getting scheduled to try to "catch up" to SETI. I also have Einstein set for a lower share, 100 vs the 125 I have set for SETI. All my other projects have been set at No Further Tasks for the last couple of weeks.
The Einstein FGRP4-SSE2 processes are not stuck: they are progressing fine, completing, new ones getting started. I have done a couple of reboots during this period, for various reasons, so it is not stuck processes. Also, I upgraded BOINC during the time, so it is not a property of a single BOINC version.
Any ideas why FGRP4-SSE2 are consistently being handed the scheduler and SETI CPU use is (mostly) being edge out?
OS-X Yosemite, BOINC Manager 7.4.42, FGRP4 1.05
FGRP4-SSE2 hogging scheduler vs SETI
)
It's probably because of the resource share between the two projects and the transition by Boinc to use REC (Recent estimated credit) instead of debt for calculating the given resource share. You have the resource share set to almost equal and Seti gets to use the GPU by it self. So Boinc schedules Einstein to use the CPU. As Seti has a little higher resource share Boinc downloads the occasional CPU task, but as you see it don't get scheduled to run for quite some time, might even be that it has to wait until EDF mode sets in.
For more quite technical info see this page. Your case seems to match "Example 1" on that page.
Thank you for the
)
Thank you for the feedback.
I can see now, looking at the document you linked to, how one project could be given all the CPU and another all of the GPU. The process seems like a design flaw, though: BOINC should not be relying on EDF to get the CPU tasks done for the project that also has GPU apps.
In the particular case of my system, over the last two weeks, the granted credit for Einstein has been about 520000, whereas the granted credit for SETI (including GPU jobs) has been about 14000, which is 1/30th of that for Einstein.
Before the SETI Astropulse (GPU) jobs were downloaded, SETI was not being scheduled, at least not since whatever it was that triggered and ran through some of the older SETI tasks about 5 days ago. Likewise now that the SETI Astropulse (GPU) jobs are complete, the SETI CPU jobs are not being scheduled.
Within the framework offered in that design document, the only hypothesis I can offer at the moment (besides "bug") would be if BOINC's estimate of the credit for SETI jobs is very different than actual credits. SETI is not exactly generous with credits.
The "Recent estimated credit"
)
The "Recent estimated credit" is not and will probably never be the same as RAC (Recent average credit) unless all projects start using a version of "Credit new" that actually works.
If you were to run both CPU and GPU tasks from both projects then you would probably see Boinc scheduling task from both projects more even on both resources.