6.12.26: insufficient nVidia tasks

Author	Message
Joseph Stateson Volunteer tester Send message Joined: 27 Jun 08 Posts: 641	Message 38073 - Posted: 31 May 2011, 22:16:30 UTC I have one system with ATI 5850 and nVidia gtx570. For preferences, Collatz and Milkway are set to allow both ATI and nVidia tasks and PrimeGrid is set for nVidia only. As shown below, I have 44 gpu tasks ready to start, all of which are ATI. There is only 1 nVidia task running, and when it complete it gets only one more task. My work buffer is set for one (1) day. Unaccountably, I have no backlog for nVidia tasks. There should be about equal number of nVidia and ATI. I checked the servers and there are plent of WU's ready to be downloaded. Something seems wrong. from message log jstateson2quad 11171 Milkyway@home 2011-05-31 5:12:25 PM Reporting 1 completed tasks, requesting new tasks for ATI GPU 11172 Milkyway@home 2011-05-31 5:12:26 PM Scheduler request completed: got 1 new tasks 11173 PrimeGrid 2011-05-31 5:13:01 PM Computation for task pps_sr2sieve_21291464_1 finished 11174 PrimeGrid 2011-05-31 5:13:01 PM Sending scheduler request: To fetch work. 11175 PrimeGrid 2011-05-31 5:13:01 PM Requesting new tasks for NVIDIA GPU 11176 PrimeGrid 2011-05-31 5:13:03 PM Started upload of pps_sr2sieve_21291464_1_0 11177 PrimeGrid 2011-05-31 5:13:04 PM Finished upload of pps_sr2sieve_21291464_1_0 11178 PrimeGrid 2011-05-31 5:13:15 PM Scheduler request completed: got 1 new tasks 11179 PrimeGrid 2011-05-31 5:13:17 PM Starting pps_sr2sieve_21291865_1 11180 PrimeGrid 2011-05-31 5:13:17 PM Starting task pps_sr2sieve_21291865_1 using pps_sr2sieve version 139 11181 Milkyway@home 2011-05-31 5:13:30 PM Sending scheduler request: To fetch work. 11182 Milkyway@home 2011-05-31 5:13:30 PM Requesting new tasks for ATI GPU 11183 Milkyway@home 2011-05-31 5:13:31 PM Scheduler request completed: got 0 new tasks 11184 Milkyway@home 2011-05-31 5:13:31 PM No tasks sent 11185 Milkyway@home 2011-05-31 5:13:31 PM This computer has reached a limit on tasks in progress I do not see milkyway requesting any nVidia at all, and, as shown, PrimeGrid only gets 1 task every time it completes the previous. ID: 38073 ·

Pepo Send message Joined: 3 Apr 06 Posts: 547	Message 38086 - Posted: 1 Jun 2011, 9:07:56 UTC - in response to Message 38073. Could you replace your screenshot with another one, which would include the "Remaining time" column? This would show the amount of cached/available work. You are using 6.12.26 and it still seems to request new tasks prior to uploading finished ones: 11173 PrimeGrid 2011-05-31 5:13:01 PM Computation for task pps_sr2sieve_21291464_1 finished 11174 PrimeGrid 2011-05-31 5:13:01 PM Sending scheduler request: To fetch work. 11175 PrimeGrid 2011-05-31 5:13:01 PM Requesting new tasks for NVIDIA GPU 11176 PrimeGrid 2011-05-31 5:13:03 PM Started upload of pps_sr2sieve_21291464_1_0 11177 PrimeGrid 2011-05-31 5:13:04 PM Finished upload of pps_sr2sieve_21291464_1_0 11178 PrimeGrid 2011-05-31 5:13:15 PM Scheduler request completed: got 1 new tasks I'm not sure when this was finally fixed (if really already?), could you try 6.12.28 for this? Now with 6.12.26, if you manually update PrimeGrid after having uploaded a finished result, does it get some more tasks, or not? What is (e.g. just now) your available amount of work (sum of their remaining times) separately for CPU, nVidia, ATI? Peter ID: 38086 ·

Pepo Send message Joined: 3 Apr 06 Posts: 547	Message 38087 - Posted: 1 Jun 2011, 9:22:05 UTC - in response to Message 38086. You are using 6.12.26 and it still seems to request new tasks prior to uploading finished ones: [...] I'm not sure when this was finally fixed (if really already?), could you try 6.12.28 for this? :-( Just noticed the same on my 6.12.28 client: Finished a SETI CUDA task, requested a new one and started to upload the finished result. Peter ID: 38087 ·

Jord Volunteer tester Help desk expert Send message Joined: 29 Aug 05 Posts: 15482	Message 38088 - Posted: 1 Jun 2011, 9:29:57 UTC - in response to Message 38086. I'm not sure when this was finally fixed (if really already?), could you try 6.12.28 for this? 6.12.28 has the same problem, still. I just found it in my own logs and sent it off to David. ID: 38088 ·

Joseph Stateson Volunteer tester Send message Joined: 27 Jun 08 Posts: 641	Message 38099 - Posted: 1 Jun 2011, 11:53:40 UTC Last modified: 1 Jun 2011, 12:08:28 UTC Peter - Proth Prime CUDA on gtx570 typically run 15 minutes and generate 4500 credits This system has two nvidia gpu's: a gtx280 and a gtx460 and is maxed out at 24 tasks each for poth prime and milkyway The following is what you asked for. The system has ati 5850 and nvidia 570. It should also have 24 or so poth primes but only has 1. ID: 38099 ·

Joseph Stateson Volunteer tester Send message Joined: 27 Jun 08 Posts: 641	Message 38106 - Posted: 1 Jun 2011, 16:07:15 UTC Last modified: 1 Jun 2011, 16:22:22 UTC I am certain that this is a boinc problem or feature. The system with both ati and nvidia is now processing only milkyway tasks. According to this thread, there is a limit of 12 tasks per gpu for milkyway. I am actually seeing this right now as I have 15 Milkyway nVidia, 7 Milkyway ATI and two Milkway gpu tasks are running, one is nVidia, the other is ATI. Appearantly, the boinc scheduler is not letting me process any PrimeGrid or Collatz and thinks I need to do more Milkyway. It took 24 hours for this to stableize. My very first post showed 24 milkyway tasks and 24 collatz, all ATI, with a single PothPrime CUDA, and a day later I am down to only 24 Milkyway with a mix of nVidia and ATI and no other gpu tasks. So perhaps this is the way it is supposed to be. To get some other projects (Prime or Collatz) I guess I will have to set preferences at Milkyway for ATI only. So what caused the scheduler to select milkyway only? I think I know how this happened. I let all work finish (set NNT) before I went on vacation and shut the farm off until I got back. When I got back last week I turn the farm on and enabled new work. I suspect each project maxed out with no regards to other gpu projects on the initial request after startup. However, my other systems with similar gpu's (all nVidia) seemed to handle the restart better and collatz, milkyway, prime, einstein adn seti were all represented somewhat equally. ID: 38106 ·

Pepo Send message Joined: 3 Apr 06 Posts: 547	Message 38111 - Posted: 1 Jun 2011, 18:29:09 UTC - in response to Message 38106. Maybe it is really something with debts to be paid off - at least between the projects. But CUDA is a separate GPU resource, in addition to ATI GPUs and all CPUs, thus I think it would be nice to feed it (and cache for it) comparably to other resources. (But BOINC devs have right to have a different opinion :) maybe there would never be enough cached tasks when a project is in a large debt (unless the resource gets idle - your case with PrimeGrid CUDA tasks?)) BTW, on jstateson2quad, the 22+1 ATI Milkyway tasks are estimated to take 46:29 minutes, the PrimeGrid CUDA task is estimated to take 8:53 more minutes. Both values are rather similar, when compared to your requested 1440:00 minutes (= 1 day) - far far away. Peter ID: 38111 ·

Richard Haselgrove Volunteer tester Help desk expert Send message Joined: 5 Oct 06 Posts: 5082	Message 38112 - Posted: 1 Jun 2011, 18:33:36 UTC - in response to Message 38106. You would need to look in detail at the debt values - "Work Fetch Priority" x3, on each project properties page in BOINC Manager - and track it over time. I did that for months on end while we were trying to get GPU scheduling right - I think it's pretty good now, and respects requested resource share over the long term. If you think you're seeing something different, best to set a well-chosen set of those logging flags Jord suggested, and present the evidence. ID: 38112 ·

Claggy Send message Joined: 23 Apr 07 Posts: 1112	Message 38114 - Posted: 1 Jun 2011, 19:21:43 UTC - in response to Message 38073. Last modified: 1 Jun 2011, 19:23:30 UTC When i ran collatz only on my E8500/GTX460/HD5770 machine's GPU's (last year using Boinc 6.10.58), i did tend to find it biased one way or other, then it would meet the Wu limits in place, so no more Wu's until i completed one, when i lowered the cache enough, then it steadied out 50/50 Nvidia/ATI, so it was cache limited instead of limited by total number of Wu's, Claggy ID: 38114 ·

Copyright © 2024 University of California.
Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.