Message boards :
Number crunching :
Delay in getting new work units untill all work units have cleared
Message board moderation
Previous · 1 · 2 · 3
Author | Message |
---|---|
Send message Joined: 13 Dec 12 Posts: 101 Credit: 1,782,658,327 RAC: 0 |
So I've been away from Milkyway@Home for a while with summer around and may have missed something but it appears the issue of fast GPU's running out of WU's still exists? Numerous times within a few hours I found my PC's with Radeon VII's not crunching tasks. It's a shame that a Radeon VII can't crunch a full 24hrs. Any fix on the horizon? This will only get worse when newer cards are released. |
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,142,956 RAC: 2 |
So I've been away from Milkyway@Home for a while with summer around and may have missed something but it appears the issue of fast GPU's running out of WU's still exists? Yes it's annoying. And all it takes is a slight reconfiguration of the server. Why are no admins reading this? Anyway, all I do is have another project (in my case Einstein) for the GPUs to work on inbetween. You can either give the projects equal priority so they do some of each, or give Einstein 0 priority, so it only grabs just enough to run in the 10 minute gap taken to get more MW tasks. |
Send message Joined: 16 Mar 10 Posts: 208 Credit: 105,439,871 RAC: 36,925 |
@Peter Hucker Yes it's annoying. And all it takes is a slight reconfiguration of the server. Why are no admins reading this? Peter, if you happen to know what that reconfiguration might be, I'm sure the admins would be delighted to implement it if you'd care to share it! :-) But seriously, if you're assuming that it must be an easy change because other projects don't have the same problem, that may not be the whole story (unfortunately!) - I think you'll find that different projects are often on different releases of the server software, and there have been issues with some versions on upgrading. For example, reading around the problems they had at SETI@home when they did an upgrade round the end of last year, there seemed to be an issue associated with resending lost tasks, one manifestation of which was very similar to what we see here. (So perhaps if that's enabled disabling it might be a fix? Who knows - I don't!) Perhaps there's a combination of configuration and code bugs? If that's the case, it could be a non-obvious configuration choice causing the problem, or it could simply be a server code bug! Whatever it is, it may be very version-specific... And, as far as I am aware, we (the end users) can't find out what the server's configuration options have been set to, so we can only speculate anyway! With the imminent hibernation of SETI@home, there may be more folks with lots of GPU power dropping in to see if MilkyWay@home is worth the time and effort -- one or two are here already -- and those users will not like this at all! So I, too, hope something can be done about this; however, unless they can find someone expert on the server software and willing/able to help... Cheers - Al. |
Send message Joined: 13 Dec 12 Posts: 101 Credit: 1,782,658,327 RAC: 0 |
So I've been away from Milkyway@Home for a while with summer around and may have missed something but it appears the issue of fast GPU's running out of WU's still exists? Hi Peter, From memory I tried this last time and the issue was that the Einstein wu had to finish before it would start the Milkyway work. So I lost time while finishing the odd E@H and the new MW@H wu was queued. Thx for the suggestion though. Edit - Or I might have used PrimeGrid as the fill in. I can't recall. |
Send message Joined: 8 May 09 Posts: 3315 Credit: 519,941,644 RAC: 22,449 |
So I've been away from Milkyway@Home for a while with summer around and may have missed something but it appears the issue of fast GPU's running out of WU's still exists? Yes so the key is to choose a Project with a short but not too short workunit, ie PrimeGrid and then maybe the the smaller GFN numbers which on a 10series gpu can do them in about 10 to 15 minutes. Einstein units can be longer than that, some of them anyway, so I stay away from them for that. What I did to fix the problem for me was to put two gpu's in a machine and only let MilkyWay and Einstein gpu workunits on the machine, that way I don't care if MilkyWay doesn't send me work right away or not as it runs Einstein which I want to run anyway on that pc. On all my other pc's I have PrimeGrid as the backup gpu Project because of the shorter workunits. |
Send message Joined: 13 Dec 12 Posts: 101 Credit: 1,782,658,327 RAC: 0 |
G'day Mikey, Yes I did look at GFN 15 or 16. Might as well go 16 as at least a prime number is large enough to be recordable vs GFN 15. I'll give it a try some time. |
Send message Joined: 5 Jul 11 Posts: 990 Credit: 376,142,956 RAC: 2 |
So I've been away from Milkyway@Home for a while with summer around and may have missed something but it appears the issue of fast GPU's running out of WU's still exists? Yes I think it sometimes tends to let the Einstein WU finish, it's rather weird the way the Boinc scheduler works. Sometimes it will interrupt the Einstein halfway through, once it gets MW ones. I can tell it to "switch between applications" every x minutes, but it just ignores me. Doesn't really concern me though, as I like Einstein too. If it uses Einstein for 20 minutes (they take 20 minutes running two WUs at once per R9 280X card) instead of 10, oh well, MW is on for 2.5 hours anyway. I just prefer MW on these cards as I deliberately got them as they're fast at double precision work, and MW is the only project that uses that. |
Send message Joined: 18 Nov 08 Posts: 291 Credit: 2,461,693,501 RAC: 0 |
Should have mentioned this earlier: The Milkyway "fix" I put together here was built with VS2013 and requires VS2013 runtime. https://www.microsoft.com/en-us/download/details.aspx?id=40784 It will not work with the newer runtimes VS2015..VS2019. Just found out as I tried using my program on a win7 system and it did not work and there were no error messages. If you want to use my program to avoid the 10-20 minute delay then for a new install put in 7.14.2 or the newer 7.16.x and then install that VS2013 runtime. The VS2015..VS2019 runtimes did not work. The Linux version does not have any special requirements other than setting attributes and ownerships as explained here https://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=4532&postid=69562#69562 VietOZ has scripts that essentially do the same thing and no need to replace the boinc client. https://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=4532&postid=69283#69283 |
©2024 Astroinformatics Group