POEM GPU

[LawryB]

VeeBee.   Had the same problem, same config, platform until I put in an app_info.xml file.  Then all was rosy.   I did a bit of a study of XML elements and now have to do a little tickling here and there, mainly regarding some of the figures shown in the app_info.xml file I downloaded from the Moo forum.
Got to look at flops, api version and gpu ram and set them for my PC and think (?) I have located them in the client_state.xml file in the BOINC/data directory.

I would be interested to know the content of your app_info.xml file should you be using one for MOO.  Maybe your could PM me with it?

[kashi]

Just a handy hint with version 7.0.2. Unlike other versions of BOINC I have used it will still download tasks when you have some of them suspended. If you have lots of temporarily suspended CPU project tasks in order to do reissued tasks first or for other reasons, if you click Update you can end up downloading more than your cache is set to.

Can you elaborate on that at all?  Does it download more for the suspended project or the unsuspended one?  If the former, that sounds broken (unless you're just suspending particular tasks rather than the project as a whole), but if the latter, that happens for me and I'm still on 6.10.60 (ie, if I suspend one project, other projects will have tasks downloaded to fill the 'empty' cache).

No it's project specific, for example if I have some WCG tasks suspended, BOINC 7.0.2 will still download more WCG tasks. With previous BOINC versions I used if you had some tasks suspended for a project, BOINC would not download any new tasks for that project.

[kashi]

Hmmm, never seen that download directory before.. thanks, +1 !

so are all the Windows versions there "stable" or tested yet ? as I am still having work fetch issues with these computers at times, with POEM gpu WU's.....

No definitely not all stable versions, versions get posted there as soon as they are released for testing. That's why I said "Usually wait until someone else has tried a new version first though, sometimes the beta versions are very buggy and get withdrawn."

As for the work fetch issues I worked out the trick with the new BOINC 7.x.x versions. It goes on your "Connect about every x.xx days" setting now. In other words if you leave that value at zero as was previously recommended for an always on connection BOINC will not fetch any tasks until your cache is dry, no matter how high you set your "Additional work buffer" setting.

So because of the current trickle feed of the POEM server, I have set "Connect about every" to 9 days and "Additional work buffer" to 10 days. I have also set <report_results_immediately>1</report_results_immediately> in a cc_config.xml file. So now BOINC asks for new POEM tasks every time one completes. It causes all my running GPU and CPU tasks to go into "High Priority" mode, but none are starting out of order so far so no harm done there. I have to put the values back to 1 day once a day and resume work fetch for WCG to keep a 1-2 day cache there. DNA@Home is self limiting to 4 tasks per CPU so that is no trouble when/if work becomes available again. If they ever fix up the POEM server BOINC will download a day's quota of tasks, probably 500-600 but that won't matter as my GPU will gobble that many up in about 4 days and there's no quorum so I won't hold anyone up.

[veebee]

I have Moo! running fine on this machine (with  6950/4770 cards) by going back to version 6.12.34, so no real problem there - though i would rather run POEM , both for the credit and the science.

My other i7 is still running 7.0.2 and I just modified the cc_config file as per Kashi's post (added the "report_results_immediately" line) , did a restart of the client and suddenly there were 3 POEM Wu's crunching, though I think they were hidden in there somewhere.. somehow, as they were all at different run-times...

have suspended Primegrid on that machine until I see if POEM will send down some more work..

[veebee]

I have given the machine running client 7.0.2 well over 24 hours now but it only downloads 1 POEM WU at a time.. and THEN, it doesn't even report it straight away.. I am not sure how long it takes to go and report it, but I end up doing it pretty much every time.

I edited the cc-config file as per your advice Kashi, in the hope that THAT was why I was not getting work, yet when I get the BOINC client to "read config file", the event log gives me a worrying "no config file found, using defaults". It is in the correct BOINC directory...cannot work out why it is doing it.

I read your post/ reply to someone on POEM's site regarding the "your app_info file doesn't not conatain..." and noted what you said about that meaning that there were no work units to "get", but how is it that with pretty much the same settings you are making a killing there at POEM and I cannot get any work  ?

Also have the app_info.xml file in the POEM folder....

nothing is working anymore, so I am going back to 6.12.34 I think... and run Moo!... sadly, it is not the type of project i would prefer to run, but there is pretty limited choice for GPU's.

[LawryB]

VeeBee.   In the DATA Directory

[kashi]

Yes for Windows 7, cc_config.xml goes in C:\ProgramData\BOINC.

I know what you mean about getting work though, even with report results immediately it is hard to keep a cache in POEM for my single GPU. For those with multiple fast GPUs it would be worse. It may not get any better soon either, I was hoping the application would be improved so that it was not necessary to run multiple concurrent tasks to get higher than 50% GPU load, but the main developer of the GPU version has just announced he is leaving.

If you are reporting completed tasks immediately and thus requesting new work frequently you have a much greater chance of picking up resends though. Without resends and/or the occasional loosening of the severe trickle setting there is currently no way to avoid almost constantly running out of POEM tasks other than wearing out your update finger unless you're an officer and command your .batman to do it for you.  

Resends bypass the trickle setting of the POEM server so instead of only getting 1 or 2 tasks each request you can get up to 33 at once. As the resends get sent out in batches if you snag some early in the batch you can get more on the next request. So you sometimes get 50-60 resent tasks in a few requests.

It might not make much difference now as the trickle feed seems to be further reduced today. Previously BOINC would make another work request when it successfully got work but hadn't yet reached the minimum cache setting, but today it's not doing that. With a server throttled that heavily the only contributors remaining will be mainly those running a script. Much fairer to limit the total tasks in progress to a smaller number per GPU like MilkyWay does. Then it annoys everyone equally not just some.;D

Of course it also disbars anyone with an intermittent internet connection from contributing at all, but it seems as if BOINC/projects have chosen to abandon those people anyway for many GPU projects.

I got fed up with MilkyWay's unreliable server a fair while ago. I've had a good little run on POEM but just about had enough of fiddling with user unfriendly BOINC development versions now, so if POEM continues to not have enough work for much longer I will detach there also. Same goes for WCG GPU when they get it going, if they don't have enough work then I'll probably have a little dabble then give it a miss.

[veebee]

maybe a batch file (is that the right name??) where you can make the project keep "updating" every X minutes ?

but, every time I hit update, it says "not requesting new tasks" ANYWAY, so what is the point ?

I have gone over to Moo! with both machines now (and back to 6.12.34 client) but am looking forward to WCG having a go at GPU applications !!

[kashi]

maybe a batch file (is that the right name??) where you can make the project keep "updating" every X minutes ?

Yep a batch file is denoted by the file type ending .bat. It's the ".batman" I was cryptically alluding to, the "command" referred to the BOINC command tool  boinccmd.exe, the program that is called. It is also the "script" that I referred to later that is being used by many of the current top POEM hosts.

but, every time I hit update, it says "not requesting new tasks" ANYWAY, so what is the point ?

It's caused by a combination of the BOINC 7.0.xx work fetch policy working with the POEM server trickle settings. Needs high "Connect about every" setting as a workaround and even with that it often only requests ATI work every 2nd server contact.

I have gone over to Moo! with both machines now (and back to 6.12.34 client) but am looking forward to WCG having a go at GPU applications !!

Many others have also quickly moved to other GPU projects or stopped GPU crunching when they discovered an unreliable supply of work after they attached. I'm only using report results immediately and I've got about 1.5 days cache left but it's gradually dropping. If I don't snag a few decent size batches of resends in the next few days I'll be out of work and detaching too.

With new projects, CreditNew initial spikes and the potential of a reduced credit rate in the future prompt some to attempt to make hay while the sun shines. Very hard to move up the rankings later if the credit rate gets reduced to a half or a quarter of what it is now. In a way I'd like to use a script to get to 20 million but it's unfair on others struggling to get any work at all. Besides, it's a bit silly anyway, why use electricity on a project that already has more contributor GPU resources than it can handle. Better for me to turn off GPU processing, reduce my electricity bill and start up a few more CPU cores on GO Fight Against Malaria or DNA@Home when there is more work. Have to keep Travis busy to stop him freezing in North Dakota.;D

Sometimes it appears the different projects do not share their knowledge and experience about the methods of dealing with the high load placed on project servers by GPU processing. Unless the servers are really heavy duty, server load needs to be reduced by sending out longer running GPU tasks. If you have tasks that run for 1 minute like MilkyWay or under 10 minutes average like POEM, it's not really surprising that the servers are put under heavy load. Also need sufficient capacity to store, and sufficient staff resources to analyse the large amount of data returned by even a relatively small number of fast GPUs.

[LawryB]

I wonder if this has something to do with it all.   Found this on the MOO forum posted by Mikey.


The version 7 series of Boinc handles work requests MUCH different that the older version of Boinc, they are almost totally different!! Under Network Usage there are two setting...one for "connect about every" and one for "additional work buffer". They are NOT treated as they say, they treated as hard count numbers now!! So if you put 1.25 in the 2nd one, "additional work buffer" you are telling Boinc to NOT connect to a project for 1.25 days! It is a learning curve and you MUST work with it to get it working like you want it to. Remember programmers think they KNOW what you want and give it to you whether you really want it or not! They also have stuff they are working towards that the rest of us have no clue about.

I guess it are trial versions after all.

[tazzduke]

Greetings All

Due to the rarity of the POEM GPU tasks, I have got the HD 5850 crunching on other ATI GPU projects, and have upgraded to BOINC 7.0.12 (x64), still tweaking it to suit my setup, well I am attached to 8 ATI GPU projects and all are processing and getting new workunits bar POEM.

Just gotta watch out the cache does not get to big, also what might work for some projects doesnt necessarily work for other projects, ie grabbing 10 workunits on one project and the other project grabs 50 workunit on the same setting.

Work in progress, but at least I am crunching away.

Happy Crunching to all

Tazzduke

[Sean]

Which are the eight projects you are running? I know Milkyway, Primegrid, Collatz, Poem, Moo and DNETC, but what are the other two?  ???

[veebee]

Which are the eight projects you are running? I know Milkyway, Primegrid, Collatz, Poem, Moo and DNETC, but what are the other two?  ???

... even then, DnetC isn't "active" any more.. and I cannot think of 8 projects for ATI either !?

[tazzduke]

Hi Sean

Milkyway, Primegrid, Collatz, POEM, Moo, there is also

Albert

I also run SETI and SETI Beta using the lunatics app.

DNETC is now offline.

Regards
Tazzduke

[Sean]

Thanks Tazzduke, I will look into those two tomorrow if I get time. The idea of running Seti on a gpu is appealing.   :drool: