| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Support
Forum: BOINC Agent Support
Thread: BOINC doesn't seem to follow it's own preferences |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 9
|
|
| Author |
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Hey there!
I crunch at three different PC's at the moment (all with Windows XP and BOINC 5.10.28 installed). I've set my preferences to Buffer 0,5 days of work ahead + "is online about every" 0,0100 days. However, BOINC seems to just ignore that and generally doesn't make much sense. On one Machine (Dual Core), the buffered extra work doesn't add up, it downloads quite a bit more than half a day ahead (even considering it would get work for both CPU's, it still downloads way too much), while on the other machines it gets either too less or no work ahead at all. Also, when a WU is finished it uploads it, but doesn't start a scheduler request every 0,0100 days like it's told to. There will be a scheduler request eventually, but it seems quite random and not in correspondance with the every 0,0100 days. Now why would all that happen and how can I make it to behave the way I want? best regards, gmw |
||
|
|
Sgt.Joe
Ace Cruncher USA Joined: Jul 4, 2006 Post Count: 7849 Status: Offline Project Badges: 
|
BOINC learns. If you just leave it alone for a while it will eventually stabilize close to what you want. Such things as a change in the amount of time you leave the computer on or a change in the length of the work units will throw it out of whack for a while, but it will eventually return to about what you set.
----------------------------------------Cheers
Sgt. Joe
*Minnesota Crunchers* |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Also, it will only send a scheduler request when it needs to in order to keep a half-day long queue. So it will not send 100 requests a day.
|
||
|
|
Sekerob
Ace Cruncher Joined: Jul 24, 2005 Post Count: 20043 Status: Offline |
Hey there! I crunch at three different PC's at the moment (all with Windows XP and BOINC 5.10.28 installed). I've set my preferences to Buffer 0,5 days of work ahead + "is online about every" 0,0100 days. However, BOINC seems to just ignore that and generally doesn't make much sense. On one Machine (Dual Core), the buffered extra work doesn't add up, it downloads quite a bit more than half a day ahead (even considering it would get work for both CPU's, it still downloads way too much), while on the other machines it gets either too less or no work ahead at all. Also, when a WU is finished it uploads it, but doesn't start a scheduler request every 0,0100 days like it's told to. There will be a scheduler request eventually, but it seems quite random and not in correspondance with the every 0,0100 days. Now why would all that happen and how can I make it to behave the way I want? best regards, gmw It's as what the Sergeant says. While crunching, BOINC continuously recomputes what in the buffers and how long it will take. It looks at computing history and is very very conservative in stocking work if it thinks that deadlines cannot be met.... makes no sense, well I've been watching it for longer to know that it does its job best when not watching. With multiple machines, catch a copy of BOINCview and connect it to all computers, local, remote, lan/wan. Periodically check the tasks and host tabs to see the CPU efficiency is good and the tasks in progress lines are green. It installs on all windows, including Vista, and isable to also monitor any other client on any other OS. BTW, the Additional Buffer option is now also available on the WCG device-profile pages and is called "Cache X extra days of work". If not visited there, the whole page has been reorganised.
WCG
----------------------------------------Please help to make the Forums an enjoyable experience for All! [Edit 1 times, last edit by Sekerob at Nov 3, 2007 12:18:48 PM] |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Well thanks for all the answers, I guess I'll just keep an eye on it then and see how it does. The machines are running 24/7 anyway so it shouldn't take BOINC that long to adjust. Just got wondering when suddenly the dual core machine would download strange amounts of work without reporting any of them completed for a while (half day + or sth).
----------------------------------------Hmm.. no I gott figure out the Admin pass for hat other machine to install BOINC there aswell :) [Edit 1 times, last edit by Former Member at Nov 3, 2007 5:52:39 PM] |
||
|
|
XS_olympic
Cruncher Joined: Sep 20, 2007 Post Count: 19 Status: Offline |
This is a bug with the new "short" DDDT work units. There is a large difference between the estimated time to completion and the actual time to completion. This makes BOINC think it's overcommitted so it refuses to communicate with WCG to report results and get new work unitl the "backlog" is cleared out. I've seen some of my machines estimate that they have 2-3 times the amount of work queued up that they should have and refuse to report results for over 24 hours.
So far I've dealt with it by simply hitting the update button once a day to report all the finished results. It's not a big deal, just an annoyance. Even an unattended machine would report in eventually when it's queue got short enough. |
||
|
|
Sekerob
Ace Cruncher Joined: Jul 24, 2005 Post Count: 20043 Status: Offline |
Hi XS_olympic,
----------------------------------------This is not a bug! As posted elsewhere, the servers continuously compute the average flops of actual work returned and assigns those flops in the header of new work circulated. The techs advised it would take a few days to normalize given the mix of the long and short jobs coming back, which it has done now more or less. All new jobs here come in with about 1:10 in estimated time, consequently lots more work unit is being buffered. The logic facilitates the best possible estimations of flops, but that said, with non-deterministic calculations it always varies. This exceptional step change is rare, so indeed we just need to work thru to find it regularised again for the larger buffer operators. If such an announcement is seen, the workaround is to set the work-fetch off, again if operating large buffer, process all cached and start fetching again. If you suffer from high build-up of "Ready to Report" (RtR), consider changing/reducing the connect time and employ the new 'cache' option on the WCG device profiles and observe if RtR's get cleared more frequently. My experience with permanent internet connection is that it always backfills that way even if the RtR's have not been send. If one of the 6 or 7 conditions have not occurred when work is needed reporting, they wont. One of these reporting rules is "24 hour after Sending Result". (See Wiki , point 4 ). cheers.
WCG
Please help to make the Forums an enjoyable experience for All! |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
XS_olympic, WCG now have an improved method for estimating the size of work units. Size is based on the size of recent units for that project.
This means that if you are seeing poor time estimates, there is likely to be a different cause. Time estimates are a function of work unit size, benchmark and DCF. Assuming your benchmark is accurate (and even if it is not) the figure to check is the DCF. Again, this value is corrected based on recent performance, but it may change slowly. Queue size is based on so many other factors that I'm not going to attempt an explanation here. Besides, different versions of the scheduling algorithm are more aggressive in queuing work, while others are more cautious. |
||
|
|
XS_olympic
Cruncher Joined: Sep 20, 2007 Post Count: 19 Status: Offline |
Thanks for pointing out the "cache" option in my BOINC profile, I hadn't noticed it yet. I set the connect time to 0.25 days and the cache size to 2.5 days. I'll run it for a few days and see what happens.
|
||
|
|
|