| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Support
Forum: BOINC Agent Support
Thread: Request to Download More Work Units |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 50
|
|
| Author |
|
|
Sekerob
Ace Cruncher Joined: Jul 24, 2005 Post Count: 20043 Status: Offline |
great insight Ingleside. Have brought down the buffer to 2 days and see what happens. Keeping it on exclusive FAAH. If the device starts pulling work again before depletion, I'll be re-affirming the monkey's uncle theorem.
----------------------------------------jgolsti, these frac values u can have life on your screen with BOINview on a by-project basis.
WCG
Please help to make the Forums an enjoyable experience for All! |
||
|
|
Ingleside
Veteran Cruncher Norway Joined: Nov 19, 2005 Post Count: 974 Status: Offline Project Badges: 
|
Thanks for another very helpful and insightful posting! I think you have hit the root cause of the problem on my machine. This machine is a laptop that I suspend overnight and at some points over the weekend vs three of my other machines which are running 24x7. The values in client_xml for this machine are: <on_frac>0.946552</on_frac> <connected_frac>0.093653</connected_frac> <active_frac>0.728136</active_frac> <cpu_efficiency>0.248532</cpu_efficiency> It looks like the root cause problem is the low CPU efficiency being calculated for this machine. This number looks out of sync with the completion speed of the work units. Is there any way to check how it is being calculated? Well, no wonder you're not getting much work, when you're only running 17.1% of the time. ;) For cpu_efficiency, add this line to cc_config.xml as another log-flag:
On_frac and active_frac has a 7-day half-life, so is slowly affected by changes. cpu_efficiency on the other hand has a 1-day half-life, so whatever done last day has greatest effect. cpu_efficiency = fraction of cpu-time to real-time science-application is getting. Using the throttle-capability, or running another cpu-intensive application like a game or something can give a huge drop...  "I make so many mistakes. But then just think of all the mistakes I don't make, although I might." |
||
|
|
cio_redulla
Advanced Cruncher Philippines Joined: Apr 24, 2006 Post Count: 130 Status: Offline Project Badges: 
|
Hello Ingleside,
----------------------------------------I'm running 5.4.11 right now because 5.8.2 and 5.8.1 didn't work right. Does 5.4.11 include the same logging capabilities as that of the later versions? I tried the config file you've mentioned earlier in your posts but I did not get the same output as that of Sekerob and the others. I guess you're right about the cpu_efficiency value, my machine obtained a high value because it's running almost 24/7. Right now, I don't have any problem. I'm just waiting for my machine to finish crunching the WUs that it has downloaded. The earliest deadline is Jan 22 and I think my machine will finish it on time. I wonder what 'overcommitted' really means. I mean, is BOINC really correct all the time when it says that the computer is overcommitted?  |
||
|
|
Sekerob
Ace Cruncher Joined: Jul 24, 2005 Post Count: 20043 Status: Offline |
There's 2 types of over-commit:
----------------------------------------1. The project, thinking that based on all the general values and project specific weight, the work cannot be finished in time: 2. The overall agent over-commitment. Item 1 i demonstrated already with the SIMAP attachment and work retrieval to hold true. With 2. The machine will revert to pure Earliest Deadline First processing. Now the answer is tricky because the algorithms between 5.4.11 and 5.8.2 were significantly amended and EDF supposedly even taken out (under its old title so to speak)? BOINC will do the best it can to balance things and finish as much it can before normality resumes and new work et al is allowed in. If one sees this and it's clear the last WU's wont make it, best to cancel those, so someone else can crunch them for u....i.e. they go back into the distribution queue. As for the testing progress, the 1 thread machine on 5.8.2 is still doing great. It's got a mix of work, multi DC, NO FAAH, which was taken out of it's WCG profile... it got 19 of WCG in the buffer on a 2.5 day buffer size and the TC actually add up to near 2.5 days considering the weight for WCG. The work trickles in 1, 2, 3 at a time as it learns from the last work completed. For the 2 thread machine, i just took out FAAH and replaced with HDC+GC. Still 4 FAAH to go and the usual depleting continuing. Than I'll see if behavior normalizes when the HDC/GC's come in. I think having even a single FAAH in the buffer is enough to cause the wacky circumstance, but that's speculation...... I'll run the diagnostic when the situation arises later today and see if it still reports this 100k seconds deadline missing..... Is BOINC right all the time.... hmmm, I think user actions have a lot to do with things going right or wrong..... fortunately the agent development evolves quickly and confirmed bugs are fixed very fast..... 5.9.0 pre-alpha is already out for self-compilation and testing. ketch u laeter
WCG
----------------------------------------Please help to make the Forums an enjoyable experience for All! [Edit 1 times, last edit by Sekerob at Jan 16, 2007 12:52:04 PM] |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Looks like cpu_efficiency is steadily increasing on my machine....thanks again for your help! I think I had run a couple of cpu intensive jobs (google desktop indexing mainly) that threw off the cpu_efficiency on this machine.
|
||
|
|
cio_redulla
Advanced Cruncher Philippines Joined: Apr 24, 2006 Post Count: 130 Status: Offline Project Badges:
|
Hello,
----------------------------------------I noticed that I get more FAAH units than the other projects. Are FAAH units more available than the others? I like these units because they give larger credits than the other projects, :-). |
||
|
|
Sekerob
Ace Cruncher Joined: Jul 24, 2005 Post Count: 20043 Status: Offline |
Distribution is set on basis of total body of available work & personal preferences. People opting to run for instance HDC or GC only get 1st served. The rest is distributed to the various OSses and to the multi-project crunchers. One value i remember quoted by the tech room a few months ago was 8:1 for FAAH:HDC. From my HDC:GC BOINC machine i know that the weight is about 50:50. On the box with FAAH:GC the balance was about 9:1
----------------------------------------It's all designed to ensure that no-one will be out of work when wanting to crunch for WCG, but some concessions needed to be made. E.g. BOINC has an 'alternate work' box. If your preferred projects run out, it will start pulling work from other projects that have availability. cheers
WCG
Please help to make the Forums an enjoyable experience for All! |
||
|
|
Ingleside
Veteran Cruncher Norway Joined: Nov 19, 2005 Post Count: 974 Status: Offline Project Badges:
|
I'm running 5.4.11 right now because 5.8.2 and 5.8.1 didn't work right. Does 5.4.11 include the same logging capabilities as that of the later versions? I tried the config file you've mentioned earlier in your posts but I did not get the same output as that of Sekerob and the others. I guess you're right about the cpu_efficiency value, my machine obtained a high value because it's running almost 24/7. Old BOINC-clients used log_flags.xml, not sure if v5.4.xx had started to use cc_config.xml or if this was introduced by v5.5.xx... In any case, there's been many changes/improvements to the logging-capabilities starting with the v5.5.xx-series. As for "didn't work right", v5.4.xx and earlier did not handle 1 CPDN-wu in deadline-mode on multi-cpu, since instead of running the CPDN-wu in time-trouble all the shorter wu was run instead... Also, this normally also meant you only asked for more work when idle cpu, regardless of cache-size. v5.6.xx fixed these problems, and with a couple later tweaks, v5.8.2 seems to have no problems handling 1 CPDN-wu running in deadline-mode on one cpu, while the other cpu(s) can have a full cache of other work. Now, the re-write also includes some less popular features, like suspending wu in a project to download even more work in same project doesn't work any longer. Also, client always supposes "worse-case", in practice meaning a cache-setting of 5 days will give less WCG-work than a 2-day cache will. Anyway, v5.8.3 for windows has hit alpha, so time for me to run-through the various tests... "I make so many mistakes. But then just think of all the mistakes I don't make, although I might." [Edit 1 times, last edit by Ingleside at Jan 16, 2007 5:36:31 PM] |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
My cpu_efficiency has steadily increased to .38 and BOINC pulled down some work before the last job finished. I found out that I was running a beta test version of another program that had a cpu utilization bug. Getting this fixed has speeded up cpu_efficiency increasing.
Ingelside, thanks again for your help and insights into how BOINC scheduling works.  |
||
|
|
cio_redulla
Advanced Cruncher Philippines Joined: Apr 24, 2006 Post Count: 130 Status: Offline Project Badges:
|
I'm running 5.4.11 right now because 5.8.2 and 5.8.1 didn't work right. Does 5.4.11 include the same logging capabilities as that of the later versions? I tried the config file you've mentioned earlier in your posts but I did not get the same output as that of Sekerob and the others. I guess you're right about the cpu_efficiency value, my machine obtained a high value because it's running almost 24/7. Old BOINC-clients used log_flags.xml, not sure if v5.4.xx had started to use cc_config.xml or if this was introduced by v5.5.xx... In any case, there's been many changes/improvements to the logging-capabilities starting with the v5.5.xx-series. As for "didn't work right", v5.4.xx and earlier did not handle 1 CPDN-wu in deadline-mode on multi-cpu, since instead of running the CPDN-wu in time-trouble all the shorter wu was run instead... Also, this normally also meant you only asked for more work when idle cpu, regardless of cache-size. v5.6.xx fixed these problems, and with a couple later tweaks, v5.8.2 seems to have no problems handling 1 CPDN-wu running in deadline-mode on one cpu, while the other cpu(s) can have a full cache of other work. Now, the re-write also includes some less popular features, like suspending wu in a project to download even more work in same project doesn't work any longer. Also, client always supposes "worse-case", in practice meaning a cache-setting of 5 days will give less WCG-work than a 2-day cache will. Anyway, v5.8.3 for windows has hit alpha, so time for me to run-through the various tests... Hello Ingleside, Are there benefits in using the new client? I mean, I'm using 5.4.11 right now and I've noticed that I'm not having any problem downloading additional WUs. Should I switch to a new client? |
||
|
|
|