Switch Between Applications Every (x) Minutes

Message boards : BOINC client : Switch Between Applications Every (x) Minutes
Message board moderation

To post messages, you must log in.

AuthorMessage
DarkStar
Avatar

Send message
Joined: 7 Jan 06
Posts: 14
United States
Message 2550 - Posted: 12 Jan 2006, 18:01:26 UTC

I'm curious what the ramifications are of the "Switch Between Applications Every (x) Minutes" setting, with a 60 minute default.

Does the scheduler stop and restart to see what it's doing every (x) minutes? Or does it just check to see if there's something else it's supposed to be doing?

Or is it more like if the scheduler realizes it needs to do something else, it won't happen until (x)?

If I have a resource share of 90/10 between project A and project B, when project B starts and requires a *bit* more time than (x) to fill its quota, will it continue to run until 2(x)? Conversely, if project A is running and about to finish some work, and (x) rolls around (and there's some work for project B due), will A be terminated precisely at (x) or will it be allowed to finish its work (or at least to its next checkpoint?

Just trying to understand the interactions of the various settings - thanks for any help.
ID: 2550 · Report as offensive
Paul D. Buck

Send message
Joined: 29 Aug 05
Posts: 225
Message 2553 - Posted: 12 Jan 2006, 19:51:27 UTC

In very simple terms, it checks at that interval and decides what to run next. There are minor issues with it, particularly on multi-CPU systems and JM VII is on the case.

If work is forcing EDF, that changes things slightly...

There are other activities that can cause the CPU scheduler to switch, including downloading work, user updates, end of a result computation, etc. BUt, to put a "cap" on it so that it does not stay there forever, the time was placed there ...

WIth multiple CPUs, with 60 min setting the average switch time was about 15 minutes, 120 got you out to about 20 minutes ... 4 hours to about 30 min or so ... so, it is not a perfect thing yet.

I don't recall where we talk about this in the Wiki, I know I had a blurp on it in the old Web Site Owner's Manual (which is linked to from the Wiki front page) and it was the preferences page where you set it ... start your reading there ... hopefully some day soon I will finish moving the old Web Site Owner's Manual into the Wiki ...
ID: 2553 · Report as offensive
DarkStar
Avatar

Send message
Joined: 7 Jan 06
Posts: 14
United States
Message 2557 - Posted: 12 Jan 2006, 22:57:07 UTC

Thanks, Paul - that helps. I'm just trying to determine the most efficient settings for my specific setup ... as the old adage goes, "one size fits all usually doesn't".

Playing with these things is part of the fun.
ID: 2557 · Report as offensive
Paul D. Buck

Send message
Joined: 29 Aug 05
Posts: 225
Message 2563 - Posted: 13 Jan 2006, 10:18:02 UTC

*MY* experience is that 60 min is a "reasonable" setting for most purposes. Since I would rather "drive" work to completion I usually have a higher setting. Unfortunately, the effect I described earlier thwarts my effort somewhat.

If you have a single CPU system and set this to, say, 4 hours, for the most part you would complete work before changing project to work on something new. Obvious exceptsions would be for work that takes longer than 4 hours.

If you would tell us more ... we might be able to tell you more in return ... like, speed of the system, projects, etc.

For example, I have 9 systems with 1 to 4 CPUs (20 total) each machine has 4-7 projects running with various resource shares with either 2 or 4 hour switch times and resource shares allocated to projects in the basic order EAH, RAH, WCG, CPDN, SAH, SDG, and PG ... with values from 1% to 77% (well, 2% to 77% for a limited time I think).
ID: 2563 · Report as offensive
DarkStar
Avatar

Send message
Joined: 7 Jan 06
Posts: 14
United States
Message 2578 - Posted: 13 Jan 2006, 22:57:46 UTC - in response to Message 2563.  

Thanks again, Paul, that's very helpful - I'd also prefer to complete than switch in the middle. Because I don't leave projects in memory when suspended, reducing the frequency of switching should also reduce the (albeit little) time lost as a result. I was concerned that there might be unforeseen negative effects of increasing the time.

I'm also interested in your comment about "machines running with various resource shares with either a 2 or 4 hour switch time" - does this mean that you're able to set different resource shares and switch time on a per machine basis, or are you using the "Home/Work/School" configuration options to achieve the effect?

Currently, I'm running 8 boxes ranging in capability from a P3/1000 to P4/3.4's, so that might be a good option for me to investigate, also.

By the way, as it turns out the majority of the EDF was a result of joining a project from BOINCMGR, instead of manually creating an account and setting the resource share first. By the time I got to reducing the resource share for the backup project, BOINC had already downloaded a day's worth of work (assuming the default share) which in the case of the system involved (the P3) would hold it for a couple of months, if not longer. Somebody will probably say bad things about me for aborting a number of results, but by the time the system decided to go into EDF most of them wouldn't have completed before the deadline anyway.

Thanks again for the help!
ID: 2578 · Report as offensive
Aaron Finney

Send message
Joined: 2 Sep 05
Posts: 45
Message 2581 - Posted: 14 Jan 2006, 4:11:11 UTC - in response to Message 2550.  

The only "bad" thing that this setting could theoretically cause is if you ever have to manually abort alot of work for some reason. The core would wait 10 minutes between marking each workunit as finished.

ALTHOUGH - This may have been fixed in a recent version.

However - Thanks for reminding me about this.. I need to change mine back from 1 minute to 30 minutes. >.< This may explain why work has been so slow the last few days. I changed it to 1 minute cuz I had to cancel a whole bunch of work and was too lazy to click all that stuff manually and mickey-moused it.
ID: 2581 · Report as offensive
Paul D. Buck

Send message
Joined: 29 Aug 05
Posts: 225
Message 2585 - Posted: 14 Jan 2006, 10:51:08 UTC

Darkstar,

Well, those that complain are, ahem, well, we won't call them names. But, that is anticipated in the design and why the option is there...

I use default, School and Work as allocations with my PowerMac G5 in Work, my two Xeons in School and the rest work off default (no point of creating home), which would be the P4s and AMD systems.

On the G5, I am 77% EAH, 10% SAH, 10% RAH and 3% SDG
Xeons are 25% CPDN, 25% EAH, 25% RAH, 20% WCG, 3% SDG, and 2% PG
Rest are 10% CPDN, 5% SAH, 35% EAH, 25% RAH, 3% SDG, 20% WCG, and 2% PG

My current goal is to get WCG over SDG and PG, and EAH over SAH. When I get the first, I will reduce WCG, probably to 5-10% and put that additional back to EAH on all but the G5 (which cannot do WCG yet). When LHC has work, not sure what I will do ... but they will get time from somewhere ... :)

In general, I don't have the switch time higher as the "turbulence" in the systems makes it pointless right now. On the Xeons I am using the 1/4 splits to try to limit switching and is works somewhat. But, the tendency is to switch in and out of projects with the system tending to be running CPDN on one CPU and the other 3 on one of the other projects until it switches because of the time or completion of one work unit.

Hopefully, when JM VII finishes the next set of CPU Scheduler fixes the "Only switch Idle CPU" fix will be in place and I can get into testing that. At that point I will put one of the Xeons back on the 2 hour schedule and one on a longer switch time to test the change. If it does as I expect, I will likely increase switching time on all my systems ...

On leaving work in memory, if the computer is on and running and has 512 or more memory, there is likely little penalty except on task change, and that penalty will be less than paging it in off disk ...
ID: 2585 · Report as offensive
Michael Roycraft
Avatar

Send message
Joined: 24 Nov 05
Posts: 129
United States
Message 2594 - Posted: 14 Jan 2006, 15:46:25 UTC - in response to Message 2578.  
Last modified: 14 Jan 2006, 15:47:00 UTC

Somebody will probably say bad things about me for aborting a number of results, but by the time the system decided to go into EDF most of them wouldn't have completed before the deadline anyway.

Thanks again for the help!


DarkStar,

What's bad to say about someone who's considerate enough to abort work??? If it's aborted, it goes back to be re-assigned almost immediately, whereas if it were ignored instead, you co-crunchers would have to wait until it timed-out, got re-assigned and crunched.

Regards,

Michael



"The arc of history is long, but it bends toward Justice"
ID: 2594 · Report as offensive

Message boards : BOINC client : Switch Between Applications Every (x) Minutes

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.