Suspend - memory resident

Message boards : Questions and problems : Suspend - memory resident
Message board moderation

To post messages, you must log in.

AuthorMessage
Professor Ray

Send message
Joined: 31 Mar 08
Posts: 59
United States
Message 33529 - Posted: 26 Jun 2010, 0:22:44 UTC

If BOINC is snoozed, or projects are suspended, and the leave memory resident option is enabled, does the project resume at the point in computation where it left off, or does it resume from the point of last checkpoint for WU being resumed?
ID: 33529 · Report as offensive
Professor Ray

Send message
Joined: 31 Mar 08
Posts: 59
United States
Message 33612 - Posted: 30 Jun 2010, 21:16:31 UTC

Empirical evidence suggests that suspend - leave memory resident while suspended - computation is resumed from the same state as at the point of suspend.
ID: 33612 · Report as offensive
Profile Jord
Volunteer tester
Help desk expert
Avatar

Send message
Joined: 29 Aug 05
Posts: 14948
Netherlands
Message 33613 - Posted: 30 Jun 2010, 21:28:53 UTC - in response to Message 33612.  

Yes. When Leave apps in memory is not on, the task resumes from the last checkpoint (or start if the app doesn't checkpoint).
ID: 33613 · Report as offensive
Professor Ray

Send message
Joined: 31 Mar 08
Posts: 59
United States
Message 33614 - Posted: 30 Jun 2010, 21:56:18 UTC

I wonder what happens if the system is hibernated. Based on the rules it would appear that the system state, RAM and page-file state, is written to disk. Restarting from hibernate would seem to resume computation from the point of hibernation.

This may be useful for those crunching interminahly long WU w/out checkpoint and a power down is unavoidable.
ID: 33614 · Report as offensive
Les Bayliss
Help desk expert

Send message
Joined: 25 Nov 05
Posts: 1636
Australia
Message 33615 - Posted: 30 Jun 2010, 22:03:38 UTC - in response to Message 33614.  

Hibernating is, effectively, similar to turning off the power. You may lose the work if the OS doesn't allow enough time to save everything.
It's not recommended for climate models.

ID: 33615 · Report as offensive
Professor Ray

Send message
Joined: 31 Mar 08
Posts: 59
United States
Message 33617 - Posted: 1 Jul 2010, 2:55:57 UTC

Can anybody tell me what happens to GPU process that is suspended - if BOINC is snoozed, project manually suspended, or scheduler begins crunching on different GPU task due to switch task config - if leave memory resident option is checked?

Clearly there are implications concerning that regarding the AGP aperature segment, TurboCache (including ATI equivalent), system video memory, and shared system memory allocations.
ID: 33617 · Report as offensive
Profile Jord
Volunteer tester
Help desk expert
Avatar

Send message
Joined: 29 Aug 05
Posts: 14948
Netherlands
Message 33618 - Posted: 1 Jul 2010, 3:57:18 UTC - in response to Message 33617.  

GPU applications are always removed from GPU RAM when suspended, but for when BOINC is running benchmarks. The "Leave applications in memory" switch doesn't have any effect on this. That one is for CPU applications using system RAM only.

(Yes, GPU applications run on the CPU as well. What little system RAM they use will be kept used when LAIM is on. But you're talking about what happens to the kernels in the videocard's memory. They get flushed.)
ID: 33618 · Report as offensive
Professor Ray

Send message
Joined: 31 Mar 08
Posts: 59
United States
Message 33619 - Posted: 1 Jul 2010, 4:46:04 UTC - in response to Message 33618.  
Last modified: 1 Jul 2010, 4:56:35 UTC

I'm unclear what you mean with
GPU applications are always removed from GPU RAM when suspended, but for when BOINC is running benchmarks. The "Leave applications in memory" switch doesn't have any effect on this. That one is for CPU applications using system RAM only.


Sounds like the app remains resident in VRAM only when BOINC runs benchmarks.

However, this nugget is buried in the cc_config.xml portion of the CUDA FAQ:

There is a problem when you have your preferences set to Leave applications in memory while suspended. Any task still in the video card's RAM will stay there until you resume BOINC. Although it will not use your GPU, the amount of memory that a Seti CUDA tasks takes up (200MB+ of the VRAM) may still have impact on your other activities.


Not meaning to trip you up, but I seen the link in your signature and checked it out. While I'd already perused that tome of knowlege, I overlooked the aforementioned citation because cc_config.xml didn't seem to be germane to my question.

EDIT: found Rich Haselgrove's response to you here:

http://einstein.phys.uwm.edu/forum_thread.php?id=8254

Succinctly: VRAM is always unloaded on suspend except for benchmarks. Ergo, the FAQ needs updating.
ID: 33619 · Report as offensive
SekeRob

Send message
Joined: 25 Aug 06
Posts: 1596
Message 33620 - Posted: 1 Jul 2010, 7:31:45 UTC - in response to Message 33619.  
Last modified: 1 Jul 2010, 7:33:04 UTC

LAIM 'on' or 'off', when my devices hibernate on instruction of the UPS due frequent and enduring power outs, they've always come out successfully resuming CPU tasks lossless, including the now 6 months running CPDN model... lucky me. Never engaged in GPU crunching... it's explicitly switched off in the cc_config.xml so can't share experience on that.
Coelum Non Animum Mutant, Qui Trans Mare Currunt
ID: 33620 · Report as offensive
Profile Jord
Volunteer tester
Help desk expert
Avatar

Send message
Joined: 29 Aug 05
Posts: 14948
Netherlands
Message 33624 - Posted: 1 Jul 2010, 13:54:41 UTC - in response to Message 33619.  

Not meaning to trip you up, but I seen the link in your signature and checked it out.

I can't edit the FAQs at the moment anyway, it's returning all kinds of errors. So sent word to the keeper of.. waiting for him to tell me what's wrong.
ID: 33624 · Report as offensive

Message boards : Questions and problems : Suspend - memory resident

Copyright © 2022 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.