Advanced search

Message boards : Number Crunching : Request for project preferences choice.

Author Message
marmot
Send message
Joined: 13 Nov 15
Posts: 6
Combined Credit: 3,077,446
DNA@Home: 24,454
SubsetSum@Home: 185,737
Wildlife@Home: 2,867,256
Wildlife@Home Watched: 0s
Wildlife@Home Events: 0
Climate Tweets: 1
Images Observed: 0

      
Message 6107 - Posted: 25 Jan 2016, 19:16:34 UTC

Some of my machines have been running out of RAM and get stuck in a swapfile crisis and after a few days I finally caught the situation (I had noticed it from performance degrading in WU reporting on other projects). The Gibbs packets are variable in size with some being 950 MB a piece.

I went through and added app_config.xml with the entry:

<project_max_concurrent>2</project_max_concurrent>


(or N=1 on some machines) to guard against swapping crisis and on another machine with only 4GB RAM I even told the manager to only grab 0.1 days of WU in order to try and prevent the d/l of the 4.5 hour/ 950MB WU's but this kind of micromanaging takes up a lot more time than if we had some selection on the preferences.

I'd request that the Gibbs WU be divided into versions with known maximum WU sizes each individually selectable. The larger WU can have a higher credit reward just like some PrimeGrid long work units have better rewards for lengthy time requirements. This way we can still perform Citizen Grid WU on machines with less RAM even though we get less credit/WU and not have to go into extensive micromanaging of Citizengrid WU's on all our machines. A simple selection in the project preferences is all it takes.

NFS@Home has a setup like this with work units choices of 500MB or 1,000MB where the higher RAM requirements pay better credit.

Profile JumpinJohnny
Avatar
Send message
Joined: 24 Sep 13
Posts: 237
Combined Credit: 10,275,610
DNA@Home: 192,548
SubsetSum@Home: 201,740
Wildlife@Home: 9,881,323
Wildlife@Home Watched: 55,997,833s
Wildlife@Home Events: 15,584
Climate Tweets: 336
Images Observed: 351

              
Message 6111 - Posted: 27 Jan 2016, 0:05:13 UTC - in response to Message 6107.

Some of my machines have been running out of RAM and get stuck in a swapfile crisis.... I'd request that the Gibbs WU be divided into versions with known maximum WU sizes each individually selectable. The larger WU can have a higher credit reward just like some PrimeGrid long work units have better rewards for lengthy time requirements. This way we can still perform Citizen Grid WU on machines with less RAM even though we get less credit/WU and not have to go into extensive micromanaging of Citizengrid WU's on all our machines... .


Very good Idea IF it can be done with consideration to how the present Gibbs WU's depend on each other for results. Seperating them may not be possible.

Because I need a free core to run GPU and ASIC's, I use the following:

<app_config> <app> <name>subset_sum</name> <max_concurrent>1</max_concurrent> </app> <app> <name>gibbs</name> <max_concurrent>2</max_concurrent> </app> </app_config>

marmot
Send message
Joined: 13 Nov 15
Posts: 6
Combined Credit: 3,077,446
DNA@Home: 24,454
SubsetSum@Home: 185,737
Wildlife@Home: 2,867,256
Wildlife@Home Watched: 0s
Wildlife@Home Events: 0
Climate Tweets: 1
Images Observed: 0

      
Message 6114 - Posted: 27 Jan 2016, 19:14:44 UTC - in response to Message 6111.
Last modified: 27 Jan 2016, 19:15:00 UTC

Since subset seems to never go above 25 MB of Private bytes and Gibbs can send down 950 MB WU's, I'll go with this on my low RAM 4 core system:

<app_config> <app> <name>subset_sum</name> <max_concurrent>3</max_concurrent> </app> <app> <name>gibbs</name> <max_concurrent>1</max_concurrent> </app> </app_config>


That system has a hard drive that's clunking and steadily losing sectors now.
I don't have the cash to order more RAM for it and hopefully I can find a used 2.5" 120GB for about $10.
I didn't notice that it was in swapfile crisis for days (it's an extra quiet laptop drive) and that drive took a beating being used as if it were RAM.

Citizengrid and Mindmodeling both had hidden 1 GB size WU coming down hitting that machine that I wasn't expecting. They all looked like under 250MB when I would check the Process Explorer that runs all day.


Post to thread

Message boards : Number Crunching : Request for project preferences choice.