Message boards : Number crunching : Short or long WUs?
Author | Message |
---|---|
Hypermarkup Send message Joined: 3 Mar 06 Posts: 7 Credit: 112,275 RAC: 0 |
Hi Rosetters! I've got an important amount of computers located in a computer room, so they are constantly switched on and off... my question: what should we consider better: to work on short WUs (like 1-2 hours/WU) or on long ones (i.e. 6-12-24 hours)?. Let's say that when we switch Boinc or the PC off, the work is saved, so why use short ones -more downloads, more switching-between for the CPU-? Should we try different proteins with shorter docking times for each one instead of receiving one WU and work on it for 24 hours of work time? I know, a WU consists of a group of models, so more work time per WU won't necessary led to a better prediction... or will it? What do you think? Hypermarkup Fotowing |
Mod.Sense Volunteer moderator Send message Joined: 22 Aug 06 Posts: 4018 Credit: 0 RAC: 0 |
With machines that are powered on and off all the time, your main concern is about how much of your crunch time is preserved, and how much is lost. Work is preserved at the end of a model, and you don't have any control over how long the individual models take. Work is also preserved at checkpoints established by the project, so again, you don't have any control over that; other then the setting for how frequently to write to disk. I think the default is 60 seconds. Just make sure you don't set it extremely high. The checkpoint doesn't complete until a time when you allow BOINC to write to disk, so if you set it to 3600 seconds or something, then you would be losing an average of a half hour of crunching each time your power off (assuming you were on for that long). You can only control your preferred runtime for a task. This will basically determine how many models you get done. More runtime, more models. It does not effect the accuracy of the models produced, nor the amount of work preserved when you power off the machine. So, your decision about runtime preference should be based on the other factors you mentioned, such as download bandwidth, and reporting completed results back prior to deadlines. You will want to assure you have the permission of the true owner of these machines. You will probably want to run BOINC as a service so it starts when the machine starts, and runs the whole time the machine is on. You may also want to establish network and CPU times of day when BOINC should run. That way all the BOINC network traffic can occur during off-peak time. By setting CPU time of day you can control your air conditioning and peak electricity costs somewhat, or assure that BOINC is not running during times when the machines are likely in-use or undergoing frequent reboots. Rosetta Moderator: Mod.Sense |
Message boards :
Number crunching :
Short or long WUs?
©2024 University of Washington
https://www.bakerlab.org