Einstein@home has been pretty conservative wrt. the BOINC server components we use. Some of them are pretty old, and heavily customized.
To be able to follow recent fixes and developments in BOINC we are now heading for a major upgrade, most prominently of the scheduler code.
I'd like to have this done before the next GW run.
Ideally the participants shouldn't notice much difference in the behavior of the server, but unfortunately we can't test all possible situations beforehand.
* Some (short) time after the actual upgrade of the code we will tell the validators to ignore the pre-assigned credit, and switch to the current default credit & runtime estimation system of BOINC. This will relieve us from having to manually adjust credit for every new workunit type, plan class, application version etc. and make us compatible to other BOINC projects.
* We plan to update the locality scheduling code to something more efficient and flexible. Again, ideally participants shouldn't notice any (negative) difference, it should just make it easier for us to manage the files involved and reduce the things to adjust, and thus the number of possible errors we could make, when setting up a run.
* I'm also working on reading plan classes from a configuration file instead of hardcoding these in the scheduler. This should relieve us from having to compile and replace the scheduler every time we just change a plan class parameter (such as the GPU memory requirement).
This is more a blog-type of news, as this is work pretty much in progress.
BM
BM
Copyright © 2024 Einstein@Home. All rights reserved.
BOINC server components (scheduler) upgrade
)
Update:
The new features of the new server code (credit, run time estimation etc.) need more investigation, testing and possibly fixing than I first thought. I doubt that the upgrade will happen before the next GW run.
The "plan class configuration file" has been implemented in a way that can be used even in the current scheduler, too. In fact it's already in the scheduler that is currently running, though not being used yet.
BM
BM