Download presentation
Presentation is loading. Please wait.
Published byBerenice McCoy Modified over 9 years ago
1
PROOF work progress
2
Progress on PROOF The TCondor class was rewritten. Tested on a condor pool with 44 nodes. Monitoring with Ganglia page. The tests we have done: Single user with multi-sessions. Multiple users with single session. Multiple users with multiple sessions. Various scheduling parameters tests.
3
The Developing PROOF+COD Model Normal Production Condor jobs PROOF jobs Condor Master PROOF Master Condor + Xrootd + PROOF pool COD requests PROOF requests 3 The local storage on each machine Neng Xu, University of Wisconsin-Madison 11/29/2007
4
New Batch PROOF+COD Model Normal Production Condor jobs Interactitive PROOF jobs Condor Master PROOF Master Condor + Xrootd + PROOF pool COD requests PROOF requests 4 The local storage on each machine Neng Xu, University of Wisconsin-Madison 11/29/2007 Batch PROOF jobs Condor Scheduler for PROOF
5
Advantages of Batch PROOF system Users don’t really need always “watch” the PROOF session. Recent PROOF scheduling system doesn’t remember the historical usage of users. This can be done with Monaliza plug-in but it’s complicated to setup. Batch PROOF jobs can be schedule by Condor scheduling system. Condor will release the jobs to PROOF master based on the users’ priority. Condor also can control how many session can be released by setting up the job slots on the Scheduler. Interactive PROOF jobs still get the highest priority because they directly talk to PROOF Master. (Here we suggest users only use interactive jobs for example testing.)
6
Problems of Batch PROOF system Condor still can not fully control the PROOF session. Job resume problem.(If Scheduler crashed, job will still run. Condor resume the job but it can not fine the old one.) Interactive jobs don’t affect the users’ priority. No preemption and suspension at this point. High priority jobs need to wait until the running jobs to finish.
7
Multi-user scheduling for PROOF Control of the number of workers assigned each session: Roundrobin: every session gets same amount of workers. This is good for Batch PROOF jobs because Condor scheduler controls number of sessions. The system won’t be over loaded. Random: Session get random number of workers but each worker only be assigned once. Load based: taking account of the load. Based on following parameters. Fraction: Optwkr: Minforquery:
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.