Last modified 12 years ago Last modified on 01/29/09 15:15:15

Timur's summary of the discussion of SRM speed improvements topic Jan 28, 2009.

  • After consideration of the list of messages used in SRM-DCACHE communication the following changes were proposed:
    1. Make srmls asynchronous and scheduled.
    2. Add information about the client side time allocated for reply wait available in the PnfsManager messages. Make PnfsManager consider its queue and provide a failed reply to the requests that could not be executed in time. Modify SRM code to take advantage of this functionality and start returning SRM_BUSY code to srm clients for the requests that could not be executed after certain number of retries due to the pnfs being busy.
    3. AutoCreate indexes for SRM DBs (list of indexes was sent by Gerd earlier)
    4. Optimize PutDone, do not use PnfsManagerGetStorageInfo, rather use PnfsManagerMapMessage, do not release file reservation in space (CancelUseSpace) if the transfer was successful.
  • The following ideas were also discussed, so they were not yet scheduled to be implemented:
    1. Make TURL unique, make doors aware of transfer related work already accomplished by SRM. Make SRM somehow pass negotiated StorageInfo, space reservations, pool info, etc to the door.
    2. Cache PnfsManager replies.
    3. Combine to SRM space reserve and use messages for implicit reservation in PrepareToPut.
    4. Make dCache services aware of system load information (CPU/Memory/Network etc.) and use it in scheduling.
    5. Make PnfsHandler aware of the pnfs load problems and hold messages or return failed messages when appropriate. Make PnfsHandler capable of sending messages with methods that provide asynchronous message reply notification.
  • Other recommendations:
  1. Allow NFS caching on SRM Node.

Last modified by Timur at Fri Feb 26 22:03:02 2021