HEADS UP - HAMMER destabilized for a few days (update 2)

Matthew Dillon dillon at apollo.backplane.com
Sun Jun 8 11:35:36 PDT 2008


    People testing HAMMER are going to have to update both the kernel
    and newfs_hammer utility, and newfs_hammer their test filesystems again
    I'm afraid.  There was a critical bug in newfs_hammer and a CRC
    generation bug in the filesystem core.

    HAMMER is not yet fully stabilized.  I've whacked most of the issues
    but FSX still fails occassionally under heavy loads and there
    are two assertion panics, all related to the new write-performance
    code.  I expect to get the panics fixed today and will hopefully track
    the bug FSX is revealing down by tomorrow.

    Write performance as of commit 53B is now very good.  The filesystem
    frontend is now able to directly reserve space on the media and flush
    its dirty buffers out whereas before it had to queue the dirty buffers
    to the backend.

    Blogbench did a complete flip.. now random write performance is
    incredible but random read performance is suffering (probably due to
    the write performance eating up so much of the disk bandwidth). 

    In general, I am very happy with HAMMER's performance now and do not
    expect to have to introduce any new major bits of code on that front.
    
					-Matt
					Matthew Dillon 
					<dillon at backplane.com>





More information about the Kernel mailing list