HAMMER problems with ciss

Matthew Dillon dillon at apollo.backplane.com
Sat Aug 8 15:23:08 PDT 2009

:Hm, Matt didn't you completely disable SYNCHRONIZE CACHE in 2.3?
:I thought you did! May this results in performance degradation.
:Just my 5 cents :)
:   Michael

    No, SYNCHRONIZE CACHE was only disabled for USB attachments in
    2.3 (because most USB devices don't handle it).  It is enabled
    in all other cases.

    I did find another problem which might be related.  It turns out
    our bioqdisksort() was still starving requests.  HAMMER is particularly
    vulnerable to this because it tends to lay out huge swaths of meta
    data before (or after) bulk data.  UFS orders the two types of data
    too, but on a much smaller cluster-by-cluster basis.

    So with HAMMER large pipelined sequential writes can seriously
    starve other BIOs, sometimes for many seconds.  In anycase, I just
    fixed it in head.  It might or might not be related to that particular
    benchmark.  I had already done a similar fix in the NATA driver but
    now with AHCI and SILI, which use CAM, I also had to fix the generic
    bioqdisksort() as well.

    Anything that fsync()'s is going to be nasty in HAMMER no matter what.

					Matthew Dillon 
					<dillon at backplane.com>

More information about the Kernel mailing list