Improvements in HAMMER parallel I/O performance

Matthew Dillon dillon at
Wed Aug 19 23:44:22 PDT 2009

    With the latest master people should be seeing some serious improvement
    in HAMMER's parallel I/O performance.  e.g. simultanious reads and
    writes, benchmarks, and so forth.  The improvements will be more
    pronounced if you are running SCSI, SILI, or the AHCI driver, but
    NATA will see a good chunk of the improvements too.

    These improvements are not in random I/O... e.g. find . or ls -lR on an
    uncached directory tree.  They are specifically for parallel I/O loads,
    aka blogbench and friends.

    I did some major work controlling buffer cache locking as it relates
    to HAMMER, to reduce lock contention, and some major work on
    bioqdisksort() to add write bleedover and write bursting to the
    read priority scheme.

					Matthew Dillon 
					<dillon at>

More information about the Kernel mailing list