HAMMER problems with ciss

Matthew Dillon dillon at apollo.backplane.com
Sun Aug 9 11:14:34 PDT 2009


:The performance is also lower when reading, here are some lines from
:the random read test:
:
:32              4.81    153.94  0.0000  0.0065  0.8895  0.0103
:64              10.59   169.49  0.0000  0.0059  2.5563  0.0085
:128             15.54   124.35  0.0001  0.0080  3.3134  0.0117
:256             29.90   119.61  0.0001  0.0083  2.2407  0.0116
:
:32              6.34    203.13  0.01    4.92    570.37  7.88
:64              11.66   186.58  0.02    5.36    516.09  8.40
:128             20.30   162.41  0.04    6.15    517.97  9.41
:256             34.13   136.54  0.07    7.32    667.05  10.74
:
:This is with a P400, 512 MiB BBWC and a 15k RPM SAS disk.

    Yah, that is more in-line with what I would expect.  UFS will win
    for data-uncached random reads as long as the namespace itself is
    cached (fits in UFS's dirhash cache).  This is primarily due to
    the direct data blocks stored in the inode itself, so there is one
    less I/O.

    Also, HAMMER's B-Tree tends to get a bit trashed by those tests,
    since no cleanup occurs while the test is running.  It is probably
    storing additional history as well.

					-Matt
					Matthew Dillon 
					<dillon at backplane.com>





More information about the Kernel mailing list