Comparison of PgSQL Performance on HAMMER and UFS
dillon at apollo.backplane.com
Fri Oct 16 11:03:40 PDT 2009
I'd like to see how those results compare with the various
MPSAFE sysctls turned on, and also with fsync disabled (so
we can get an idea how badly fsync effects write transactions,
since we know fsync is very expensive on HAMMER):
And also with the recent buffer cache work on master.
The meta-data for a 5GB dataset should be cachable but
prior to the work buffer cache data was being thrown out
of the backing VM page caches too quickly.
Another thing we know about HAMMER is that the initial
B-Tree layout is pretty horrible. It takes a reblocking
pass to fix it up. It shouldn't effect a dataset that
small with the cache priorities fixed in master, though.
More information about the Users