Hi,
I'm preparing a personal TPC-H benchmark. The goal is not to measure or
optimize the database performance, but to compare ZFS to UFS in similar
configurations.
At the moment I'm preparing the tests at home. The test setup is as
follows:
. Solaris snv_37
. 2 x AMD Opteron 252
. 4 GB RAM
. 2 x 80 GB ST380817AS
. Oracle 10gR2 (small SGA (320m))
The disks also contain the OS image (mirrored via SVM). On the remaining
space I have created one zpool (one disk) resp. one MD-Volume with an UFS
filesystem ontop (the other disk)
Later I want to rerun the tests on an old E3500 (4x400MHz, 2GB RAM) with
two A5200 attached (~15 still alive 9GB disks each).
The first results at home are not very promising for ZFS.
I measured:
. database creation
. catalog integration (catalog + catproc)
. tablespace creation
. loading data into the database from dbgen with sqlldr
I can provide all the scripts (and precompiled binaries for qgen and dbgen
(SPARC + x86) if anyone wants to verify my tests.
In most of these tests UFS was considerable faster than ZFS. I tested
. ZFS with default options
. ZFS with compression enabled
. ZFS without checksums
. UFS (newfs: -f 8192 -i 2097152; tunefs: -e 6144; mount: nologging)
Below the (preliminary) results (with a 1GB dataset from dbgen), runtime
in minutes:seconds
UFS ZFS (default) ZFS+comp ZFS+nochksum
db creation 0:38 0:42 0:18 0:40
catalog 6:19 12:05 11:55 12:04
ts creation 0:13 0:14 0:04 0:16
data load[1] 8:49 26:20 25:39 26:19
index creation 0:48 0:38 0:31 0:36
key creation 1:55 1:31 1:18 1:25
[1] dbgen writes into named pipes, which are read back by sqlldr. So no
interim files are created
Esp. on catalog creation and loading data into the database UFS is by factor
2-3 faster than ZFS (regardless of ZFS options)
Only for read intensive tasks and for file creation if compression is enabled
ZFS is faster than UFS. This is to no surprise, since the machine has 4GB
RAM of which at least 3GB are unused, so ZFS has plenty of space for
caching (all datafiles together use just 2.8GB disk space). If I enlarge
the dataset I suspect that then also on the tests where ZFS does perform
better, UFS will again gain the lead.
I will now prepare the query benchmark to see how ZFS performs with a larger
amount of parallelism in the database. In order to test also read throughput
of ZFS vs. UFS, instead of using a larger dataset I will cut the memory the
OS uses by setting physmem to 1GB.
--
Daniel
_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
