• No replies
6 Posts

Pinned topic DB workload on mirrored storage, like to compare numbers

‏2014-04-24T19:55:14Z | db mirror performance


Hi all,
we are using an emc vplex metro cluster with 2 emc vmax 40k underneath.
looking at a DB2 workload (Data Warehouse usage) on AIX. DBMS seems to be writing to database files with a single writer thread per DB filesystem (16k write-size, random, direct-io).
The filesystems are JFS2 and have been created on vplex Distributed volumes (sync mirror, write through).
The write rate is about 10 Mbyte/s.
With a synthetic benchmark (fio, see config below) we can duplicate the scenario, yielding the same write rate.
Of course the write rate can be increased by using larger blocks, using more writer threads or even dropping direct-io in favour of cached io.
DBA guys are investigating whether they can tune this in the DBMS config. SAN switch parameters also are in discussion.
I'd appreciate if anyone would run below specified benchmark on a mirrored storage (be it host-based or storage-based) and offer their results for comparison.
thanks in advance
Joachim Gann
here is the fio config file I use, fio is a publicly available OSS io generator written by Jens Axboe, should be able to find it in your OSS repository for linux, aix, etc. ( for example)
you run the above mentioned workload with fio --section=dbwriter1 <filename_of_file_as_mentioned_below>
it will produce IO for 6 minutes in directory as specified by "directory" on a file with "size" size (50g in my example, the amount should be free in the filesystem).
invalidate=0    # invalidate cache not needed for dio
time_based=1    # run for specified amount of time
direct=1        # directio
ramp_time=5     # short delay before measuring
runtime=360     # 6min

size=50g        # size of the file being created for the benchmark
bs=16k          # block size
iodepth=32      # how much async ios to keep in-flight
numjobs=1       # 1 thread
Updated on 2014-04-24T19:58:13Z at 2014-04-24T19:58:13Z by j.gann