At 12:46 AM 12/28/2006, Guy Rouillier wrote:
I don't want to violate any license agreement by discussing
performance, so I'll refer to a large, commercial
PostgreSQL-compatible DBMS only as BigDBMS here.
I'm trying to convince my employer to replace BigDBMS with
PostgreSQL for at least some of our Java applications. As a proof
of concept, I started with a high-volume (but conceptually simple)
network data collection application. This application collects
files of 5-minute usage statistics from our network devices, and
stores a raw form of these stats into one table and a normalized
form into a second table. We are currently storing about 12 million
rows a day in the normalized table, and each month we start new
tables. For the normalized data, the app inserts rows initialized
to zero for the entire current day first thing in the morning, then
throughout the day as stats are received, executes updates against
existing rows. So the app has very high update activity.
In my test environment, I have a dual-x86 Linux platform running the
application, and an old 4-CPU Sun Enterprise 4500 running BigDBMS
and PostgreSQL 8.2.0 (only one at a time.) The Sun box has 4 disk
arrays attached, each with 12 SCSI hard disks (a D1000 and 3 A1000,
for those familiar with these devices.) The arrays are set up with
RAID5. So I'm working with a consistent hardware platform for this
comparison. I'm only processing a small subset of files (144.)
BigDBMS processed this set of data in 20000 seconds, with all
foreign keys in place. With all foreign keys in place, PG took
54000 seconds to complete the same job. I've tried various
approaches to autovacuum (none, 30-seconds) and it doesn't seem to
make much difference. What does seem to make a difference is
eliminating all the foreign keys; in that configuration, PG takes
about 30000 seconds. Better, but BigDBMS still has it beat significantly.
If you are using pg configured as default installed, you are not
getting pg's best performance. Ditto using data structures optimized
for BigDBMS.
A= go through each query and see what work_mem needs to be for that
query to be as RAM resident as possible. If you have enough RAM, set
work_mem for that query that large. Remember that work_mem is =per
query=, so queries running in parallel eat the sum of each of their work_mem's.
B= Make sure shared buffers is set reasonably. A good rule of thumb
for 8.x is that shared buffers should be at least ~1/4 your RAM. If
your E4500 is maxed with RAM, there's a good chance shared buffers
should be considerably more than 1/4 of RAM.
C= What file system are you using? Unlike BigDBMS, pg does not have
its own native one, so you have to choose the one that best suits
your needs. For update heavy applications involving lots of small
updates jfs and XFS should both be seriously considered.
D= Your table schema and physical table layout probably needs to
change. What BigDBMS likes here is most likely different from what pg likes.
E= pg does not actually update records in place. It appends new
records to the table and marks the old version invalid. This means
that things like pages size, RAID stripe size, etc etc may need to
have different values than they do for BigDBMS. Another consequence
is that pg likes RAID 10 even more than most of its competitors.
F= This may seem obvious, but how many of the foreign keys and other
overhead do you actually need? Get rid of the unnecessary.
G= Bother the folks at Sun, like Josh Berkus, who know pq inside and
out +and+ know your HW (or have access to those that do ;-) )inside
and out. I'll bet they'll have ideas I'm not thinking of.
H= Explain Analyze is your friend. Slow queries may need better
table statistics, or better SQL, or may be symptoms of issues "C" or
"D" above or ...
I've got PG configured so that that the system database is on disk
array 2, as are the transaction log files. The default table space
for the test database is disk array 3. I've got all the reference
tables (the tables to which the foreign keys in the stats tables
refer) on this array. I also store the stats tables on this
array. Finally, I put the indexes for the stats tables on disk
array 4. I don't use disk array 1 because I believe it is a software array.
I= With 4 arrays of 12 HDs each, you definitely have enough spindles
to place pg_xlog somewhere separate from all the other pg tables. In
addition, you should analyze you table access patterns and then
scatter them across your 4 arrays in such as way as to minimize head
contention.
I'm out of ideas how to improve this picture any further. I'd
appreciate some suggestions. Thanks.
Hope this helps,
Ron Peacetree
