DBT-3 PostgreSQL performance on 2.6

[Mary Edie Meredith <maryedie@osdl.org>]

At OSDL we compared the performance of 2.4.24 versus 2.6 base 
kernels using one of our database workloads, DBT3, on PostgreSQL. 
(This models a Decision Support type of database, aka 'DSS').  
We used the deadline scheduler on 2.6 kernels. 

DBT-3 has a single stream phase and a multi-stream phase. In the 
single stream phase a single query runs on the system, whereas 
the multi-stream phase has an active query per CPU. So we are 
talking about always a small number of active processes. 
This workload does not swap on either kernel. It consumes a 
great deal of memory, and drives a lot of 8k IO's to the disk, 
mostly read IO.

The 2.6 kernels performed better than 2.4.24 for the single stream test,
but worse for the multi-stream test for both 4-CPU and 8-CPU systems. 

Here is the 8-way data (all numbers are percent improvements 
relative to 2.4.24):

8-way
kernel   single   multi
2.4.24   base     base
2.6.2     1.3     -4.26
2.6.1-mm5 0.29    -12.38
2.6.1     6.78    -7.02
2.6.0    7.04     -15.15  

Since PostgreSQL does not support direct I/O, we are currently
investigating the possibility that system defaults (e.g. vm 
tuning parameters like dirty_expire_centisec) are not optimal
for this workload. Other suggestions welcome.

For more data (4-way, AS scheduler, latest 4-way and 8-way runs) see
http://developer.osdl.org/maryedie/DBT3_PGSQL/investigation.html

For the complete report describing the system layout and workload see:
http://developer.osdl.org/maryedie/DBT3_PGSQL/Study_DBT3_PGSQL_wkld.html

-- 
Mary Edie Meredith <maryedie@osdl.org>
Test and Performance Group
Open Source Development Lab