Re: DRBD and Postgres: how to improve the perfomance?

Maila Fatticcioni wrote:
> Hello.
> We have made some performance tests with DRBD and Postgresql 8.2.3. We
> have two identical servers in a cluster (Dell 2950) with a partition of
> 100 GB managed by DRBD: once we checked Postgres keeping his data folder
> in a local partition, the second time we moved the data folder in the
> shared partition. The two servers are connected point to point using a
> cross cable to reduce their latency.
> The partition is mounted with the option noatime in order to not update
> the inode access time in case of read access.
> We used pgbench for the testings, creating a dabase of about 3GB with a
> scale of 200. After we perfomed 10 tests for each configuration,
> simulating the usage of 100 clients with 500 transactions each.
>
> DRBD configuration:
> --------------------------------------------------------------------------------------------
> resource drbd0 {
>
> protocol C;
> incon-degr-cmd "halt -f";
>
> on db-node1 {
> device /dev/drbd0;
> disk /dev/sda2;
> address 10.0.0.201:7788;
> meta-disk internal;
> }
>
> on db-node2 {
> device /dev/drbd0;
> disk /dev/sda2;
> address 10.0.0.202:7788;
> meta-disk internal;
> }
> syncer {
> rate 700000K;
> }
> }
> --------------------------------------------------------------------------------------------
>
> Pgbench
>
> --------------------------------------------------------------------------------------------
> pgbench -i pgbench -s 200
> pgbench -c 100 -t 500 pgbench
> --------------------------------------------------------------------------------------------
>
> The results were that the TPS (transaction per second) with Postgres
> running in the local partition is almost double than the one with the DRDB:
>
> Postgres in shared DRBD partition: 60.863324 TPS
> Postgres in local partition: 122.016138 TPS
>
> Obviously, working with the database in DRBD, we had two writes instead
> of only one but we are a bit disappointed about the low results. We
> would like to know if there is any way to improve the performance in
> order to have a 3/4 rate instead of the 1/2 one.

You seem to be limited by the speed you can fsync the WAL to the DRBD
device. Using a RAID controller with a battery-backed up cache in both
servers should help, with and without DRBD. You might find that the
difference between local and shared partition just gets bigger, but you
should get better numbers.

In 8.3, you could turn synchronous_commit=off, if you can accept the
loss of recently committed transactions in case of a crash.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com