Thread: latest pgbench results

latest pgbench results

From
Robert Haas
Date:
Here are the latest pgbench results from the IBM POWER7 machine.
These results were gathered about two weeks ago.  I ran each test
configuration three times; below I report the median of the three
results.  For all runs, I used scale factor = 300, clients = jobs, and
the following non-default configuration parameters: shared_buffers =
8GB, maintenance_work_mem = 4GB, synchronous_commit = off,
checkpoint_segments = 300, checkpoint_timeout = 15min,
checkpoint_completion_target = 0.9, log_line_prefix = '%t [%p] '.
Commits in use were cce5d681be7abfd9f48c28151ebf2b242f8ba438,
dd8ea2eb5e996f3f3dfd928e20aa2462c4bd9c63,
cce5d681be7abfd9f48c28151ebf2b242f8ba438.

Clients: 1
9.2 tps = 1355.953202 (including connections establishing)
9.3 tps = 1376.949633 (including connections establishing)
9.4 tps = 1283.168055 (including connections establishing)

Clients: 8
9.2 tps = 9165.548715 (including connections establishing)
9.3 tps = 9168.803482 (including connections establishing)
9.4 tps = 9130.962750 (including connections establishing)

Clients: 32
9.2 tps = 14456.061411 (including connections establishing)
9.3 tps = 14226.033279 (including connections establishing)
9.4 tps = 14932.841344 (including connections establishing)

The 9.2 and 9.3 numbers are very close, but things do seem to have
changed a bit in 9.4 - for worse at 1 client, and for better, perhaps,
at 32.

Make of these what you will; I'm not posting these as a way of
advocating anything in particular.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company



Re: latest pgbench results

From
Fabien COELHO
Date:
Hello,

> Here are the latest pgbench results from the IBM POWER7 machine. These 
> results were gathered about two weeks ago.  I ran each test 
> configuration three times; below I report the median of the three 
> results.  For all runs, I used scale factor = 300,

This means a database size of about 4.5 GB

> clients = jobs,

I understand that it means one thread per client.

> and the following non-default configuration parameters: shared_buffers = 
> 8GB, maintenance_work_mem = 4GB, synchronous_commit = off, 
> checkpoint_segments = 300, checkpoint_timeout = 15min, 
> checkpoint_completion_target = 0.9, log_line_prefix = '%t [%p] '. 
> Commits in use were cce5d681be7abfd9f48c28151ebf2b242f8ba438, 
> dd8ea2eb5e996f3f3dfd928e20aa2462c4bd9c63, 
> cce5d681be7abfd9f48c28151ebf2b242f8ba438.

How long did it run?

As there has been some changes in pgbench, would it make sense to run the 
same pgbench version (whatever) against the different servers?

To help appreciate wether the change is significant, it would help to
have all three runs so as to have an idea of the performance spread. 
Maybe you can try my patch for better pgbench measures:

http://archives.postgresql.org/message-id/alpine.DEB.2.02.1308061042030.29999@localhost6.localdomain6

My experience is that performance can varry significantly during a run.

> Clients: 1
> 9.2 tps = 1355.953202 (including connections establishing)
> 9.3 tps = 1376.949633 (including connections establishing)
> 9.4 tps = 1283.168055 (including connections establishing)

> Clients: 8
> 9.2 tps = 9165.548715 (including connections establishing)
> 9.3 tps = 9168.803482 (including connections establishing)
> 9.4 tps = 9130.962750 (including connections establishing)
>
> Clients: 32
> 9.2 tps = 14456.061411 (including connections establishing)
> 9.3 tps = 14226.033279 (including connections establishing)
> 9.4 tps = 14932.841344 (including connections establishing)
>
> The 9.2 and 9.3 numbers are very close, but things do seem to have
> changed a bit in 9.4 - for worse at 1 client, and for better, perhaps,
> at 32.

-- 
Fabien.



Re: latest pgbench results

From
Josh Berkus
Date:
On 08/06/2013 08:26 AM, Robert Haas wrote:
> Here are the latest pgbench results from the IBM POWER7 machine.
> These results were gathered about two weeks ago.  I ran each test
> configuration three times; below I report the median of the three
> results.  For all runs, I used scale factor = 300, clients = jobs, and
> the following non-default configuration parameters: shared_buffers =
> 8GB, maintenance_work_mem = 4GB, synchronous_commit = off,
> checkpoint_segments = 300, checkpoint_timeout = 15min,
> checkpoint_completion_target = 0.9, log_line_prefix = '%t [%p] '.
> Commits in use were cce5d681be7abfd9f48c28151ebf2b242f8ba438,
> dd8ea2eb5e996f3f3dfd928e20aa2462c4bd9c63,
> cce5d681be7abfd9f48c28151ebf2b242f8ba438.

What's the length of the run each time?

-- 
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com



Re: latest pgbench results

From
Robert Haas
Date:
On Tue, Aug 6, 2013 at 12:04 PM, Fabien COELHO <coelho@cri.ensmp.fr> wrote:
> How long did it run?

Each run was 30 minutes.

> As there has been some changes in pgbench, would it make sense to run the
> same pgbench version (whatever) against the different servers?

I used the master branch's version of pgbench for all tests.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company