Re: Postgres benchmarking with pgbench

On 3/19/09 2:25 PM, "ml@bortal.de"  wrote:

> 
> Here is my config (maybe with some odd setting):
> http://pastebin.com/m5d7f5717
> 
> I played around with:
> - max_connections
> - shared_buffers
> - work_mem
> - maintenance_work_mem
> - checkpoint_segments
> - effective_cache_size
> 
> ..but whatever i do, the graph looks the same. Any hints or tips what my
> config should look like? Or are these results even okay? Maybe i am
> driving myself crazy for nothing?
> 
> Cheers,
> Mario
> 

I'm assuming this is linux: What linux version?  What is your kernel's
dirty_ratio and background_dirty_ratio?

The default for a long time was 40 and 10, respectively.  This is far too
large for most uses on today's servers, you would not want 40% of your RAM
to have pages not yet flushed to disk except perhaps on a small workstation.
See
Current kernels default to 10 and 5, which is better.

What is best for your real life workload will differ from pg_bench here.
I don't know if this is the cause for any of your problems, but it is
related closely to the checkpoint_segments and checkpoint size/time
configuration.

Is your xlog on the same device as the data?  I have found that for most
real world workloads, having the xlog on a separate device helps
tremendously.  Even more so for 'poor' RAID controllers like the PERC5 --
your sync writes in xlog will be interfering with the RAID controller cache
of your data due to bad design.
But my pg_bench knowledge with respect to this is limited.