Great info, I really appreciate the insight. Is there a FAQ/recommended
setup for running pgbench to determine where this might be? (Is there a
reason to setup pgbench differently based on the server's cores/memory/etc?)
Sorry if this detracts from the OP's original question.
-AJ
On 3/4/2013 9:36 AM, Scott Marlowe wrote:
> On Mon, Mar 4, 2013 at 7:04 AM, AJ Weber<aweber@comcast.net> wrote:
>> Apologies for the tangential question, but how would pgpool2 "increase
>> throughput"? Wouldn't the same number of statements be issued by your
>> application? It would likely reduce the number of concurrent connections,
>> but that doesn't necessarily equate to "increased throughput".
> This is a pretty common subject. Most servers have a "peak
> throughput" that occurs at some fixed number of connections. for
> instance a common throughput graph of pgbench on a server might look
> like this:
>
> conns : tps
> 1 : 200
> 2 : 250
> 4 : 400
> 8 : 750
> 12 : 1200
> 16 : 2000
> 24 : 2200
> 28 : 2100
> 32 : 2000
> 40 : 1800
> 64 : 1200
> 80 : 800
> 100 : 400
>
> So by concentrating your connections to be ~24 you would get maximum
> throughput. Such a graph is typical for most db servers, just a
> different "sweet spot" where the max throughput for a given number of
> connections. Some servers fall off fast past this number, some just
> slowly drop off.
