On Tue, Oct 19, 2010 at 11:00 AM, Brian Hirt <bhirt at me.com> wrote:
> On Oct 19, 2010, at 8:18 AM, Vick Khera wrote:
>
>> On Mon, Oct 18, 2010 at 3:48 PM, Aleksey Tsalolikhin
>> <atsaloli.tech at gmail.com> wrote:
>>> I've increased the alert threshold to 30 minutes as an immediate
>>> measure, but is there any way to reduce pg_dump's impact on
>>> replication?  or will we pretty much have to live with it?
>>>
>>
>> Your hardware is approaching the limits of the workload you are
>> throwing at it.  I'd recommend adding more RAM and/or tuning postgres
>> to use more of it somehow.  Either that or faster disks, or both.
>>
>
> I think the problem is actually the exclusive lock mentioned in the previous post.  Upgrading hardware might make the dump run faster, but it will still block replication.   We've noticed the exact same thing on our systems.

Yeah, I think you need to look at IO utilization outside of when dumps
are occurring to really say whether or not you're maxing out IO.  Keep
in mind pg_dump is simple copies which can really hammer most IO
subsystems hard, but if you're only using 5% of your IO the rest of
the time then you're probably ok.

-- 
To understand recursion, one must first understand recursion.