On Tue, Aug 30, 2011 at 3:00 AM, Boszormenyi Zoltan <zb@cybertec.at> wrote:
> Hi,
>
> 2011-08-29 22:36 keltezéssel, Lonni J Friedman írta:
>> ... I read that
>> (max_connections * work_mem) should never exceed physical RAM, and if
>> that's accurate, then I suspect that's the root of my problem on
>> systemA (below).
>
> work_mem is process-local memory so
>
> (max_connections * work_mem) < (physical RAM - shared_buffers)
>
> Some queries may allocate multiples of work_mem, too.
>
> Also, the kernel uses some memory for internal accounting, caching
> and you need to account for the process binary in memory.
>
>> However, I'd like confirmation before I start
>> tweaking things, as one of these servers is in production, and I can't
>> easily tweak settings to experiment (plus this problem takes a few
>> weeks before swapping gets bad enough to impact performance).
>>
>> A few examples:
>>
>> 0) system A: 56GB RAM, running postgresql-8.4.8 with the following parameters:
>> maintenance_work_mem = 96MB
>> effective_cache_size = 40GB
>> work_mem = 256MB
>> wal_buffers = 16MB
>> shared_buffers = 13GB
>> max_connections = 300
>
> RAM (56GB) - shared_buffers (13GB) = 43GB
>
> which is less than
>
> work_mem * max_connections = 300 * 0.25GB = 75GB
>
> The system would start swapping before 43GB/0.25GB = 172 clients.
>
>> 1) system B: 120GB RAM, running postgresql-9.0.4 with the following parameters:
>> maintenance_work_mem = 1GB
>> effective_cache_size = 88GB
>> work_mem = 576MB
>> wal_buffers = 4MB
>> shared_buffers = 28GB
>> max_connections = 200
>
> Similarly:
>
> 120GB - 28GB = 92GB
>
> is less than
>
> work_mem * max_connections = 200 * 576MB = 112.5GB
>
> Also, if you run anything else on the machine then the system would start
> swapping much sooner than hitting max_connections number of clients.
>
> I would never set work_mem that high by default. 8 - 16MB is usually
> enough for the common case and you can set work_mem for special
> queries from the client and then reset it.
Thanks for your reply. I've reduced shared_buffers to 8GB everywhere,
which should definitely help.