Thread: shared_buffers vs Linux file cache
Hi All
I thought 'shared_buffers' sets how much memory that is dedicated to PostgreSQL to use for caching data, therefore not available to other applications.
However, as shown in the following screenshots, The server (CentOS 6.6 64bit) has 64GB of RAM, and 'shared_buffer' is set to 32GB, but the free+buffer+cache is 60GB.
Shouldn't the maximum value for free+buffer+cache be 32GB ( 64 - 32)?
Is 'shared_buffers' pre allocated to Postgres, and Postgres only?
Thanks
Huan


Attachment
> From: Huan Ruan <huan.ruan.it@gmail.com> >To: pgsql-performance@postgresql.org >Sent: Thursday, 15 January 2015, 11:30 >Subject: [PERFORM] shared_buffers vs Linux file cache > > > >Hi All > > >I thought 'shared_buffers' sets how much memory that is dedicated to PostgreSQL to use for caching data, therefore not availableto other applications. > > >However, as shown in the following screenshots, The server (CentOS 6.6 64bit) has 64GB of RAM, and 'shared_buffer' is setto 32GB, but the free+buffer+cache is 60GB. > > >Shouldn't the maximum value for free+buffer+cache be 32GB ( 64 - 32)? >Is 'shared_buffers' pre allocated to Postgres, and Postgres only? > I've not looked at the images, but I think you're getting PostgreSQL shared_buffers and the OS buffercache mixed up; theyare not the same. PostgreSQL shared_buffers is specific to postgres, whereas the OS buffercache will just use free memory to cache data pagesfrom disk, and this is what you're seeing. Some reading for you: http://www.tldp.org/LDP/sag/html/buffer-cache.html Glyn
On Thu, Jan 15, 2015 at 3:30 AM, Huan Ruan <huan.ruan.it@gmail.com> wrote:
Hi AllI thought 'shared_buffers' sets how much memory that is dedicated to PostgreSQL to use for caching data, therefore not available to other applications.However, as shown in the following screenshots, The server (CentOS 6.6 64bit) has 64GB of RAM, and 'shared_buffer' is set to 32GB, but the free+buffer+cache is 60GB.Shouldn't the maximum value for free+buffer+cache be 32GB ( 64 - 32)?Is 'shared_buffers' pre allocated to Postgres, and Postgres only?
While PostgreSQL has reserves the right to use 32GB, as long as PostgreSQL has not actually dirtied that RAM yet, then the kernel is free to keep using it to cache files.
Cheers,
Jeff
Jeff Janes <jeff.janes@gmail.com> writes: > On Thu, Jan 15, 2015 at 3:30 AM, Huan Ruan <huan.ruan.it@gmail.com> wrote: >> I thought 'shared_buffers' sets how much memory that is dedicated to >> PostgreSQL to use for caching data, therefore not available to other >> applications. > While PostgreSQL has reserves the right to use 32GB, as long as PostgreSQL > has not actually dirtied that RAM yet, then the kernel is free to keep > using it to cache files. Another thing to keep in mind is that, even if Postgres *has* used the RAM, the kernel might decide to swap parts of it out if it's not being used heavily. This is pretty disastrous from a performance standpoint, so it's advisable to not make shared_buffers very much larger than what your application will keep "hot". Ideally we'd lock the shared buffer arena into RAM to prevent that, but such facilities are often unavailable or restricted to root. regards, tom lane
Thanks very much, Glyn, Jeff, and Tom. That was very clearly explained.
A related case, see the following top dump. The Postgres process is using 87g residential memory, which I thought was the physical memory consumed by a process that can't be shared with others. While, the free+cached is about 155gb. But, (87 + 155) is bigger than the total available 198g RAM. Does this mean some of the residential memory used by Postgres is actually shareable to others?
>> Mem: 198311880k total, 183836408k used, 14475472k free, 8388k buffers >> Swap: 4194300k total, 314284k used, 3880016k free, 141105408k cached >> >> PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND >> 15338 postgres 20 0 97.9g 87g 87g S 0.3 46.4 21:47.44 >> postgres: checkpointer process >> 27473 postgres 20 0 98.1g 29g 29g S 0.0 15.8 2:14.93 >> postgres: xxxx idle >> 4710 postgres 20 0 98.1g 24g 23g S 0.0 12.7 1:17.41 >> postgres: xxxx idle
>> 26587 postgres 20 0 98.0g 15g 15g S 0.0 8.0 1:21.24
Huan,
Residential memory is part of the process memory that is now swapped and is in RAM. This includes also memory shared with other processes so sum of RES for all processes may be greater that total physical memory.
I recommend this article http://www.depesz.com/2012/06/09/how-much-ram-is-postgresql-using/ to get better understanding how linux manages memory.
Regards,
Roman
On Fri, Jan 16, 2015 at 5:32 AM, Huan Ruan <huan.ruan.it@gmail.com> wrote:
Thanks very much, Glyn, Jeff, and Tom. That was very clearly explained.A related case, see the following top dump. The Postgres process is using 87g residential memory, which I thought was the physical memory consumed by a process that can't be shared with others. While, the free+cached is about 155gb. But, (87 + 155) is bigger than the total available 198g RAM. Does this mean some of the residential memory used by Postgres is actually shareable to others?>> Mem: 198311880k total, 183836408k used, 14475472k free, 8388k buffers >> Swap: 4194300k total, 314284k used, 3880016k free, 141105408k cached >> >> PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND >> 15338 postgres 20 0 97.9g 87g 87g S 0.3 46.4 21:47.44 >> postgres: checkpointer process >> 27473 postgres 20 0 98.1g 29g 29g S 0.0 15.8 2:14.93 >> postgres: xxxx idle >> 4710 postgres 20 0 98.1g 24g 23g S 0.0 12.7 1:17.41 >> postgres: xxxx idle>> 26587 postgres 20 0 98.0g 15g 15g S 0.0 8.0 1:21.24