Re: Intermittent hangs with 9.2 - Mailing list pgsql-performance

On 10/09/13 20:04, David Whittaker wrote:
> On Tue, Sep 10, 2013 at 11:26 AM, Andrew Dunstan <andrew@dunslane.net
> <mailto:andrew@dunslane.net>> wrote:
>
>
>     On 09/10/2013 11:04 AM, David Whittaker wrote:
>
>
>         Hi All,
>
>         I've been seeing a strange issue with our Postgres install for
>         about a year now, and I was hoping someone might be able to help
>         point me at the cause. At what seem like fairly random intervals
>         Postgres will become unresponsive to the 3 application nodes it
>         services. These periods tend to last for 10 - 15 minutes before
>         everything rights itself and the system goes back to normal.
>
>         During these periods the server will report a spike in the
>         outbound bandwidth (from about 1mbs to about 5mbs most
>         recently), a huge spike in context switches / interrupts (normal
>         peaks are around 2k/8k respectively, and during these periods
>         they‘ve gone to 15k/22k), and a load average of 100+. CPU usage
>         stays relatively low, but it’s all system time reported, user
>         time goes to zero. It doesn‘t seem to be disk related since
>         we’re running with a shared_buffers setting of 24G, which will
>         fit just about our entire database into memory, and the IO
>         transactions reported by the server, as well as the disk reads
>         reported by Postgres stay consistently low.
>
>         We‘ve recently started tracking how long statements take to
>         execute, and we’re seeing some really odd numbers. A simple
>         delete by primary key, for example, from a table that contains
>         about 280,000 rows, reportedly took 18h59m46.900s. An update by
>         primary key in that same table was reported as 7d 17h 58m
>         30.415s. That table is frequently accessed, but obviously those
>         numbers don't seem reasonable at all.
>
>         Some other changes we've made to postgresql.conf:
>
>         synchronous_commit = off
>
>         maintenance_work_mem = 1GB
>         wal_level = hot_standby
>         wal_buffers = 16MB
>
>         max_wal_senders = 10
>
>         wal_keep_segments = 5000
>
>         checkpoint_segments = 128
>
>         checkpoint_timeout = 30min
>
>         checkpoint_completion_target = 0.9
>
>         max_connections = 500
>
>         The server is a Dell Poweredge R900 with 4 Xeon E7430
>         processors, 48GB of RAM, running Cent OS 6.3.
>
>         So far we‘ve tried disabling Transparent Huge Pages after I
>         found a number of resources online that indicated similar
>         interrupt/context switch issues, but it hasn’t resolve the
>         problem. I managed to catch it happening once and run a perf
>         which showed:
>
>         |
>         +  41.40%       48154  postmaster  0x347ba9         f 0x347ba9
>         +   9.55%       10956  postmaster  0x2dc820         f
>         set_config_option
>         +   8.64%        9946  postmaster  0x5a3d4          f writeListPage
>         +   5.75%        6609  postmaster  0x5a2b0          f
>         ginHeapTupleFastCollect
>         +   2.68%        3084  postmaster  0x192483         f
>         build_implied_join_equality
>         +   2.61%        2990  postmaster  0x187a55         f
>         build_paths_for_OR
>         +   1.86%        2131  postmaster  0x794aa          f
>         get_collation_oid
>         +   1.56%        1822  postmaster  0x5a67e          f
>         ginHeapTupleFastInsert
>         +   1.53%        1766  postmaster  0x1929bc         f
>         distribute_qual_to_rels
>         +   1.33%        1558  postmaster  0x249671         f cmp_numerics|
>
>         I‘m not sure what 0x347ba9 represents, or why it’s an address
>         rather than a method name.
>
>         That's about the sum of it. Any help would be greatly
>         appreciated and if you want any more information about our
>         setup, please feel free to ask.
>
>
>
>     I have seen cases like this with very high shared_buffers settings.
>
>     24Gb for shared_buffers is quite high, especially on a 48Gb box.
>     What happens if you dial that back to, say, 12Gb?
>
>
> I'd be willing to give it a try.  I'd really like to understand what's
> going on here though.  Can you elaborate on that?  Why would 24G of
> shared buffers be too high in this case?  The machine is devoted
> entirely to PG, so having PG use half of the available RAM to cache data
> doesn't feel unreasonable.

Here is what I have recently learned.

The root cause is crash safety and checkpoints. This is certainly
something you want. When you write to the database these operations
first occur in the buffer cache and the particular buffer you write to
is marked dirty. The cache is organized in chunks of 8kb. Additionally
write operations are also committed to the WAL.

A checkpoint iterates over all dirty buffers writing them to the
database. After that all buffers are clean again.

Now, if you write to a clean buffer it gets entirely written to the WAL.
That means after a checkpoint since every buffer is clean every write
triggers an 8kb write to the WAL. (Already dirty buffers are written
only partially)

And the more shared buffers you have the more can be dirtied immediately
after a checkpoint, hence the spike.

To mitigate that lower shared_buffers to the 12GB Andrew mentioned or
even lower (8GB) but also adjust effective_cache_size. This should
reflect the free space you have when the database is NOT running. I
expect in your case that would be something between 40GB and 46GB.

Please correct me if I'm wrong!
Torsten