On 10/7/16 10:42 AM, Andres Freund wrote:
> Hi,
>
> On 2016-10-06 20:52:22 -0700, Alfred Perlstein wrote:
>> This contention on WAL reminds me of another scenario I've heard about that
>> was similar.
>>
>> To fix things what happened was that anyone that the first person to block
>> would be responsible for writing out all buffers for anyone blocked behind
>> "him".
> We pretty much do that already. But while that's happening, the other
> would-be-writers show up as blocking on the lock. We don't use kind of
> an odd locking model for the waiters (LWLockAcquireOrWait()), which
> waits for the lock to be released, but doesn't try to acquire it
> afterwards. Instead the wal position is rechecked, and in many cases
> we'll be done afterwards, because enough has been written out.
>
> Greetings,
>
> Andres Freund
>
>
Are the batched writes all done before fsync is called?
Are you sure that A only calls fsync after flushing all the buffers from
B, C, and D? Or will it fsync twice? Is there instrumentation to show
that?
I know there's a tremendous level of skill involved in this code, but
simply asking in case there's some tricks.
Another strategy that may work is actually intentionally
waiting/buffering some few ms between flushes/fsync, for example, make
sure that the number of flushes per second doesn't exceed some
configurable amount because each flush likely eats at least one iop from
the disk and there is a maximum iops per disk, so might as well buffer
more if you're exceeding that iops count. You'll trade some latency,
but gain throughput for doing that.
Does this make sense? Again apologies if this has been covered. Is
there a whitepaper or blog post or clear way I can examine the algorithm
wrt locks/buffering for flushing WAL logs?
-Alfred