Re: WAL Insertion Lock Improvements

On Mon, May 08, 2023 at 05:57:09PM +0530, Bharath Rupireddy wrote:
> I ran performance tests on the patch with different use-cases. Clearly
> the patch reduces burden on LWLock's waitlist lock (evident from perf
> reports [1]). However, to see visible impact in the output, the txns
> must be generating small (between 16 bytes to 2 KB) amounts of WAL in
> a highly concurrent manner, check the results below (FWIW, I've zipped
> and attached perf images for better illustration along with test
> setup).
>
> When the txns are generating a small amount of WAL i.e. between 16
> bytes to 2 KB in a highly concurrent manner, the benefit is clearly
> visible in the TPS more than 2.3X improvement. When the txns are
> generating more WAL i.e. more than 2 KB, the gain from reduced burden
> on waitlist lock is offset by increase in the wait/release for WAL
> insertion locks and no visible benefit is seen.
>
> As the amount of WAL each txn generates increases, it looks like the
> benefit gained from reduced burden on waitlist lock is offset by
> increase in the wait for WAL insertion locks.

Nice.

> test-case 1: -T5, WAL ~16 bytes
> test-case 1: -t1000, WAL ~16 bytes

I wonder if it's worth doing a couple of long-running tests, too.

--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com