Sorry for the late reply.
On Tue, Jul 6, 2021 at 3:15 PM r.takahash...@fujitsu.com
<r.takahash...@fujitsu.com> wrote:
>
> Hi,
>
>
> I'm interested in this patch and I also run the same test with Ikeda-san's
> fxact_update.pgbench.
Thank you for testing!
> In my environment (poor spec VM), the result is following.
>
> * foreign_twophase_commit = disabled
> 363tps
>
> * foreign_twophase_commit = required (It is necessary to set -R ${RATE} as
> Ikeda-san said)
> 13tps
>
>
> I analyzed the bottleneck using pstack and strace.
> I noticed that the open() during "COMMIT PREPARED" command is very slow.
>
> In my environment the latency of the "COMMIT PREPARED" is 16ms.
> (On the other hand, the latency of "COMMIT" and "PREPARE TRANSACTION" is 1ms)
> In the "COMMIT PREPARED" command, open() for wal segment file takes 14ms.
> Therefore, open() is the bottleneck of "COMMIT PREPARED".
> Furthermore, I noticed that the backend process almost always open the same
> wal segment file.
>
> In the current patch, the backend process on foreign server which is
> associated with the connection from the resolver process always run "COMMIT
> PREPARED" command.
> Therefore, the wal segment file of the current "COMMIT PREPARED" command
> probably be the same with the previous "COMMIT PREPARED" command.
>
> In order to improve the performance of the resolver process, I think it is
> useful to skip closing wal segment file during the "COMMIT PREPARED" and
> reuse file descriptor.
> Is it possible?
Not sure but it might be possible to keep holding an xlogreader for
reading PREPARE WAL records even after the transaction commit. But I
wonder how much open() for wal segment file accounts for the total
execution time of 2PC. 2PC requires 2 network round trips for each
participant. For example, if it took 500ms in total, we would not get
benefits much from the point of view of 2PC performance even if we
improved it from 14ms to 1ms.
Regards,
--
Masahiko Sawada
EDB: https://www.enterprisedb.com/
