Re: eXtensible Transaction Manager API

On 02.11.2015 06:17, Craig Ringer wrote:
> On 31 October 2015 at 17:22, konstantin knizhnik
> <k(dot)knizhnik(at)postgrespro(dot)ru> wrote:
>
>> Waiting for your feedback
> For anyone wondering about performance impact, there are some graphs
> on page 23 of the PDF presentation. I didn't see anything else, and
> the graphs don't seem to cover comparison of Pg with the XTM
> transaction manager hooks and no DTM enabled vs Pg without the hooks,
> i.e. the hook overhead its self.
>
> Have you done much work on that? Personally I wouldn't expect to see
> any meaningful overhead, but I'd really like to have numbers behind
> that.
>
Overhead of indirect call is negligible - see for example
https://gist.github.com/rianhunter/0be8dc116b120ad5fdd4

But we have certainly performed comparison of PostgreSQL with/without
XTM patch.
Pgbench results are almost the same - within the measurement error:

With XTM:
transaction type: TPC-B (sort of)
scaling factor: 70
query mode: simple
number of clients: 144
number of threads: 24
duration: 600 s
number of transactions actually processed: 12275179
latency average: 7.037 ms
latency stddev: 46.787 ms
tps = 20456.945469 (including connections establishing)
tps = 20457.164023 (excluding connections establishing)

Without XTM:
transaction type: TPC-B (sort of)
scaling factor: 70
query mode: simple
number of clients: 144
number of threads: 24
duration: 600 s
number of transactions actually processed: 12086367
latency average: 7.156 ms
latency stddev: 48.431 ms
tps = 20114.491785 (including connections establishing)
tps = 20116.074391 (excluding connections establishing)