[Postgres-xl-developers] 答复: reducing the duration of regression tests(Internet mail)

Tomas Vondra tomas.vondra at 2ndquadrant.com
Sun Jul 9 13:20:42 PDT 2017



On 7/9/17 6:49 PM, Mason Sharp wrote:
> 
> 
> On Fri, Jul 7, 2017 at 12:14 PM, Tomas Vondra 
> <tomas.vondra at 2ndquadrant.com <mailto:tomas.vondra at 2ndquadrant.com>> wrote:
> 
>     Hi,
> 
>     On 7/7/17 5:33 PM, Pavan Deolasee wrote:
> 
> 
> 
>         On Fri, Jul 7, 2017 at 8:34 PM, Tomas Vondra
>         <tomas.vondra at 2ndquadrant.com
>         <mailto:tomas.vondra at 2ndquadrant.com>
>         <mailto:tomas.vondra at 2ndquadrant.com
>         <mailto:tomas.vondra at 2ndquadrant.com>>> wrote:
> 
>              Hi,
> 
>              On 7/7/17 5:23 AM, jasonysli(李跃森) wrote:
> 
>                  Hi,
>                  In fact, when running the PGXL regress on an advanced
>         server
>                  with 24 cores, 64GB memory. But it still take several
>         minutes,
>                  so I think
>                  maybe it indeed need so much time :)
> 
> 
>              But doesn't that actually suggest the opposite - that the main
>              resources (CPU, I/O, memory) are not the limiting factor
>         here? I
>              mean, if it takes minutes both on a laptop and a server
>         with much
>              more resources, then it probably is not about the amount of
>         work, no?
> 
> 
>         For long I wanted to compare run time for individual tests
>         against PG and see if there are specific tests which are much
>         slower in XL. Do you see slowdown across all tests or just
>         specific ones? May be DDLs are much slower in XL?
> 
> 
>     It's definitely a slowdown across all regression tests, although the
>     slowdown varies depending on the test of course. Attached is a
>     spreadsheet with timings for one "make check" run on XL 9.5 and
>     PostgreSQL 9.5, and two CDF charts (cumulative distribution function).
> 
>     On PostgreSQL 95% of tests completes in under 1 second, while on XL
>     it's only 40%. On XL the 95%-percentile is only achieved at ~10 seconds.
> 
>     I haven't repeated those measurements on the XL 10 code, but it
>     doesn't seem to have changed significantly.
> 
> 
> We are adding latency going between the datanodes and GTM.
> 
> Running with the coordinator generating XIDs and snapshots may speed 
> things up a bit, mainly for internal testing.
> 

But those tests are running on the same machine, through localhost. So 
the latency is like 0.05 ms. I wouldn't expect that to make tests 10x 
slower, but maybe I'm wrong.

-- 
Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services


More information about the Postgres-xl-developers mailing list