Обсуждение: BUG #8326: pgbench tps wrong BUG

The following bug has been logged on the website:

Bug reference:      8326
Logged by:          digoal.zhou
Email address:      digoal@126.com
PostgreSQL version: Unsupported/Unknown
Operating system:   CentOS 5.8 x64
Description:

when client thread big, the tps without connection time wrong!!!!
pg94@db-172-16-3-33-> pgbench  -M prepared -n -r -f ./test.sql -T 10 -c 800
-j 100 digoal
transaction type: Custom query
scaling factor: 1
query mode: prepared
number of clients: 800
number of threads: 100
duration: 10 s
number of transactions actually processed: 2037960
tps = 199534.267305 (including connections establishing)
tps = 675603.691013 (excluding connections establishing)
statement latencies in milliseconds:
        1.019046        select 1;


when client threads small, the wrong is little.
pg94@db-172-16-3-33-> pgbench  -M prepared -n -r -f ./test.sql -T 10 -c 800
-j 10 digoal
transaction type: Custom query
scaling factor: 1
query mode: prepared
number of clients: 800
number of threads: 10
duration: 10 s
number of transactions actually processed: 1035345
tps = 102285.117970 (including connections establishing)
tps = 115616.181725 (excluding connections establishing)
statement latencies in milliseconds:
        6.812371        select 1;

> when client thread big, the tps without connection time wrong!!!!
> number of threads: 100
> duration: 10 s

My 0.02 €:

I cannot test that on my laptop! What kind of server is this?

The thread start time is currently taken before thread creation. The
overhead of creating 100 threads compared to 10 seconds of run is quite
significant and may impact the result. The second issue is that 800
connections and as many server processes must be created, and the time for
that must be quite significant as well against the 10 second run.

So I would mostly classify that as a feature rather than a bug, because
the run is so short.

BTW, I'm planning to submit an enhancement patch to pgbench to improve
measurements, that will include moving the start time after the thread is
started, and possibly help getting more consistent figures.

--
Fabien.
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