Re: Improving connection scalability (src/backend/storage/ipc/procarray.c)

On 5/25/22 11:07, Ranier Vilela wrote:
> Em qua., 25 de mai. de 2022 às 00:46, Andres Freund <andres@anarazel.de
> <mailto:andres@anarazel.de>> escreveu:
> 
> Hi Andres, thank you for taking a look.
> 
> 
>     On 2022-05-24 12:28:20 -0300, Ranier Vilela wrote:
>     > Linux Ubuntu 64 bits (gcc 9.4)
>     > ./pgbench -M prepared -c  $conns -j $conns -S -n -U postgres
>     >
>     > conns         tps head                 tps patched
>     > 1          2918.004085              3190.810466
>     > 10      12262.415696            17199.862401
>     > 50      13656.724571            18278.194114
>     > 80      14338.202348            17955.336101
>     > 90      16597.510373            18269.660184
>     > 100    17706.775793            18349.650150
>     > 200    16877.067441            17881.250615
>     > 300    16942.260775            17181.441752
>     > 400    16794.514911            17124.533892
>     > 500    16598.502151            17181.244953
>     > 600    16717.935001            16961.130742
>     > 700    16651.204834            16959.172005
>     > 800    16467.546583            16834.591719
>     > 900    16588.241149            16693.902459
>     > 1000  16564.985265            16936.952195
> 
>     17-18k tps is pretty low for pgbench -S. For a shared_buffers
>     resident run, I
>     can get 40k in a single connection in an optimized build. If you're
>     testing a
>     workload >> shared_buffers, GetSnapshotData() isn't the bottleneck. And
>     testing an assert build isn't a meaningful exercise either, unless
>     you have
>     way way higher gains (i.e. stuff like turning O(n^2) into O(n)).
> 
> Thanks for sharing these hits.
> Yes, their 17-18k tps are disappointing.
> 
> 
>     What pgbench scale is this and are you using an optimized build?
> 
> Yes this optimized build.
> CFLAGS='-Wall -Wmissing-prototypes -Wpointer-arith
> -Wdeclaration-after-statement -Werror=vla -Wendif-labels
> -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type
> -Wformat-security -fno-strict-aliasing -fwrapv
> -fexcess-precision=standard -Wno-format-truncation
> -Wno-stringop-truncation -O2'
> from config.log
> 

That can still be assert-enabled build. We need to see configure flags.

> pgbench was initialized with:
> pgbench -i -p 5432 -d postgres
> 
> pgbench -M prepared -c 100 -j 100 -S -n -U postgres

You're not specifying duration/number of transactions to execute. So
it's using just 10 transactions per client, which is bound to give you
bogus results due to not having anything in relcache etc. Use -T 60 or
something like that.

> pgbench (15beta1)
> transaction type: <builtin: select only>
> scaling factor: 1
> query mode: prepared
> number of clients: 100
> number of threads: 100
> 
> The shared_buffers is default:
> shared_buffers = 128MB
> 
> Intel® Core™ i5-8250U CPU Quad Core
> RAM 8GB
> SSD 256 GB
> 

Well, quick results on my laptop (i7-9750H, so not that different from
what you have):

1 = 18908.080126
2 = 32943.953182
3 = 42316.079028
4 = 46700.087645

So something is likely wrong in your setup.


regards

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company