2022-05-30 07:38:38

by Yacan Liu

[permalink] [raw]
Subject: SMC-R problem under multithread

Hi experts,

I recently used memcached to test the performance of SMC-R relative to TCP, but the results
are confusing me. When using multithread on the server side, the performance of SMC-R is not as good as TCP.

Specifically, I tested 4 scenarios with server thread: 1\2\4\8. The client uses 8threads fixedly.

server: (smc_run) memcached -t 1 -m 16384 -p [SERVER-PORT] -U 0 -F -c 10240 -o modern
client: (smc-run) memtier_benchmark -s [SERVER-IP] -p [SERVER-PORT] -P memcache_text --random-data --data-size=100 --data-size-pattern=S --key-minimum=30 --key-maximum=100 -n 5000000 -t 8

The result is as follows:

SMC-R:

server-thread ops/sec client-cpu server-cpu
1 242k 220% 97%
2 362k 241% 128%
4 378k 242% 160%
8 395k 242% 210%

TCP:
server-thread ops/sec client-cpu server-cpu
1 185k 224% 100%
2 435k 479% 200%
4 780k 731% 400%
8 938k 800% 659%

It can be seen that as the number of threads increases, the performance increase of SMC-R is much slower than that of TCP.

Am I doing something wrong? Or is it only when CPU resources are tight that SMC-R has a significant advantage ?

Any suggestions are welcome.


Thanks & Regards,
Yacan.



2022-05-30 09:07:58

by Tony Lu

[permalink] [raw]
Subject: Re: SMC-R problem under multithread

On Mon, May 30, 2022 at 11:16:04AM +0800, [email protected] wrote:
> Hi experts,
>
> I recently used memcached to test the performance of SMC-R relative to TCP, but the results
> are confusing me. When using multithread on the server side, the performance of SMC-R is not as good as TCP.
>
> Specifically, I tested 4 scenarios with server thread: 1\2\4\8. The client uses 8threads fixedly.
>
> server: (smc_run) memcached -t 1 -m 16384 -p [SERVER-PORT] -U 0 -F -c 10240 -o modern
> client: (smc-run) memtier_benchmark -s [SERVER-IP] -p [SERVER-PORT] -P memcache_text --random-data --data-size=100 --data-size-pattern=S --key-minimum=30 --key-maximum=100 -n 5000000 -t 8
>
> The result is as follows:
>
> SMC-R:
>
> server-thread ops/sec client-cpu server-cpu
> 1 242k 220% 97%
> 2 362k 241% 128%
> 4 378k 242% 160%
> 8 395k 242% 210%
>
> TCP:
> server-thread ops/sec client-cpu server-cpu
> 1 185k 224% 100%
> 2 435k 479% 200%
> 4 780k 731% 400%
> 8 938k 800% 659%
>
> It can be seen that as the number of threads increases, the performance increase of SMC-R is much slower than that of TCP.
>
> Am I doing something wrong? Or is it only when CPU resources are tight that SMC-R has a significant advantage ?
>
> Any suggestions are welcome.

Hi Yacan,

This result matches some of our scenarios to some extent. Let's talk
about this result first.

Based on your benchmark, the biggest factor affecting performance seems
that the CPU resource is limited. As the number of threads increased,
neither CPU usage nor performance metrics improved, and CPU is limited
to about 200-250%. To make it clear, could you please give out more
metrics about per-CPU (usr / sys / hi / si) and memcached process usage.

Secondly, it seems that there is lots of connections in this test.
If it takes too much time to establish a connection, or the number of
final connections does not reach the specified value, the result will be
greatly affected. Could you please give out more details about the
connections numbers during benchmark?

We have noticed SMC has some limitations in multiple threads and many
connections. This benchmark happens to be basically in line with this
scenario. In general, there are some aspects in brief:
1. control path (connection setup and dismiss) is not as fast as TCP;
2. data path (lock contention, CQ spreading, etc.) needs further improvement;

About CPU limitation, SMC use one CQ and one core to handle data
transmission, which cannot spread workload over multiple cores. There is
is an early temporary solution [1], which also need to improve (new CQ
API, WR refactor). With this early solution, it shows several times the
performance improvement.

About the improvement of connection setup, you can see [2] for more
details, which is still a proposal now, and we are working on it now.
This show considerable performance boost.

[1] https://lore.kernel.org/all/[email protected]/
[2] https://lore.kernel.org/all/[email protected]/

Thanks,
Tony LU