Re: [PATCH v4 net-next 2/3] net/udp: Add 4-tuple hash list basis

[Philo Lu <lulie@linux.alibaba.com>]

On 2024/10/16 15:45, Paolo Abeni wrote:
> On 10/16/24 08:30, Philo Lu wrote:
>> On 2024/10/14 18:07, Paolo Abeni wrote:
>>> It would be great if you could please share some benchmark showing the
>>> raw max receive PPS performances for unconnected sockets, with and
>>> without this series applied, to ensure this does not cause any real
>>> regression for such workloads.
>>>
>>
>> Tested using sockperf tp with default msgsize (14B), 3 times for w/ and
>> w/o the patch set, and results show no obvious difference:
>>
>> [msg/sec]  test1    test2    test3    mean
>> w/o patch  514,664  519,040  527,115  520.3k
>> w/  patch  516,863  526,337  527,195  523.5k (+0.6%)
>>
>> Thank you for review, Paolo.
> 
> Are the value in packet per seconds, or bytes per seconds? Are you doing 
> a loopback test or over the wire? The most important question is: is the 
> receiver side keeping (at least) 1 CPU fully busy? Otherwise the test is 
> not very relevant.
> 
> It looks like you have some setup issue, or you are using a relatively 
> low end H/W: the expected packet rate for reasonable server H/W is well 
> above 1M (possibly much more than that, but I can't put my hands on 
> recent H/W, so I can't provide a more accurate figure).
> 
> A single socket, user-space, UDP sender is usually unable to reach such 
> tput without USO, and even with USO you likely need to do an over-the- 
> wire test to really be able to keep the receiver fully busy. AFAICS 
> sockperf does not support USO for the sender.
> 
> You could use the udpgso_bench_tx/udpgso_bench_rx pair from the net 
> selftests directory instead.
> 
> Or you could use pktgen as traffic generator.
> 

I test it again with udpgso_bench_tx/udpgso_bench_rx. In server, 2 cpus 
are involved, one for udpgso_bench_rx and the other for nic rx queue so 
that the si of nic rx cpu is 100%. udpgso_bench_tx runs with payload 
size 20, and the tx pps is larger than rx ensuring rx is the bottleneck.

The outputs of udpgso_bench_rx:
[without patchset]
udp rx:     20 MB/s  1092546 calls/s
udp rx:     20 MB/s  1095051 calls/s
udp rx:     20 MB/s  1094136 calls/s
udp rx:     20 MB/s  1098860 calls/s
udp rx:     20 MB/s  1097963 calls/s
udp rx:     20 MB/s  1097460 calls/s
udp rx:     20 MB/s  1098370 calls/s
udp rx:     20 MB/s  1098089 calls/s
udp rx:     20 MB/s  1095330 calls/s
udp rx:     20 MB/s  1095486 calls/s

[with patchset]
udp rx:     21 MB/s  1105533 calls/s
udp rx:     21 MB/s  1105475 calls/s
udp rx:     21 MB/s  1104244 calls/s
udp rx:     21 MB/s  1105600 calls/s
udp rx:     21 MB/s  1108019 calls/s
udp rx:     21 MB/s  1101971 calls/s
udp rx:     21 MB/s  1104147 calls/s
udp rx:     21 MB/s  1104874 calls/s
udp rx:     21 MB/s  1101987 calls/s
udp rx:     21 MB/s  1105500 calls/s

The averages w/ and w/o the patchset are 1104735 and 1096329, the gap is 
0.8%, which I think is negligible.

Besides, perf shows ~0.6% higher cpu consumption of __udp4_lib_lookup() 
with this patchset (increasing from 5.7% to 6.3%).

Thanks.
-- 
Philo