[dpdk-users] [ConnectX-5 MCX515A-CCAT / MCX516A-CCAT] Can only generate 53Gb/s with 64B packets
Yan Lei
l.yan at epfl.ch
Sat Mar 14 19:37:16 CET 2020
Hi Sachin,
Thanks a lot for the answer. The issue is resolved, was able to get 98Gb/s with 64B packets after set pci maxReadRequest to 1024 and turn off NIC flow control. These optimization settings are actually posted in the mlx5 PMD guide, my bad to have ignored them...
Cheers,
Lei
________________________________
From: sachin gupta <sachingg at yahoo.com>
Sent: Thursday, March 12, 2020 7:45:31 AM
To: users at dpdk.org; Yan Lei
Subject: Re: [dpdk-users] [ConnectX-5 MCX515A-CCAT / MCX516A-CCAT] Can only generate 53Gb/s with 64B packets
Hi Lei,
The smaller the Packet size, the more the number of Packets per second. I believe this is the inherent problem in all systems, even the ones which have proprietary hardware.
In general applications which uses such small packets are rare and you will see a mix of traffic in the system.
Regards
Sachin
On Thursday, March 12, 2020, 5:10:14 AM GMT+5:30, Yan Lei <l.yan at epfl.ch> wrote:
Hi, I am currently struggling in getting more than 53Gb/s with 64B packets on both of the MCX515A-CCAT and MCX516A-CCAT adapter when running a DPDK app that generates and transmits packets. WIth 256B packets I can get 98Gb/s.
Has anyone saw the same performance on these NICs? I checked the perf. report on https://core.dpdk.org/perf-reports/ but there are no numbers of these NICs.
Is this inherent limitation of these NICs (only reach 100Gb/s with larger packets)? If not, which firmware/driver/DPDK/system configurations could I tune to get 100Gb/s with 64B packets? My setup is as following: - CPU: E5-2697 v3 (14 cores, SMT disabled, CPU frequency fixed @ 2.6 GHz) - NIC: Mellanox MCX515A-CCAT / MCX516A-CCAT (Using only one port for TX, installed on PCIe Gen3 x16) - DPDK: 19.05 - RDMA-CORE: v28.0 - Kernel: 5.3.0 - OS: Ubuntu 18.04 - Firmware: 16.26.1040 I measured the TX rate with DPDK's testpmd: $ ./testpmd -l 3-13 -n 4 -w 02:00.0 -- -i --port-topology=chained --nb-ports=1 --rxq=10 --txq=10 --nb-cores=10 --burst=128 --rxd=512 --txd=512 --mbcache=512 --forward-mode=txonly So 10 cores generating and transmits 64B packets on 10 NIC queues. Your feedbacks will be much appreciated.
Thanks, Lei
More information about the users
mailing list