[dpdk-users] Bad IO-latency when sending one rte_mbuf at a time

[Sungho Hong]

Hello, I am testing single round-trip latency of a single message using
DPDK and POSIX.

The round-trip latency that I am talking about is that
I send one message in other words rte_mbuf *m[1] from client to server,
and the server echos back to the client .


I have tested this same thing on both POSIX and DPDK, and DPDK performance
is really bad when I do this.

for example when using POSIX
the total round-trip latency is 1275.666667 usec

while when I use DPDK
the total round-trip latency is 61322


In the past, I have only tested DPDK based on run-time on sending a bulk of
data, for example 10 Gigabyte of files, in that case, I remember that DPDK
outperforms POSIX.

I believe that I am using the DPDK in the wrong way, or missing something
very critical. the test cases that I have built can be viewed here.

https://github.com/SungHoHong2/Ceph-Experiment/tree/master/DPDK-FUSE/FUSE-2nd


Would it be possible to know how I can improve the performance of a
round-trip latency of a single message?
(Or is this not ideal for DPDK?)







Best
Sungho Hong