From: "Tomáš Jánský" <tomas.jansky@flowmon.com>
To: Paul T <paultop6@outlook.com>
Cc: "users@dpdk.org" <users@dpdk.org>
Subject: Re: [dpdk-users] X710 DA2 (2x10G) performance 64B packet
Date: Thu, 21 Mar 2019 15:44:52 +0100 [thread overview]
Message-ID: <CAPP7y6wic9-Ca_GpXtTRvLe8L9PAzEO5fEHto3AjRd8os6wytg@mail.gmail.com> (raw)
In-Reply-To: <DB7PR06MB482725FDAEEF8AB3F61126B889420@DB7PR06MB4827.eurprd06.prod.outlook.com>
Thanks Paul for another suggestion.
This is my boot setup now:
default_hugepagesz=1G hugepagesz=1G isolcpus=1-4 nohz_full=1-4 rcu_nocbs=1-4
and I am using 16x 1GB hugepages on the NUMA node.
But so far no improvement.
Tomas
On Thu, Mar 21, 2019 at 3:31 PM Paul T <paultop6@outlook.com> wrote:
> 1GB huge page chunks instead of 2MB would also be worth a try
>
> ------------------------------
> *From:* Tomáš Jánský <tomas.jansky@flowmon.com>
> *Sent:* 21 March 2019 14:00
> *To:* Paul T
> *Cc:* users@dpdk.org
> *Subject:* Re: [dpdk-users] X710 DA2 (2x10G) performance 64B packet
>
> Hi Paul,
>
> thank you for your suggestion.
> I tried isolating the cores; however, the improvement was negligible.
>
> Tomas
>
> On Thu, Mar 21, 2019 at 12:50 PM Paul T <paultop6@outlook.com> wrote:
>
> Hi Tomas,
>
> I would isolate the CPUs in which the dpdk threads are running from the
> linux schedular. The low packet drop at 64B makes me thing its context
> switching happen on the core because of the linux scheduler.
>
> Use the following command in the linux command line params in your grub
> config:
> isolcpus=cpus to isolate, e.g. 1,3,4 or 1-4
>
> Regards
>
> Paul
>
> Message: 3
> Date: Thu, 21 Mar 2019 10:53:34 +0100
> From: Tom?? J?nsk? <tomas.jansky@flowmon.com>
> To: users@dpdk.org
> Subject: [dpdk-users] X710 DA2 (2x10G) performance 64B packets
> Message-ID:
> <CAPP7y6z13qFR-34+-Xn97ru5jOnaVAV7s=6WPgk_j=
> 9CLMQrSQ@mail.gmail.com>
> Content-Type: text/plain; charset="UTF-8"
>
> Hello DPDK users,
>
> I am having an issue concerning the performance of X710 DA2 (2x10G) NIC
> when using testpmd (and also l2fwd) application on both ports.
>
> HW and SW parameters:
> CPUs: Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHz x16
> Disabled hyperthreading.
> All used lcores and ports are on the same NUMA node (0).
> Hugepages: 1024x 2MB on the NUMA node 0.
> RAM: 64 GB
>
> DPDK version: 18.05.1
> Modue: IGB UIO
> GCC version: 4.8.5
>
> When using testpmd application only on one port:
> ./testpmd -b 0000:04:00.0 -n 4 --lcore=0@0,2@2 -- --socket-num=0
> --nb-cores=1 --nb-ports=1 --numa --forward-mode=rxonly
>
> 14.63 Mpps (64B packet length) - 0.01% packets dropped
>
> When using testmpd on both ports:
> ./testpmd -n 4 --lcore=0@0,2@2,4@4 -- --socket-num=0 --nb-cores=2
> --nb-ports=2 --numa --forward-mode=rxonly
>
> 28.08 Mpps (64B packet length) - 3.47% packets dropped
>
> Does anybody have an explanation why am I experiencing this performance
> drop?
> Any suggestion would be much appreciated.
>
> Thank you
> Tomas
>
>
prev parent reply other threads:[~2019-03-21 14:45 UTC|newest]
Thread overview: 4+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-03-21 11:50 Paul T
2019-03-21 14:00 ` Tomáš Jánský
2019-03-21 14:31 ` Paul T
2019-03-21 14:44 ` Tomáš Jánský [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAPP7y6wic9-Ca_GpXtTRvLe8L9PAzEO5fEHto3AjRd8os6wytg@mail.gmail.com \
--to=tomas.jansky@flowmon.com \
--cc=paultop6@outlook.com \
--cc=users@dpdk.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).