[dpdk-users] Pktgen performance tricks

[Wiles, Keith]

On 6/23/16, 12:33 PM, "users on behalf of Marko Niinimaki" <users-bounces at dpdk.org on behalf of manzikki at gmail.com> wrote:

>Hi, apologies for such a simple question:
>I'm able to send about 4M custom packets by Pktgen using the following
>parameters:
>pktgen -c 0x3 -n 4 --  -m "1.0"
>Hoping to speed up, I added some more cores with
>pktgen -c 0x7 -n 4 --  -m "1.0, 2.1"
>but this does not really improve the performance. I've been tinkering a
>bit, but no great breaktrough. What would be the best approach/parameters?
>I'm only interested in TX this time.

Can you tell me more about the system like how many cores maybe run the cpu_layout.py script in the tools directory?
Also the type of NIC being used 1G, 10G or 40G?

Normally the best performance is to split the TX and RX of a single port on to two different physical cores. If you have hyper-threading enabled then make sure the two lcores being used are on the same socket, but two different cores.

Pktgen is able to send close to 10GB 64 byte frames on a single XEON core.

Try pktgen –l 1-3 –n 4 -- -m “[2:3].0” This assume lcore 2 and 3 are on different physical cores on the same socket. Note skipping lcore 0 for Linux and lcore 1 is used for Pktgen display and timers.

If you need two ports pktgen –l 1-5 –n 4 -- -m”[2:3].0, [4:5].1” 

>
>Thanks in advance,
>Mark
>