[dpdk-users] multi core performance problem

[dpdk-users] multi core performance problem

[Sepehr Nazary]

Hi
I run dpdk multi_process (symetric_mp) example and see this result:

NumberOfCore  Throughput
1              9986
2              9987
3              9808
4              9988
5              9645
6              9171
7              9030
8              8517
9              8537
10              8415
11              7832
12              7985
13              7569
14              7285
15              7624
16              6994

I run pktgen with range on and packet size is 64 byte.
I sure RSS works correctly and packets are distributed uniformly between
cores. Packet drops is 0, as well.

Why performance decrease ??

configuration :
dpdk-2.2.0, pktgen 3.0.0
NIC model : 82599ES 10-Gigabit
CPU model : Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz
Linux version 3.16.0-4-amd64
Two NUMA Node:
socket 0 : 16G RAM , 18 CORE , 2 NIC ,run DPDK symetric_mp example
socket 1 : 16G RAM , 18 CORE , 2 NIC ,run Pktgen (range on)

Re: [dpdk-users] multi core performance problem

[Muhammad Zain-ul-Abideen]

I am no expert with memory bandwidth and hardware acceleration. But what i
can infer from this issue is that 1 of 2 things
1. Either QPI is being used and the maximum bandwidth is achieved
2. Or RAM has reached it's maximum bandwidth and this is the reason the
higher orders core are getting more bandwidth then lower order

But keith, and Thomas can tell you more on this subject

Re: [dpdk-users] multi core performance problem

[Masoud Moshref Javadi]

Check the CPU core frequencies as you increase the number of cores that run
in 100%.
Because of heating and dynamic frequency in Cores, as you increase the
number of cores that process packets, the frequency can go down.

Also make sure that for higher number of cores, you use large enough mem
pool.

On Mon, Jul 4, 2016 at 10:34 AM Muhammad Zain-ul-Abideen <zain2294@gmail.com>
wrote:

> I am no expert with memory bandwidth and hardware acceleration. But what i
> can infer from this issue is that 1 of 2 things
> 1. Either QPI is being used and the maximum bandwidth is achieved
> 2. Or RAM has reached it's maximum bandwidth and this is the reason the
> higher orders core are getting more bandwidth then lower order
>
> But keith, and Thomas can tell you more on this subject
>