[dpdk-users] Throughput dropping by half with 1B increase in packet size
keith.wiles at intel.com
Fri Feb 9 16:55:28 CET 2018
> On Feb 9, 2018, at 9:39 AM, Theano <theanosaurus at gmail.com> wrote:
> Hi All,
> We are eventually going to build an application over DPDK, and we've
> started out with a very simple pktgen-like application that sends and
> receives packets to a different machine running l2fwd.
> With packets up to 64B (60B headers and payload + 4B of FCS), l2fwd shows
> line rate packet forwarding of this application's packets. As soon as the
> packet size increases to 65B, the throughput (in pps) halves. With
> dpdk-pktgen we observe a much more graceful decline in packet throughput as
> packet size increases.
> I examined the dpdk-pktgen code and can't find any significant differences
> between how it constructs its packets vs. how we do it, but of course I'm
> missing something.
> Has anyone seen this before and/or has any ideas on how to identify the
Adding one byte will push you to have two cache lines used and that can reduce performance. Make sure the packets are aligned to a 64 byte boundary.
Other then that I see no reason for the 1/2 performance without comparing the code. Pktgen-dpdk I spent a fair amount of time looking at the performance, but I also had to play with burst size and the number of packet sent to the NIC they too need to be cached aligned or multiples of a cache line.
> We are using DPDK version 17.02.1; please let me know if any other
> information would be helpful.
More information about the users