LPC2018 - Path to DPDK speeds for AF XDP

speaker: Björn Töpel (Intel), Magnus Karlsson (Intel)

AF_XDP is a new socket type for raw frames to be introduced in 4.18
(in linux-next at the time of writing). The current code base offers
throughput numbers north of 20 Mpps per application core for 64-byte
packets on our system, however there are a lot of optimizations that
could be performed in order to increase this even further. The focus
of this paper is the performance optimizations we need to make in
AF_XDP to get it to perform as fast as DPDK.

We present optimization that fall into two broad categories: ones that
are seamless to the application and ones that requires additions to
the uapi. In the first category we examine the following:

Loosen the requirement for having an XDP program. If the user does
not need an XDP program and there is only one AF_XDP socket bound to
a particular queue, we do not need an XDP program. This should cut
out quite a number of cycles from the RX path.

Wire up busy poll from user space. If the application writer is
using epoll() and friends, this has the potential benefit of
removing the coherency communication between the RX (NAPI) core and
the application core as everything is now done on a single
core. Should improve performance for a number of use cases. Maybe it
is worth revisiting the old idea of threaded NAPI in this context
too.

Optimize for high instruction cache usage through batching as has
been explored in for example Cisco's VPP stack and Edward Cree in
his net-next RFC "Handle multiple received packets at each stage".

In the uapi extensions category we examine the following
optimizations:

Support a new mode for NICs with in-order TX completions. In this
mode, the completion queue would not be used. Instead the
application would simply look at the pointer in the TX queue to see
if a packet has been completed. In this mode, we do not need any
backpreassure between the completion queue and the TX queue and we
do not need to populate or publish anything in the completion queue
as it is not used. Should improve the performance of TX for in-order
NICs significantly.

Introduce the "type-writer" model where each chunk can contain
multiple packets. This is the model that e.g., Chelsio has in its
NICs. But experiments show that this mode also can provide better
performance for regular NICs as there are fewer transactions on the
queues. Requires a new flag to be introduced in the options field of
the descriptor.

With these optimization, we believe we can reach our goal of close to
40 Mpps of throughput for 64-byte packets in zero-copy mode. Full
analysis with performance numbers will be presented in the final
paper.