On 2018-10-02 20:23, William Tu wrote:
On Tue, Oct 2, 2018 at 1:01 AM Björn Töpel <bjorn.to...@gmail.com> wrote:
From: Björn Töpel <bjorn.to...@intel.com>
Jeff: Please remove the v1 patches from your dev-queue!
This patch set introduces zero-copy AF_XDP support for Intel's ixgbe
driver.
The ixgbe zero-copy code is located in its own file ixgbe_xsk.[ch],
analogous to the i40e ZC support. Again, as in i40e, code paths have
been copied from the XDP path to the zero-copy path. Going forward we
will try to generalize more code between the AF_XDP ZC drivers, and
also reduce the heavy C&P.
We have run some benchmarks on a dual socket system with two Broadwell
E5 2660 @ 2.0 GHz with hyperthreading turned off. Each socket has 14
cores which gives a total of 28, but only two cores are used in these
experiments. One for TR/RX and one for the user space application. The
memory is DDR4 @ 2133 MT/s (1067 MHz) and the size of each DIMM is
8192MB and with 8 of those DIMMs in the system we have 64 GB of total
memory. The compiler used is GCC 7.3.0. The NIC is Intel
82599ES/X520-2 10Gbit/s using the ixgbe driver.
Below are the results in Mpps of the 82599ES/X520-2 NIC benchmark runs
for 64B and 1500B packets, generated by a commercial packet generator
HW blasting packets at full 10Gbit/s line rate. The results are with
retpoline and all other spectre and meltdown fixes.
AF_XDP performance 64B packets:
Benchmark XDP_DRV with zerocopy
rxdrop 14.7
txpush 14.6
l2fwd 11.1
AF_XDP performance 1500B packets:
Benchmark XDP_DRV with zerocopy
rxdrop 0.8
l2fwd 0.8
XDP performance on our system as a base line.
64B packets:
XDP stats CPU Mpps issue-pps
XDP-RX CPU 16 14.7 0
1500B packets:
XDP stats CPU Mpps issue-pps
XDP-RX CPU 16 0.8 0
The structure of the patch set is as follows:
Patch 1: Introduce Rx/Tx ring enable/disable functionality
Patch 2: Preparatory patche to ixgbe driver code for RX
Patch 3: ixgbe zero-copy support for RX
Patch 4: Preparatory patch to ixgbe driver code for TX
Patch 5: ixgbe zero-copy support for TX
Changes since v1:
* Removed redundant AF_XDP precondition checks, pointed out by
Jakub. Now, the preconditions are only checked at XDP enable time.
* Fixed a crash in the egress path, due to incorrect usage of
ixgbe_ring queue_index member. In v2 a ring_idx back reference is
introduced, and used in favor of queue_index. William reported the
crash, and helped me smoke out the issue. Kudos!
Thanks! I tested this series and no more crash.
Thank you for spending time on this!
The number is pretty good (*without* spectre and meltdown fixes)
model name : Intel(R) Xeon(R) CPU E5-2440 v2 @ 1.90GHz, total 16 cores/
AF_XDP performance 64B packets:
Benchmark XDP_DRV with zerocopy
rxdrop 20
txpush 18
l2fwd 20
What is 20 here? Given that 14.8Mpps is maximum for 64B@10Gbit/s for
one queue, is this multiple queues? Is this xdpsock or OvS with AF_XDP?
Cheers,
Björn
Regards,
William
