On Thu, Jul 1, 2021 at 9:20 PM Jay Vosburgh [off-list ref] wrote:

joamaki@gmail.com wrote:

quoted

From: Jussi Maki <redacted>

This patchset introduces XDP support to the bonding driver.

The motivation for this change is to enable use of bonding (and
802.3ad) in hairpinning L4 load-balancers such as [1] implemented with
XDP and also to transparently support bond devices for projects that
use XDP given most modern NICs have dual port adapters.  An alternative
to this approach would be to implement 802.3ad in user-space and
implement the bonding load-balancing in the XDP program itself, but
is rather a cumbersome endeavor in terms of slave device management
(e.g. by watching netlink) and requires separate programs for native
vs bond cases for the orchestrator. A native in-kernel implementation
overcomes these issues and provides more flexibility.

Below are benchmark results done on two machines with 100Gbit
Intel E810 (ice) NIC and with 32-core 3970X on sending machine, and
16-core 3950X on receiving machine. 64 byte packets were sent with
pktgen-dpdk at full rate. Two issues [2, 3] were identified with the
ice driver, so the tests were performed with iommu=off and patch [2]
applied. Additionally the bonding round robin algorithm was modified
to use per-cpu tx counters as high CPU load (50% vs 10%) and high rate
of cache misses were caused by the shared rr_tx_counter. Fix for this
has been already merged into net-next. The statistics were collected
using "sar -n dev -u 1 10".

-----------------------|  CPU  |--| rxpck/s |--| txpck/s |----
without patch (1 dev):
  XDP_DROP:              3.15%      48.6Mpps
  XDP_TX:                3.12%      18.3Mpps     18.3Mpps
  XDP_DROP (RSS):        9.47%      116.5Mpps
  XDP_TX (RSS):          9.67%      25.3Mpps     24.2Mpps
-----------------------
with patch, bond (1 dev):
  XDP_DROP:              3.14%      46.7Mpps
  XDP_TX:                3.15%      13.9Mpps     13.9Mpps
  XDP_DROP (RSS):        10.33%     117.2Mpps
  XDP_TX (RSS):          10.64%     25.1Mpps     24.0Mpps
-----------------------
with patch, bond (2 devs):
  XDP_DROP:              6.27%      92.7Mpps
  XDP_TX:                6.26%      17.6Mpps     17.5Mpps
  XDP_DROP (RSS):       11.38%      117.2Mpps
  XDP_TX (RSS):         14.30%      28.7Mpps     27.4Mpps
--------------------------------------------------------------

        To be clear, the fact that the performance numbers for XDP_DROP
and XDP_TX are lower for "with patch, bond (1 dev)" than "without patch
(1 dev)" is expected, correct?

Yes that is correct. With the patch the ndo callback for choosing the
slave device is invoked which in this test (mode=xor) hashes L2&L3
headers (I seem to have failed to mention this in the original
message). In round-robin mode I recall it being about 16Mpps versus
the 18Mpps without the patch. I did also try "INDIRECT_CALL" to avoid
going via ndo_ops, but that had no discernible effect.