vhost_net: batch used ring update in rx
authorJason Wang <jasowang@redhat.com>
Tue, 9 Jan 2018 10:27:45 +0000 (18:27 +0800)
committerDavid S. Miller <davem@davemloft.net>
Wed, 10 Jan 2018 20:04:30 +0000 (15:04 -0500)
commite2b3b35eb9896f26c98b9a2c047d9111638059a2
tree37e02ed712f920656f9a7f5bf75b6b8ab00a0185
parent65d51f26823ee286602ae850ec575a63e41d7fd7
vhost_net: batch used ring update in rx

This patch tries to batched used ring update during RX. This is pretty
fit for the case when guest is much faster (e.g dpdk based
backend). In this case, used ring is almost empty:

- we may get serious cache line misses/contending on both used ring
  and used idx.
- at most 1 packet could be dequeued at one time, batching in guest
  does not make much effect.

Update used ring in a batch can help since guest won't access the used
ring until used idx was advanced for several descriptors and since we
advance used ring for every N packets, guest will only need to access
used idx for every N packet since it can cache the used idx. To have a
better interaction for both batch dequeuing and dpdk batching,
VHOST_RX_BATCH was used as the maximum number of descriptors that
could be batched.

Test were done between two machines with 2.40GHz Intel(R) Xeon(R) CPU
E5-2630 connected back to back through ixgbe. Traffic were generated
on one remote ixgbe through MoonGen and measure the RX pps through
testpmd in guest when do xdp_redirect_map from local ixgbe to
tap. RX pps were increased from 3.05 Mpps to 4.00 Mpps (about 31%
improvement).

One possible concern for this is the implications for TCP (especially
latency sensitive workload). Result[1] does not show obvious changes
for most of the netperf test (RR, TX, and RX). And we do get some
improvements for RX on some specific size.

Guest RX:

size/sessions/+thu%/+normalize%
   64/     1/   +2%/   +2%
   64/     2/   +2%/   -1%
   64/     4/   +1%/   +1%
   64/     8/    0%/    0%
  256/     1/   +6%/   -3%
  256/     2/   -3%/   +2%
  256/     4/  +11%/  +11%
  256/     8/    0%/    0%
  512/     1/   +4%/    0%
  512/     2/   +2%/   +2%
  512/     4/    0%/   -1%
  512/     8/   -8%/   -8%
 1024/     1/   -7%/  -17%
 1024/     2/   -8%/   -7%
 1024/     4/   +1%/    0%
 1024/     8/    0%/    0%
 2048/     1/  +30%/  +14%
 2048/     2/  +46%/  +40%
 2048/     4/    0%/    0%
 2048/     8/    0%/    0%
 4096/     1/  +23%/  +22%
 4096/     2/  +26%/  +23%
 4096/     4/    0%/   +1%
 4096/     8/    0%/    0%
16384/     1/   -2%/   -3%
16384/     2/   +1%/   -4%
16384/     4/   -1%/   -3%
16384/     8/    0%/   -1%
65535/     1/  +15%/   +7%
65535/     2/   +4%/   +7%
65535/     4/    0%/   +1%
65535/     8/    0%/    0%

TCP_RR:

size/sessions/+thu%/+normalize%
    1/     1/    0%/   +1%
    1/    25/   +2%/   +1%
    1/    50/   +4%/   +1%
   64/     1/    0%/   -4%
   64/    25/   +2%/   +1%
   64/    50/    0%/   -1%
  256/     1/    0%/    0%
  256/    25/    0%/    0%
  256/    50/   +4%/   +2%

Guest TX:

size/sessions/+thu%/+normalize%
   64/     1/   +4%/   -2%
   64/     2/   -6%/   -5%
   64/     4/   +3%/   +6%
   64/     8/    0%/   +3%
  256/     1/  +15%/  +16%
  256/     2/  +11%/  +12%
  256/     4/   +1%/    0%
  256/     8/   +5%/   +5%
  512/     1/   -1%/   -6%
  512/     2/    0%/   -8%
  512/     4/   -2%/   +4%
  512/     8/   +6%/   +9%
 1024/     1/   +3%/   +1%
 1024/     2/   +3%/   +9%
 1024/     4/    0%/   +7%
 1024/     8/    0%/   +7%
 2048/     1/   +8%/   +2%
 2048/     2/   +3%/   -1%
 2048/     4/   -1%/  +11%
 2048/     8/   +3%/   +9%
 4096/     1/   +8%/   +8%
 4096/     2/    0%/   -7%
 4096/     4/   +4%/   +4%
 4096/     8/   +2%/   +5%
16384/     1/   -3%/   +1%
16384/     2/   -1%/  -12%
16384/     4/   -1%/   +5%
16384/     8/    0%/   +1%
65535/     1/    0%/   -3%
65535/     2/   +5%/  +16%
65535/     4/   +1%/   +2%
65535/     8/   +1%/   -1%

Signed-off-by: Jason Wang <jasowang@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/vhost/net.c