This is a new set of patches to optimize the mergeable Rx code path.
No refactoring (rewrite) was made this time. It just applies some
findings from Zhihong (kudos to him!) that could improve the mergeable
Rx path on the old code.
The two major factors that could improve the performance greatly are:
- copy virtio header together with packet data. This could remove
the buubbles between the two copy to optimize the cache access.
This is implemented in patch 2 "vhost: optimize cache access"
- shadow used ring update and update them at once
The basic idea is to update used ring in a local buffer and flush
them to the virtio used ring at once in the end. Again, this is
for optimizing the cache access.
This is implemented in patch 5 "vhost: shadow used ring update"
The two optimizations could yield 40+% performance in micro testing
and 20+% in PVP case testing with 64B packet size.
Besides that, there are some tiny optimizations, such as prefetch
avail ring (patch 6) and retrieve avail head once (patch 7).
Note: the shadow used ring tech could also be applied to the non-mrg
Rx path (and even the dequeu) path. I didn't do that for two reasons:
- we already update used ring in batch in both path: it's not shadowed
first though.
- it's a bit too late too make many changes at this stage: RC1 is out.
Please help testing.
Thanks.
--yliu
Cc: Jianbo Liu <jianbo.liu at linaro.org>
---
Yuanhan Liu (4):
vhost: simplify mergeable Rx vring reservation
vhost: use last avail idx for avail ring reservation
vhost: prefetch avail ring
vhost: retrieve avail head once
Zhihong Wang (3):
vhost: remove useless volatile
vhost: optimize cache access
vhost: shadow used ring update
lib/librte_vhost/vhost.c | 13 ++-
lib/librte_vhost/vhost.h | 5 +-
lib/librte_vhost/vhost_user.c | 23 +++--
lib/librte_vhost/virtio_net.c | 193 +++++++++++++++++++++++++-----------------
4 files changed, 149 insertions(+), 85 deletions(-)
--
1.9.0
