This series tries to add basic busy polling for vhost net. The idea is
simple: at the end of tx/rx processing, busy polling for new tx added
descriptor and rx receive socket for a while. The maximum number of
time (in us) could be spent on busy polling was specified ioctl.
Test A were done through:
- 50 us as busy loop timeout
- Netperf 2.6
- Two machines with back to back connected mlx4
- Guest with 8 vcpus and 1 queue
Results:
- TCP_RR was imporved obviously (at most 27%). And cpu utilizaton was
also improved in this case.
- No obvious differences in Guest RX throughput.
- Guest TX throughput was also improved.
TCP_RR:
size/session/+thu%/+normalize%/+tpkts%/+rpkts%/+ioexits%/
1/ 1/ +27%/ 0%/ +27%/ +27%/ +25%
1/ 50/ +2%/ +1%/ +2%/ +2%/ -4%
1/ 100/ +2%/ +1%/ +3%/ +3%/ -14%
1/ 200/ +2%/ +2%/ +5%/ +5%/ -15%
64/ 1/ +20%/ -13%/ +20%/ +20%/ +20%
64/ 50/ +17%/ +14%/ +16%/ +16%/ -11%
64/ 100/ +14%/ +12%/ +14%/ +14%/ -35%
64/ 200/ +16%/ +15%/ +9%/ +9%/ -28%
256/ 1/ +19%/ -6%/ +19%/ +19%/ +18%
256/ 50/ +18%/ +15%/ +16%/ +16%/ +3%
256/ 100/ +11%/ +9%/ +12%/ +12%/ -1%
256/ 200/ +5%/ +8%/ +4%/ +4%/ +64%
512/ 1/ +20%/ 0%/ +20%/ +20%/ -2%
512/ 50/ +12%/ +10%/ +12%/ +12%/ +8%
512/ 100/ +11%/ +7%/ +10%/ +10%/ -5%
512/ 200/ +3%/ +2%/ +3%/ +3%/ -5%
1024/ 1/ +19%/ -2%/ +19%/ +19%/ +18%
1024/ 50/ +13%/ +10%/ +12%/ +12%/ 0%
1024/ 100/ +9%/ +8%/ +8%/ +8%/ -16%
1024/ 200/ +3%/ +4%/ +3%/ +3%/ -14%
Guest RX:
size/session/+thu%/+normalize%/+tpkts%/+rpkts%/+ioexits%/
64/ 1/ -12%/ -10%/ +2%/ +1%/ +42%
64/ 4/ -3%/ -5%/ +2%/ -1%/ 0%
64/ 8/ -1%/ -5%/ -1%/ -2%/ 0%
512/ 1/ +5%/ -13%/ +6%/ +9%/ +17%
512/ 4/ -3%/ -9%/ +6%/ +4%/ -14%
512/ 8/ -2%/ -7%/ 0%/ 0%/ -1%
1024/ 1/ +18%/ +31%/ -12%/ -11%/ -31%
1024/ 4/ 0%/ -9%/ -1%/ -6%/ -7%
1024/ 8/ -3%/ -8%/ -2%/ -4%/ 0%
2048/ 1/ 0%/ -1%/ 0%/ -4%/ +5%
2048/ 4/ 0%/ +2%/ 0%/ 0%/ 0%
2048/ 8/ 0%/ -6%/ 0%/ -3%/ -1%
4096/ 1/ -1%/ +2%/ -14%/ -5%/ +8%
4096/ 4/ 0%/ +1%/ 0%/ +1%/ -1%
4096/ 8/ -1%/ -1%/ -2%/ -2%/ -3%
16384/ 1/ 0%/ 0%/ +4%/ +5%/ 0%
16384/ 4/ 0%/ +5%/ +7%/ +9%/ 0%
16384/ 8/ +1%/ +1%/ +3%/ +3%/ +2%
65535/ 1/ 0%/ +12%/ -1%/ +2%/ -2%
65535/ 4/ 0%/ 0%/ -2%/ -2%/ +2%
65535/ 8/ -1%/ -1%/ -4%/ -4%/ 0%
Guest TX:
size/session/+thu%/+normalize%/+tpkts%/+rpkts%/+ioexits%/
64/ 1/ -16%/ -21%/ -2%/ -12%/ +1%
64/ 4/ -6%/ -2%/ -1%/ +6%/ -7%
64/ 8/ +4%/ +4%/ -2%/ +1%/ +30%
512/ 1/ -32%/ -33%/ -11%/ +62%/ +314%
512/ 4/ +30%/ +20%/ -22%/ -17%/ -14%
512/ 8/ +24%/ +12%/ -21%/ -10%/ -6%
1024/ 1/ +1%/ -7%/ +2%/ +51%/ +75%
1024/ 4/ +10%/ +9%/ -11%/ -19%/ -10%
1024/ 8/ +13%/ +7%/ -11%/ -13%/ -12%
2048/ 1/ +17%/ 0%/ +1%/ +35%/ +78%
2048/ 4/ +15%/ +14%/ -17%/ -24%/ -15%
2048/ 8/ +11%/ +9%/ -15%/ -20%/ -12%
4096/ 1/ +3%/ -7%/ 0%/ +21%/ +48%
4096/ 4/ +3%/ +4%/ -9%/ -19%/ +41%
4096/ 8/ +15%/ +13%/ -33%/ -28%/ -15%
16384/ 1/ +5%/ -8%/ -4%/ -10%/ +323%
16384/ 4/ +13%/ +5%/ -15%/ -11%/ +147%
16384/ 8/ +8%/ +6%/ -25%/ -27%/ -31%
65535/ 1/ +8%/ 0%/ +5%/ 0%/ +45%
65535/ 4/ +10%/ +1%/ +7%/ -8%/ +151%
65535/ 8/ +5%/ 0%/ +1%/ -16%/ -29%
Test B were done through:
- 50us as busy loop timeout
- Netperf 2.6
- Two machines with back to back connected ixgbe
- Two guests each wich 1 vcpu and 1 queue
- pin two vhost threads to the same cpu on host to simulate the cpu
contending
Results:
- In this radical case, we can still get at most 14% improvement on
TCP_RR.
- For guest tx stream, minor improvemnt with at most 5% regression in
one byte case. For guest rx stream, at most 5% regression were seen.
Guest TX:
size /-+% /
1 /-5.55%/
64 /+1.11%/
256 /+2.33%/
512 /-0.03%/
1024 /+1.14%/
4096 /+0.00%/
16384/+0.00%/
Guest RX:
size /-+% /
1 /-5.11%/
64 /-0.55%/
256 /-2.35%/
512 /-3.39%/
1024 /+6.8% /
4096 /-0.01%/
16384/+0.00%/
TCP_RR:
size /-+% /
1 /+9.79% /
64 /+4.51% /
256 /+6.47% /
512 /-3.37% /
1024 /+6.15% /
4096 /+14.88%/
16384/-2.23% /
Changes from V2:
- rename vhost_vq_more_avail() to vhost_vq_avail_empty(). And return
false we __get_user() fails.
- do not bother premmptions/timers for good path.
- use vhost_vring_state as ioctl parameter instead of reinveting a new
one.
- add the unit of timeout (us) to the comment of new added ioctls
Changes from V1:
- remove the buggy vq_error() in vhost_vq_more_avail().
- leave vhost_enable_notify() untouched.
Changes from RFC V3:
- small tweak on the code to avoid multiple duplicate conditions in
critical path when busy loop is not enabled.
- add the test result of multiple VMs
Changes from RFC V2:
- poll also at the end of rx handling
- factor out the polling logic and optimize the code a little bit
- add two ioctls to get and set the busy poll timeout
- test on ixgbe (which can give more stable and reproducable numbers)
instead of mlx4.
Changes from RFC V1:
- add a comment for vhost_has_work() to explain why it could be
lockless
- add param description for busyloop_timeout
- split out the busy polling logic into a new helper
- check and exit the loop when there's a pending signal
- disable preemption during busy looping to make sure lock_clock() was
correctly used.
Jason Wang (3):
vhost: introduce vhost_has_work()
vhost: introduce vhost_vq_avail_empty()
vhost_net: basic polling support
drivers/vhost/net.c | 79 +++++++++++++++++++++++++++++++++++++++++++---
drivers/vhost/vhost.c | 35 ++++++++++++++++++++
drivers/vhost/vhost.h | 3 ++
include/uapi/linux/vhost.h | 6 ++++
4 files changed, 118 insertions(+), 5 deletions(-)
--
2.5.0
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
