On 6/29/11 1:42 AM, "Michael S. Tsirkin" [off-list ref] wrote:

On Tue, Jun 28, 2011 at 11:08:07AM -0500, Tom Lendacky wrote:

quoted

On Sunday, June 19, 2011 05:27:00 AM Michael S. Tsirkin wrote:

quoted

OK, different people seem to test different trees.  In the hope to get
everyone on the same page, I created several variants of this patch so
they can be compared. Whoever's interested, please check out the
following, and tell me how these compare:

kernel:

git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost.git

virtio-net-limit-xmit-polling/base - this is net-next baseline to test
against virtio-net-limit-xmit-polling/v0 - fixes checks on out of capacity
virtio-net-limit-xmit-polling/v1 - previous revision of the patch
            this does xmit,free,xmit,2*free,free
virtio-net-limit-xmit-polling/v2 - new revision of the patch
            this does free,xmit,2*free,free

Here's a summary of the results.  I've also attached an ODS format
spreadsheet
(30 KB in size) that might be easier to analyze and also has some pinned VM
results data.  I broke the tests down into a local guest-to-guest scenario
and a remote host-to-guest scenario.

Within the local guest-to-guest scenario I ran:
  - TCP_RR tests using two different messsage sizes and four different
    instance counts among 1 pair of VMs and 2 pairs of VMs.
  - TCP_STREAM tests using four different message sizes and two different
    instance counts among 1 pair of VMs and 2 pairs of VMs.

Within the remote host-to-guest scenario I ran:
  - TCP_RR tests using two different messsage sizes and four different
    instance counts to 1 VM and 4 VMs.
  - TCP_STREAM and TCP_MAERTS tests using four different message sizes and
    two different instance counts to 1 VM and 4 VMs.
over a 10GbE link.

roprabhu, Tom,

Thanks very much for the testing. So on the first glance
one seems to see a significant performance gain in V0 here,
and a slightly less significant in V2, with V1
being worse than base. But I'm afraid that's not the
whole story, and we'll need to work some more to
know what really goes on, please see below.


Some comments on the results: I found out that V0 because of mistake
on my part was actually almost identical to base.
I pushed out virtio-net-limit-xmit-polling/v1a instead that
actually does what I intended to check. However,
the fact we get such a huge distribution in the results by Tom
most likely means that the noise factor is very large.


From my experience one way to get stable results is to
divide the throughput by the host CPU utilization
(measured by something like mpstat).
Sometimes throughput doesn't increase (e.g. guest-host)
by CPU utilization does decrease. So it's interesting.


Another issue is that we are trying to improve the latency
of a busy queue here. However STREAM/MAERTS tests ignore the latency
(more or less) while TCP_RR by default runs a single packet per queue.
Without arguing about whether these are practically interesting
workloads, these results are thus unlikely to be significantly affected
by the optimization in question.

What we are interested in, thus, is either TCP_RR with a -b flag
(configure with  --enable-burst) or multiple concurrent
TCP_RRs.

Michael, below are some numbers I got from one round of runs.
Thanks,
Roopa

256byte req/response.
Vcpus and irqs were pinned to 4 cores and the cpu utilization is
Avg across 4 cores.

base:
Numof concurrent TCP_RRs    Num of transactions/sec  host cpu-util(%)
1                            7982.93                        15.72
25                           67873                         28.84
50                           112534                        52.25
100                          192057                       86.54


v1
Numof concurrent TCP_RRs    Num of transactions/sec    host cpu-util(%)
1                           7970.94                       10.8
25                          65496.8                       28
50                          109858                        53.22
100                         190155                        87.5


v1a
Numof concurrent TCP_RRs    Num of transactions/sec   host cpu-util (%)
1                           7979.81                       9.5
25                          66786.1                       28
50                          109552                        51
100                         190876                        88


v2
Numof concurrent TCP_RRs    Num of transactions/sec   host cpu-util (%)
1                            7969.87                     16.5
25                           67780.1                     28.44
50                           114966                      54.29
100                          177982                      79.9