Hello,
On Wed, 05 Feb 2020 23:24:29 +0900,
Hajime Tazaki wrote:
> > Make sure you are testing with the vector network devices, the
> > legacy ones are scheduled to be obsoleted at some point
>
> I was aware of the commit to obsolete several backend with the vector
> device, but did not include in the patchset and tests. I will try to
> do it for the next round.
So I added a vector device support, tested with tap backend.
Here is a list of numbers with various configurations that v3+ patch
have.
disclaimer: the experiment is immature, not apple-to-apple in many
aspects. So this result only presents one of the parameter set that I
took. I will update/clean up later if there are interests.
- testbed
+--docker0--+
| | 10GbE
netperf +---tap0 eth0 +==========+ eth0 +---+ netserver
(client) (ixgbe) (ixgbe)
<-- Linux box (4.18.5) --> <-- Linux box (4.17.19) -->
- setup
varied client side (netperf) with different net drivers/devices.
tso,tx/rx csum are enabled if possible.
- netperf 10secs (Mbps) result
|TCP_STREAM | TCP_MAERTS
-------------------- --------------------------
UMMODE_LIB (um-tap) | 2290.42 | 1.04
UMMODE_LIB (vec-tap) | 3699.98 | 5682.40
UMMODE_LIB (virtio) | 8029.13 | 9384.78
UMMODE_KERN (um-tap) | 2233.17 | 7.85
UMMODE_KERN (vec-tap)| 5527.37 | 9414.00
# UMMODE_LIB (virtio) isn't included in v3 patches.
full output log is here;
https://gist.github.com/thehajime/a71878cccf7830a23a23f8f8e8cc8753
result of UMMODE_LIB (vec-tap) is not stable: it sometimes shows over
8Gbps (TCP_STREAM) while most of the times lower.
But I suppose UMMODE_LIB with vector driver isn't that bad, though
there is still a gap to UMMODE_KERN (vector).
-- Hajime
