Re: [Qemu-devel] [PATCH] net: add raw backend - some performance measurements

[Or Gerlitz <ogerlitz@Voltaire.com>]

Jamie Lokier wrote:
> Or Gerlitz wrote:

>> the performance (packets per second and cpu utilization) one can get 
>> with bridge+tap is much lower vs what you get with the raw mode approach.

> Have you measured it?

yes, here's some data: using 2.6.29.1 in the guest, 2.6.30 in the host, with 1Gbe 
connectivity (Intel 82575EB) between the two nodes,  I see the following results: 

with -net raw (packet socket)
		pps	cs	us	sys
vm->phys	240k	200	7	8	
phys->vm	160k	100	5	7

with -net tap (tap + bridge)
		pps	cs	us	sys
vm->phys	170k	600	5	10
phys->vm	150k	14k	5	20

where "pps" stands for packets-per-second, "cs", "us" and "sys" are taken from vmstat output, such that they represent the context switches per second, user and system time percents.  The benchmark I use is netperf 2.4.4 / UDP_STREAM with 22 bytes payload length such that there are 64(=14+20+8+22) bytes on the wire. On this setup (udp, 64 byte frames), doing phys->phys test, netperf sends/receives 450K pps and pktgen sends 900K pps, all tests done without any interrupt moderation tuning.

You can see that the raw mode has much better packets per second for the VM TX flow, and on the VM RX side, a bit better pps rate but much lower cpu utilization and context switches number.

Or.

All this on top of mainstream qemu whose head is commit 8676188b751ca28ab7c42baf20ea64391625b44d
"Work around Solaris gas problem"