Hi Yiwen,
I'm currently interested on how to improve the VSOCK performance and I
read your discussions with Jason, Michael, and Stefan on both
linux-netdev and qemu-devel mailing lists.
Are you still working on it?
Reading the discussions I understood that batching can help us a lot
to increase the VSOCK performance (especially for guest->host
direction).
As a first step, I added VSOCK support to iperf3 [1] in order to
measure the current performance of VSOCK (without any modification):
Single socket
Guest->Host ~950MB/s
Host->Guest ~2700MB/s
They are slightly different from your results but I think that depends
on the hardware.
My current setup is the following:
HW: Lenovo T480s (i7-8650U, DDR4 2133 Mhz)
OS (guest/host): Fedora 29 (Linux 4.20.4-200.fc29.x86_64)
Guest app: iperf3 --vsock -s
Host app: iperf3 --vsock -c $GUEST_CID -l 64K
Just to double check these values, can you share with me the tool that you used?
Thanks,
Stefano
#1 https://github.com/esnet/iperf/pull/840
--
Stefano Garzarella
Software Engineer @ Red Hat
