Netmap with VMWare

[Pradosh Datta]

Hi,
I am trying to use the netmap on vmware in CentOS guests, I am able to use
netmap with the e1000 driver on vmware (though the throughput is close to
line speed).

But is there any way to use the netmap with vmxnet3? Is there any patch
available to support that? Or any suggestion on how can I achieve
high-packet-rate between vm-to-vm using netmap?

Best Regards,
Pradosh

[Luigi Rizzo]

Post by Pradosh Datta
Hi,
I am trying to use the netmap on vmware in CentOS guests, I am able to use
netmap with the e1000 driver on vmware (though the throughput is close to
line speed).
But is there any way to use the netmap with vmxnet3? Is there any patch
available to support that? Or any suggestion on how can I achieve
high-packet-rate between vm-to-vm using netmap?

The vm-to-vm bottleneck is typically in the hypervisor (vmware) and the
switch (in this case it may be the linux bridge in centos).

I am afraid just using netmap in the guest won't help you much in this respect.
See our ANCS'13 paper
http://info.iet.unipi.it/~luigi/papers/20130903-rizzo-ancs.pdf
for what you can expect (look at the unmodified hypervisor/linux bridge case).

This said, you can always use netmap (in emulated mode) on any network interface
in the guest.

cheers
luigi

Post by Pradosh Datta
Best Regards,
Pradosh
_______________________________________________
https://lists.freebsd.org/mailman/listinfo/freebsd-net

[Pradosh Datta]

Hi Luigi,

Thank you for your response and the reference to the paper. After receiving
your comment, for the time being, I shifted my focus to PCI-Passthrough
because without modifying the host, it will not be possible to achieve line
rate between guest-to-guest.

With PCI-Passthrough enabled in ESXi, I am able to test with ixgbe (device:
82599) with netmap. But here again I am reaching maximum of ~5.5Gbps (with
64byte pkt, single CPU 2666.761MHz), the pkt-gen is using ~73%. I looked
into another paper related to ptnetmap, but that again needs hypervisor
change (and I do not want to go into that).

I played with different values with 'ethtool --coalesce rx-usecs' - nothing
is helping. The flow-control is also disabled.

Are there any specific CONFIG(s) I need to enable/disable while compiling
the linux-kernel (3.10 x86_64), or compiling the ixgbe driver (I am using
3.13.10)?
Also, the sections 3.X in the paper (Speeding Up Packet I/O in Virtual
Machines - Our Works) describes many ways of improvement - are those need
to be taken care in case of PCI-Passthrough as well?
If you are already aware of any experiment, is it possible to achieve line
rate using PCI-Passthrough with netmap?

Again thanks for your time and help.

Best Regards,
Pradosh

(NB: one typo correction in my last mail, with e1000 ...throughput is _not_
close to line rate)

Post by Luigi Rizzo

Post by Pradosh Datta
Hi,
I am trying to use the netmap on vmware in CentOS guests, I am able to

use

Post by Pradosh Datta
netmap with the e1000 driver on vmware (though the throughput is close to
line speed).
But is there any way to use the netmap with vmxnet3? Is there any patch
available to support that? Or any suggestion on how can I achieve
high-packet-rate between vm-to-vm using netmap?

The vm-to-vm bottleneck is typically in the hypervisor (vmware) and the
switch (in this case it may be the linux bridge in centos).
I am afraid just using netmap in the guest won't help you much in this respect.
See our ANCS'13 paper
http://info.iet.unipi.it/~luigi/papers/20130903-rizzo-ancs.pdf
for what you can expect (look at the unmodified hypervisor/linux bridge case).
This said, you can always use netmap (in emulated mode) on any network interface
in the guest.
cheers
luigi

Post by Pradosh Datta
Best Regards,
Pradosh
_______________________________________________
https://lists.freebsd.org/mailman/listinfo/freebsd-net

--
-----------------------------------------+-------------------------------
http://www.iet.unipi.it/~luigi/ . Universita` di Pisa
TEL [phone] . via Diotisalvi 2
Mobile [phone] . 56122 PISA (Italy)
-----------------------------------------+-------------------------------