Re: [Xen-users] xen network para-virtualization

[Ian Campbell <ian.campbell@xxxxxxxxxx>]

On Sun, 2014-03-09 at 20:48 +0800, åä wrote:
> Hi all, 
> 
> 
> I have a question about xen network para-virtualization. From the web
> information, it seems that xen supports copy and page flipping for
> packet transmission between netback and netfront. Does anybody know
> which method is efficient?  Why? At first, I think page flipping is
> more efficient than copy. However, when I read the linux 3.7 code, I
> found that it only supports copy. If page flipping is faster than
> copy, why the newer kernel version  does not support it. So I think my
> understanding has some problem. Does anybody give me some
> instructions? Thank you in advance!

There are actually three modes here, copy (both directions), flipping
(guest rx) and mapping (guest tx).

IIRC very old kernels (pre 2.6.18) supported page flipping for traffic
going back->front but IIRC the TLB shootdown overheads we re so high
that copy was found to be more efficient. (NB: this was before my time
so I am not 100% sure of the specifics).

The 2.6.18 ports also supported mapping for traffic going from
back->front. However this required infrastructure which was very
difficult to upstream and so a version which used copying was upstreamed
into the mainline kernel instead. Zoltan Kiss has been working recently
on reintroducing mapping mode to the upstream netback (see xen-devel@
archives).

At one point there was some speculation that copying might be faster on
newer processors which are more optimised for such cases. That doesn't
seem to be the case in practice though, at least not according to the
numbers Zoltan has been reporting.

Ian.



_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxx
http://lists.xen.org/xen-users