Re: [Xen-devel] [PATCH 0/4] Implement persistent grant in xen-netfront/netback

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Thu, Nov 15, 2012 at 11:56:17AM +0100, Roger Pau Monné wrote:
> On 15/11/12 09:38, ANNIE LI wrote:
> > 
> > 
> > On 2012-11-15 15:40, Pasi Kärkkäinen wrote:
> >> Hello,
> >>
> >> On Thu, Nov 15, 2012 at 03:03:07PM +0800, Annie Li wrote:
> >>> This patch implements persistent grants for xen-netfront/netback. This
> >>> mechanism maintains page pools in netback/netfront, these page pools is used to
> >>> save grant pages which are mapped. This way improve performance which is wasted
> >>> when doing grant operations.
> >>>
> >>> Current netback/netfront does map/unmap grant operations frequently when
> >>> transmitting/receiving packets, and grant operations costs much cpu clock. In
> >>> this patch, netfront/netback maps grant pages when needed and then saves them
> >>> into a page pool for future use. All these pages will be unmapped when
> >>> removing/releasing the net device.
> >>>
> >> Do you have performance numbers available already? with/without persistent grants?
> > I have some simple netperf/netserver test result with/without persistent 
> > grants,
> > 
> > Following is result of with persistent grant patch,
> > 
> > Guests, Sum,      Avg,     Min,     Max
> >   1,  15106.4,  15106.4, 15106.36, 15106.36
> >   2,  13052.7,  6526.34,  6261.81,  6790.86
> >   3,  12675.1,  6337.53,  6220.24,  6454.83
> >   4,  13194,  6596.98,  6274.70,  6919.25
> > 
> > 
> > Following are result of without persistent patch
> > 
> > Guests, Sum,     Avg,    Min,        Max
> >   1,  10864.1,  10864.1, 10864.10, 10864.10
> >   2,  10898.5,  5449.24,  4862.08,  6036.40
> >   3,  10734.5,  5367.26,  5261.43,  5473.08
> >   4,  10924,    5461.99,  5314.84,  5609.14
> 
> In the block case, performance improvement is seen when using a large
> number of guests, could you perform the same benchmark increasing the
> number of guests to 15?

Keep in mind that one of the things that is limiting these numbers is
that netback is very CPU intensive. So I think it could get much much
faster - but netback pegs at 100%. With Wei Liu's patches the CPU usage
did drop by 40% (this is when I tested the old netback with Wei's
netback patches)- so we should see even a further speed increase.