Re: [RFC PATCHv1 net-next] xen-netback: always fully coalesce guest Rx packets

From: Ian Campbell <Ian.Campbell@citrix.com>
To: David Vrabel <david.vrabel@citrix.com>
Cc: netdev@vger.kernel.org,
	Jonathan Davies <jonathan.davies@citrix.com>,
	Wei Liu <wei.liu2@citrix.com>,
	xen-devel@lists.xenproject.org
Subject: Re: [RFC PATCHv1 net-next] xen-netback: always fully coalesce guest Rx packets
Date: Tue, 20 Jan 2015 11:21:48 +0000	[thread overview]
Message-ID: <1421752908.10440.224.camel__12114.3642966219$1421753002$gmane$org@citrix.com> (raw)
In-Reply-To: <54BD408A.6080405@citrix.com>

On Mon, 2015-01-19 at 17:36 +0000, David Vrabel wrote:
> On 13/01/15 14:30, Wei Liu wrote:
> > On Tue, Jan 13, 2015 at 02:05:17PM +0000, David Vrabel wrote:
> >> Always fully coalesce guest Rx packets into the minimum number of ring
> >> slots.  Reducing the number of slots per packet has significant
> >> performance benefits (e.g., 7.2 Gbit/s to 11 Gbit/s in an off-host
> >> receive test).
> >>
> > 
> > Good number.
> > 
> >> However, this does increase the number of grant ops per packet which
> >> decreases performance with some workloads (intrahost VM to VM)
> > 
> > Do you have figures before and after this change?
> 
> Some better (more rigorous) results done by Jonathan Davies shows no
> regressions with full coalescing even without the grant copy
> optimization, and a big improvement to single stream receive.
> 
>                          baseline    Full coalesce
> Interhost aggregate      24 Gb/s     24 Gb/s
> Interhost VM receive      7.2 Gb/s   11 Gb/s
> Intrahost single stream  14 Gb/s     14 Gb/s
> Intrahost aggregate      34 Gb/s     34 Gb/s
> 
> We do not measure the performance of dom0 to guest traffic but my ad-hoc
> measurements suggest this may be 5-10% slower.  I don't think this is a
> very important use case though.

If you are updating your dom0 kernel to take advantage of this
improvement and you care about dom0->domU performance too then also
updating your Xen at the same is not a huge deal, I think. Or at least I
don't consider it a blocker for making progress (certainly not progress
of the order of 50% improvements!).

> So...
> 
> >> /unless/ grant copy has been optimized for adjacent ops with the same
> >> source or destination (see "grant-table: defer releasing pages
> >> acquired in a grant copy"[1]).
> >>
> >> Do we need to retain the existing path and make the always coalesce
> >> path conditional on a suitable version of Xen?
> 
> ...I think the answer to this is no.

Agreed.

> >> ---
> >>  drivers/net/xen-netback/common.h  |    1 -
> >>  drivers/net/xen-netback/netback.c |  106 ++-----------------------------------
> >>  2 files changed, 3 insertions(+), 104 deletions(-)
> > 
> > Love the diffstat!
> 
> Yes, it's always nice when you delete code and it goes faster... :)

Full-Ack to that ;-)

Ian.