Message-ID: <1431451117-70051-1-git-send-email-joao.martins@neclab.eu>
Date:	Tue, 12 May 2015 19:18:24 +0200
From:	Joao Martins <joao.martins@...lab.eu>
To:	<xen-devel@...ts.xenproject.org>, <netdev@...r.kernel.org>
CC:	<wei.liu2@...rix.com>, <ian.campbell@...rix.com>,
	<david.vrabel@...rix.com>, <boris.ostrovsky@...cle.com>,
	<konrad.wilk@...cle.com>, Joao Martins <joao.martins@...lab.eu>
Subject: [RFC PATCH 00/13] Persistent grant maps for xen net drivers

This patch implements persistent grants for xen-net{back,front}. There has
been work on persistent grants in the past[1], but the one here
described is a bit different: 1) using zerocopy skbs for RX path in
xen-netfront as opposed to memcpy every packet; 2) using a tree to store
the grants (and doing so using same interfaces as blkback/blkfront, to
hopefully share a common layer) 3) reusing TX map/unmap paths for handling
persistent grants case 4) sending the buffer on ndo_start_xmit as opposed
to bouncing them through the RX kthread.

Intrahost performance increases significantly between 16-78% (depending on
the host), and performance per-queue (tested with up to 6 queues) scales
nicely all the way to ~30 Gbits, specially on TX path. Rates for small
packet sizes increase up to 1.82 Mpps TX (1.2 Gbit/s on wire, 64 pkt_size)
and 2.78 Mpps RX (1.8 Gbit/s on wire, 64 pkt_size). This is around 1.5x to
2x improvement compared with grant copy/map. On bigger packet sizes the
improvement is even more noticeable. The only problem though, is that
performance seems to decrease for RX path on 1 queue (and 2 queues with
slower CPUs). This is because of the extra memcpy on xen-netfront (in
skb_copy_ubufs, frags > 0) for pkt len > RX_COPY_THRESHOLD.

This series are organized as follows: Patch 1 defines the routines for
managing the tree, Patch 2,11 implements feature detection, Patch 3-4 the
actual implementation of TX/RX grants on the backend, Patch 5 proposes
copying the buffer on ndo_start_xmit, Patch 7 fixes a bug when using pktgen
with burst >1 without persistent grants. Patches 12-13 implements the
frontend part. Patches 5,9,10 are only structural before introducing the
main changes. Overall, it works as the following:

On transmit path, xen-netfront will grant a new page and memcpy the skb to
it.  On netback NAPI poll, we will check for a persistent grant available
for header and frags grefs. If none is found in the tree, it will resort to
grant copy (only for the header) but still preparing the grant map and
adding them to the tree.  The frags are handled the same as before with the
exception of having to add the grant map to the tree and not having to
unmap it in the skb callback (when freeing).

On receive path we lookup for a page mapped for guest gref. If it exists in
the tree, it does a memcpy and reverting to grant copy in case of the first
failed lookup, or in case we don't have free pages to create mappings. On
xen-netfront RX we then grab responses and use zerocopy skbs to be able to
reuse the same pages for future requests, and to avoid an extra copy for
packets smaller <= RX_COPY_THRESHOLD.  Additionally I also propose copying
the buffer on ndo_start_xmit to avoid wait queue and rx_queue contention
and because no grant table batching is required with persistent grants
(besides the initial map/copy). The latter improved the pkt rates
(by ~2x) specially for smaller packet sizes, and for pkt
len <= RX_COPY_THRESHOLD in RX path.

Packet I/O Tests:

Measured on a Intel Xeon E5-1650 v2, Xen 4.5, no HT. Used pktgen "burst 1"
and "clone_skbs 100000" (to avoid alloc skb overheads) with various pkt
sizes. All tests are DomU <-> Dom0, unless specified otherwise.
Graphs:
http://cnp.neclab.eu/images/pgntperf/udprx_hostA.png
http://cnp.neclab.eu/images/pgntperf/udptx_hostA.png

                             | Baseline  | Pers. Grants       |
---------------------------------------------------------------
1q DomU TX (pkt_size 64)     | 1.24 Mpps | 1.82  Mpps (+ 46%) |
1q DomU TX (pkt_size 1496)   | 518  Kpps | 1.48  Mpps (+150%) |
1q DomU TX (pkt_size 65535)  | 66.4 Kpps | 205.1 Kpps         |
---------------------------------------------------------------
1q DomU RX (pkt_size 64)     | 1.33 Mpps | 2.78  Mpps (+109%) |
1q DomU RX (pkt_size 1496)   | 1.03 Mpps | 1.66  Mpps (+ 60%) |
1q DomU RX (pkt_size 65535)  | 52.5 Kpps | 97.8  Kpps         |
---------------------------------------------------------------

I also made a micro-benchmark with a MiniOS-based guest and it was able
to reach up to 4.17 Mpps (with pktgen burst 8, pkt_size 64) hinting that
the throughput grows with bigger batches when using xmit_more. In
this case, the guest netfront was busy looping, and not setting the ring
rsp_event which would (only) lead to not triggering the notification by
the backend. Note that my purpose with this experiment was just to see if
copying the buffer on xenvif_start_xmit was indeed performing better.

Bulk Transfer Tests A:

Measured on a Intel Xeon E5-1650 v2 @ 3.5 Ghz, Xen 4.5, no HT. Used
iperf (TCP)  with an increased number of flows following similar
methodology explained in [2]. All vif irqs are balanced across cores
in both Dom0 and DomU. Tests are between DomU <-> Dom0, unless
specified otherwise.
Graph:
http://cnp.neclab.eu/images/pgntperf/tcprx_hostA.png
http://cnp.neclab.eu/images/pgntperf/tcptx_hostA.png
http://cnp.neclab.eu/images/pgntperf/tcpintra_hostA.png

                | Baseline  | Pers. Grants      |
-------------------------------------------------
1q DomU TX      | 14.5 Gbit | 21.6 Gbit (+ 48%) |
2q DomU TX      | 17.6 Gbit | 27.4 Gbit         |
3q DomU TX      | 17.2 Gbit | 29.3 Gbit (+ 70%) |
-------------------------------------------------
1q DomU RX      | 20.9 Gbit | 17.8 Gbit (- 15%) |
2q DomU RX      | 21.1 Gbit | 24.9 Gbit         |
3q DomU RX      | 22.1 Gbit | 31.0 Gbit (+ 40%) |
-------------------------------------------------
1q DomU-to-DomU | 12.4 Gbit | 18.9 Gbit (+ 52%) |

Bulk Transfer Tests B: 

Same as before, but measured on a Intel Xeon E5-2697 v2 @ 2.7 Ghz,
to test guests with a higher number of queues (>3).
Graph:
http://cnp.neclab.eu/images/pgntperf/tcprx_hostB.png
http://cnp.neclab.eu/images/pgntperf/tcptx_hostB.png
http://cnp.neclab.eu/images/pgntperf/tcpintra_hostB.png

                | Baseline  | Pers. Grants      |
-------------------------------------------------
1q DomU TX      | 10.5 Gbit | 15.9 Gbit (+ 51%) |
2q DomU TX      | 14.0 Gbit | 20.1 Gbit         |
3q DomU TX      | 15.7 Gbit | 23.5 Gbit         |
4q DomU TX      | 15.0 Gbit | 25.9 Gbit         |
6q DomU TX      | 15.9 Gbit | 30.0 Gbit (+ 88%) |
-------------------------------------------------
1q DomU RX      | 15.1 Gbit | 13.3 Gbit (- 11%) |
2q DomU RX      | 19.5 Gbit | 18.1 Gbit (-  7%) |
3q DomU RX      | 22.2 Gbit | 22.7 Gbit         |
4q DomU RX      | 23.7 Gbit | 25.8 Gbit         |
6q DomU RX      | 24.0 Gbit | 29.8 Gbit (+ 24%) |
-------------------------------------------------
1q DomU-to-DomU | 12.5 Gbit | 14.5 Gbit (+ 16%) |
2q DomU-to-DomU | 12.6 Gbit | 20.6 Gbit (+ 63%) |
3q DomU-to-DomU | 13.7 Gbit | 24.5 Gbit (+ 78%) |

There have been recently[3] some discussions and issues raised on
persistent grants for the block layer, though the numbers above
show some significant improvements specially on more network intensive
workloads and provide a margin for comparison against future map/unmap
improvements.

Any comments or suggestions are welcome,
Thanks!
Joao

[1] http://article.gmane.org/gmane.linux.network/249383
[2] http://bit.ly/1IhJfXD
[3] http://lists.xen.org/archives/html/xen-devel/2015-02/msg02292.html

Joao Martins (13):
  xen-netback: add persistent grant tree ops
  xen-netback: xenbus feature persistent support
  xen-netback: implement TX persistent grants
  xen-netback: implement RX persistent grants
  xen-netback: refactor xenvif_rx_action
  xen-netback: copy buffer on xenvif_start_xmit()
  xen-netback: add persistent tree counters to debugfs
  xen-netback: clone skb if skb->xmit_more is set
  xen-netfront: move grant_{ref,page} to struct grant
  xen-netfront: refactor claim/release grant
  xen-netfront: feature-persistent xenbus support
  xen-netfront: implement TX persistent grants
  xen-netfront: implement RX persistent grants

 drivers/net/xen-netback/common.h    |  79 ++++
 drivers/net/xen-netback/interface.c |  78 +++-
 drivers/net/xen-netback/netback.c   | 873 ++++++++++++++++++++++++++++++------
 drivers/net/xen-netback/xenbus.c    |  24 +
 drivers/net/xen-netfront.c          | 362 ++++++++++++---
 5 files changed, 1216 insertions(+), 200 deletions(-)

-- 
2.1.3

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists