[dpdk-dev] [PATCH v6 0/8] virtio ring layout optimization and simple rx/tx processing

[Xu, Qian Q]

Some virtio-pmd optimization performance data sharing: 
1. Use simplified vhost-sample, only doing the dequeuer and free, so virtio only tx, then test the virtio tx performance improvement. Then in the VM, using one virtio to do the txonly, and let the virtio tx working. Also modified the txonly file to remove the memory copy part, then check the virtio TX rate. The performance of optimized virtio-pmd will have ~2x performance than the non-optimized virtio-pmd. 
2. Similarly as item1, but use the default txonly file, so with memory copy, then the performance of optimized virtio-pmd will have ~37% performance improvement than the non-optimized virtio-pmd. 
3. In the OVS test scenario, one physical NIC + one virtio in the VM, then let the virtio do the loopback(having rx and tx), running testpmd in the VM, then the performance will have 60% performance improvement than the non-optimized virtio-pmd. 



Thanks
Qian

-----Original Message-----
From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Huawei Xie
Sent: Thursday, October 29, 2015 10:53 PM
To: dev at dpdk.org
Subject: [dpdk-dev] [PATCH v6 0/8] virtio ring layout optimization and simple rx/tx processing

Changes in v6:
- Update release notes
- Fix the error in virtio tx ring layout ascii chart in the cover-letter

Changes in v5:
- Call __rte_pktmbuf_prefree_seg to check refcnt when free mbufs

Changes in v4:
- Fix the error in virtio tx ring layout ascii chart in the commit message
- Move virtio_xmit_cleanup ahead to free descriptors earlier
- Test merge-able feature when select simple rx/tx functions

Changes in v3:
- Remove unnecessary NULL test for rte_free
- Remove unnecessary assign of local var after free
- Remove return at the end of void function
- Remove always_inline attribute for virtio_xmit_cleanup
- Reword some commit messages
- Add TODO in the commit message of simple tx patch

Changes in v2:
- Remove the configure macro
- Enable simple R/TX processing when user specifies simple txq flags
- Reword some comments and commit messages

In DPDK based switching enviroment, mostly vhost runs on a dedicated core while virtio processing in guest VMs runs on other different cores.
Take RX for example, with generic implementation, for each guest buffer,
a) virtio driver allocates a descriptor from free descriptor list
b) modify the entry of avail ring to point to allocated descriptor
c) after packet is received, free the descriptor

When vhost fetches the avail ring, it need to fetch the modified L1 cache from virtio core, which is a heavy cost in current CPU implementation.

This idea of this optimization is:
    allocate the fixed descriptor for each entry of avail ring, so avail ring will always be the same during the run.
This removes L1M cache transfer from virtio core to vhost core for avail ring.
(Note we couldn't avoid the cache transfer for descriptors).
Besides, descriptor allocation and free operation is eliminated.
This also makes vector procesing possible to further accelerate the processing.

This is the layout for the avail ring(take 256 ring entries for example), with each entry pointing to the descriptor with the same index.
                    avail
                    idx
                    +
                    |
+----+----+---+-------------+------+
| 0  | 1  | 2 | ... |  254  | 255  |  avail ring
+-+--+-+--+-+-+---------+---+--+---+
  |    |    |       |   |      |
  |    |    |       |   |      |
  v    v    v       |   v      v
+-+--+-+--+-+-+---------+---+--+---+
| 0  | 1  | 2 | ... |  254  | 255  |  desc ring
+----+----+---+-------------+------+
                    |
                    |
+----+----+---+-------------+------+
| 0  | 1  | 2 |     |  254  | 255  |  used ring
+----+----+---+-------------+------+
                    |
                    +

This is the ring layout for TX.
As we need one virtio header for each xmit packet, we have 128 slots available.

                         ++
                         ||
                         ||
+-----+-----+-----+--------------+------+------+------+
|  0  |  1  | ... |  127 || 128  | 129  | ...  | 255  |   avail ring
+--+--+--+--+-----+---+------+---+--+---+------+--+---+
   |     |            |  ||  |      |             |
   v     v            v  ||  v      v             v
+--+--+--+--+-----+---+------+---+--+---+------+--+---+
| 128 | 129 | ... |  255 || 127  | 128  | ...  | 255  |   desc ring for virtio_net_hdr
+--+--+--+--+-----+---+------+---+--+---+------+--+---+
   |     |            |  ||  |      |             |
   v     v            v  ||  v      v             v
+--+--+--+--+-----+---+------+---+--+---+------+--+---+
|  0  |  1  | ... |  127 ||  0   |  1   | ...  | 127  |   desc ring for tx dat
+-----+-----+-----+--------------+------+------+------+
                         ||
                         ||
                         ++


Performance boost could be observed only if the virtio backend isn't the bottleneck or in VM2VM case.
There are also several vhost optimization patches to be submitted later.


Huawei Xie (8):
  virtio: add virtio_rxtx.h header file
  virtio: add software rx ring, fake_buf into virtqueue
  virtio: rx/tx ring layout optimization
  virtio: fill RX avail ring with blank mbufs
  virtio: virtio vec rx
  virtio: simple tx routine
  virtio: pick simple rx/tx func
  doc: update release notes 2.2 about virtio performance optimization

 doc/guides/rel_notes/release_2_2.rst    |   3 +
 drivers/net/virtio/Makefile             |   2 +-
 drivers/net/virtio/virtio_ethdev.c      |  12 +-
 drivers/net/virtio/virtio_ethdev.h      |   5 +
 drivers/net/virtio/virtio_rxtx.c        |  56 ++++-
 drivers/net/virtio/virtio_rxtx.h        |  39 +++
 drivers/net/virtio/virtio_rxtx_simple.c | 414 ++++++++++++++++++++++++++++++++
 drivers/net/virtio/virtqueue.h          |   5 +
 8 files changed, 532 insertions(+), 4 deletions(-)  create mode 100644 drivers/net/virtio/virtio_rxtx.h  create mode 100644 drivers/net/virtio/virtio_rxtx_simple.c

--
1.8.1.4