Re: [PATCH RFC v8 02/11] vhost: use batched get_vq_desc version

[Jason Wang <jasowang@redhat.com>]

On 2020/7/10 上午1:37, Michael S. Tsirkin wrote:
> On Thu, Jul 09, 2020 at 06:46:13PM +0200, Eugenio Perez Martin wrote:
>> On Wed, Jul 1, 2020 at 4:10 PM Jason Wang <jasowang@redhat.com> wrote:
>>>
>>> On 2020/7/1 下午9:04, Eugenio Perez Martin wrote:
>>>> On Wed, Jul 1, 2020 at 2:40 PM Jason Wang <jasowang@redhat.com> wrote:
>>>>> On 2020/7/1 下午6:43, Eugenio Perez Martin wrote:
>>>>>> On Tue, Jun 23, 2020 at 6:15 PM Eugenio Perez Martin
>>>>>> <eperezma@redhat.com> wrote:
>>>>>>> On Mon, Jun 22, 2020 at 6:29 PM Michael S. Tsirkin <mst@redhat.com> wrote:
>>>>>>>> On Mon, Jun 22, 2020 at 06:11:21PM +0200, Eugenio Perez Martin wrote:
>>>>>>>>> On Mon, Jun 22, 2020 at 5:55 PM Michael S. Tsirkin <mst@redhat.com> wrote:
>>>>>>>>>> On Fri, Jun 19, 2020 at 08:07:57PM +0200, Eugenio Perez Martin wrote:
>>>>>>>>>>> On Mon, Jun 15, 2020 at 2:28 PM Eugenio Perez Martin
>>>>>>>>>>> <eperezma@redhat.com> wrote:
>>>>>>>>>>>> On Thu, Jun 11, 2020 at 5:22 PM Konrad Rzeszutek Wilk
>>>>>>>>>>>> <konrad.wilk@oracle.com> wrote:
>>>>>>>>>>>>> On Thu, Jun 11, 2020 at 07:34:19AM -0400, Michael S. Tsirkin wrote:
>>>>>>>>>>>>>> As testing shows no performance change, switch to that now.
>>>>>>>>>>>>> What kind of testing? 100GiB? Low latency?
>>>>>>>>>>>>>
>>>>>>>>>>>> Hi Konrad.
>>>>>>>>>>>>
>>>>>>>>>>>> I tested this version of the patch:
>>>>>>>>>>>> https://lkml.org/lkml/2019/10/13/42
>>>>>>>>>>>>
>>>>>>>>>>>> It was tested for throughput with DPDK's testpmd (as described in
>>>>>>>>>>>> http://doc.dpdk.org/guides/howto/virtio_user_as_exceptional_path.html)
>>>>>>>>>>>> and kernel pktgen. No latency tests were performed by me. Maybe it is
>>>>>>>>>>>> interesting to perform a latency test or just a different set of tests
>>>>>>>>>>>> over a recent version.
>>>>>>>>>>>>
>>>>>>>>>>>> Thanks!
>>>>>>>>>>> I have repeated the tests with v9, and results are a little bit different:
>>>>>>>>>>> * If I test opening it with testpmd, I see no change between versions
>>>>>>>>>> OK that is testpmd on guest, right? And vhost-net on the host?
>>>>>>>>>>
>>>>>>>>> Hi Michael.
>>>>>>>>>
>>>>>>>>> No, sorry, as described in
>>>>>>>>> http://doc.dpdk.org/guides/howto/virtio_user_as_exceptional_path.html.
>>>>>>>>> But I could add to test it in the guest too.
>>>>>>>>>
>>>>>>>>> These kinds of raw packets "bursts" do not show performance
>>>>>>>>> differences, but I could test deeper if you think it would be worth
>>>>>>>>> it.
>>>>>>>> Oh ok, so this is without guest, with virtio-user.
>>>>>>>> It might be worth checking dpdk within guest too just
>>>>>>>> as another data point.
>>>>>>>>
>>>>>>> Ok, I will do it!
>>>>>>>
>>>>>>>>>>> * If I forward packets between two vhost-net interfaces in the guest
>>>>>>>>>>> using a linux bridge in the host:
>>>>>>>>>> And here I guess you mean virtio-net in the guest kernel?
>>>>>>>>> Yes, sorry: Two virtio-net interfaces connected with a linux bridge in
>>>>>>>>> the host. More precisely:
>>>>>>>>> * Adding one of the interfaces to another namespace, assigning it an
>>>>>>>>> IP, and starting netserver there.
>>>>>>>>> * Assign another IP in the range manually to the other virtual net
>>>>>>>>> interface, and start the desired test there.
>>>>>>>>>
>>>>>>>>> If you think it would be better to perform then differently please let me know.
>>>>>>>> Not sure why you bother with namespaces since you said you are
>>>>>>>> using L2 bridging. I guess it's unimportant.
>>>>>>>>
>>>>>>> Sorry, I think I should have provided more context about that.
>>>>>>>
>>>>>>> The only reason to use namespaces is to force the traffic of these
>>>>>>> netperf tests to go through the external bridge. To test netperf
>>>>>>> different possibilities than the testpmd (or pktgen or others "blast
>>>>>>> of frames unconditionally" tests).
>>>>>>>
>>>>>>> This way, I make sure that is the same version of everything in the
>>>>>>> guest, and is a little bit easier to manage cpu affinity, start and
>>>>>>> stop testing...
>>>>>>>
>>>>>>> I could use a different VM for sending and receiving, but I find this
>>>>>>> way a faster one and it should not introduce a lot of noise. I can
>>>>>>> test with two VM if you think that this use of network namespace
>>>>>>> introduces too much noise.
>>>>>>>
>>>>>>> Thanks!
>>>>>>>
>>>>>>>>>>>      - netperf UDP_STREAM shows a performance increase of 1.8, almost
>>>>>>>>>>> doubling performance. This gets lower as frame size increase.
>>>>>> Regarding UDP_STREAM:
>>>>>> * with event_idx=on: The performance difference is reduced a lot if
>>>>>> applied affinity properly (manually assigning CPU on host/guest and
>>>>>> setting IRQs on guest), making them perform equally with and without
>>>>>> the patch again. Maybe the batching makes the scheduler perform
>>>>>> better.
>>>>> Note that for UDP_STREAM, the result is pretty trick to be analyzed. E.g
>>>>> setting a sndbuf for TAP may help for the performance (reduce the drop).
>>>>>
>>>> Ok, will add that to the test. Thanks!
>>>
>>> Actually, it's better to skip the UDP_STREAM test since:
>>>
>>> - My understanding is very few application is using raw UDP stream
>>> - It's hard to analyze (usually you need to count the drop ratio etc)
>>>
>>>
>>>>>>>>>>>      - rests of the test goes noticeably worse: UDP_RR goes from ~6347
>>>>>>>>>>> transactions/sec to 5830
>>>>>> * Regarding UDP_RR, TCP_STREAM, and TCP_RR, proper CPU pinning makes
>>>>>> them perform similarly again, only a very small performance drop
>>>>>> observed. It could be just noise.
>>>>>> ** All of them perform better than vanilla if event_idx=off, not sure
>>>>>> why. I can try to repeat them if you suspect that can be a test
>>>>>> failure.
>>>>>>
>>>>>> * With testpmd and event_idx=off, if I send from the VM to host, I see
>>>>>> a performance increment especially in small packets. The buf api also
>>>>>> increases performance compared with only batching: Sending the minimum
>>>>>> packet size in testpmd makes pps go from 356kpps to 473 kpps.
>>>>> What's your setup for this. The number looks rather low. I'd expected
>>>>> 1-2 Mpps at least.
>>>>>
>>>> Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz, 2 NUMA nodes of 16G memory
>>>> each, and no device assigned to the NUMA node I'm testing in. Too low
>>>> for testpmd AF_PACKET driver too?
>>>
>>> I don't test AF_PACKET, I guess it should use the V3 which mmap based
>>> zerocopy interface.
>>>
>>> And it might worth to check the cpu utilization of vhost thread. It's
>>> required to stress it as 100% otherwise there could be a bottleneck
>>> somewhere.
>>>
>>>
>>>>>> Sending
>>>>>> 1024 length UDP-PDU makes it go from 570kpps to 64 kpps.
>>>>>>
>>>>>> Something strange I observe in these tests: I get more pps the bigger
>>>>>> the transmitted buffer size is. Not sure why.
>>>>>>
>>>>>> ** Sending from the host to the VM does not make a big change with the
>>>>>> patches in small packets scenario (minimum, 64 bytes, about 645
>>>>>> without the patch, ~625 with batch and batch+buf api). If the packets
>>>>>> are bigger, I can see a performance increase: with 256 bits,
>>>>> I think you meant bytes?
>>>>>
>>>> Yes, sorry.
>>>>
>>>>>>     it goes
>>>>>> from 590kpps to about 600kpps, and in case of 1500 bytes payload it
>>>>>> gets from 348kpps to 528kpps, so it is clearly an improvement.
>>>>>>
>>>>>> * with testpmd and event_idx=on, batching+buf api perform similarly in
>>>>>> both directions.
>>>>>>
>>>>>> All of testpmd tests were performed with no linux bridge, just a
>>>>>> host's tap interface (<interface type='ethernet'> in xml),
>>>>> What DPDK driver did you use in the test (AF_PACKET?).
>>>>>
>>>> Yes, both testpmd are using AF_PACKET driver.
>>>
>>> I see, using AF_PACKET means extra layers of issues need to be analyzed
>>> which is probably not good.
>>>
>>>
>>>>>> with a
>>>>>> testpmd txonly and another in rxonly forward mode, and using the
>>>>>> receiving side packets/bytes data. Guest's rps, xps and interrupts,
>>>>>> and host's vhost threads affinity were also tuned in each test to
>>>>>> schedule both testpmd and vhost in different processors.
>>>>> My feeling is that if we start from simple setup, it would be more
>>>>> easier as a start. E.g start without an VM.
>>>>>
>>>>> 1) TX: testpmd(txonly) -> virtio-user -> vhost_net -> XDP_DROP on TAP
>>>>> 2) RX: pkgetn -> TAP -> vhost_net -> testpmd(rxonly)
>>>>>
>>>> Got it. Is there a reason to prefer pktgen over testpmd?
>>>
>>> I think the reason is using testpmd you must use a userspace kernel
>>> interface (AF_PACKET), and it could not be as fast as pktgen since:
>>>
>>> - it talks directly to xmit of TAP
>>> - skb can be cloned
>>>
>> Hi!
>>
>> Here it is the result of the tests. Details on [1].
>>
>> Tx:
>> ===
>>
>> For tx packets it seems that the batching patch makes things a little
>> bit worse, but the buf_api outperforms baseline by a 7%:
>>
>> * We start with a baseline of 4208772.571 pps and 269361444.6 bytes/s [2].
>> * When we add the batching, I see a small performance decrease:
>> 4133292.308 and 264530707.7 bytes/s.
>> * However, the buf api it outperform the baseline: 4551319.631pps,
>> 291205178.1 bytes/s
>>
>> I don't have numbers on the receiver side since it is just a XDP_DROP.
>> I think it would be interesting to see them.
>>
>> Rx:
>> ===
>>
>> Regarding Rx, the reverse is observed: a small performance increase is
>> observed with batching (~2%), but buf_api makes tests perform equally
>> to baseline.
>>
>> pktgen was called using pktgen_sample01_simple.sh, with the environment:
>> DEV="$tap_name" F_THREAD=1 DST_MAC=$MAC_ADDR COUNT=$((2500000*25))
>> SKB_CLONE=$((2**31))
>>
>> And testpmd is the same as Tx but with forward-mode=rxonly.
>>
>> Pktgen reports:
>> Baseline: 1853025pps 622Mb/sec (622616400bps) errors: 7915231
>> Batch: 1891404pps 635Mb/sec (635511744bps) errors: 4926093
>> Buf_api: 1844008pps 619Mb/sec (619586688bps) errors: 47766692
>>
>> Testpmd reports:
>> Baseline: 1854448pps, 860464156 bps. [3]
>> Batch: 1892844.25pps, 878280070bps.
>> Buf_api: 1846139.75pps, 856609120bps.
>>
>> Any thoughts?
>>
>> Thanks!
>>
>> [1]
>> Testpmd options: -l 1,3
>> --vdev=virtio_user0,mac=01:02:03:04:05:06,path=/dev/vhost-net,queue_size=1024
>> -- --auto-start --stats-period 5 --tx-offloads="$TX_OFFLOADS"
>> --rx-offloads="$RX_OFFLOADS" --txd=4096 --rxd=4096 --burst=512
>> --forward-mode=txonly
>>
>> Where offloads were obtained manually running with
>> --[tr]x-offloads=0x8fff and examining testpmd response:
>> declare -r RX_OFFLOADS=0x81d
>> declare -r TX_OFFLOADS=0x802d
>>
>> All of the tests results are an average of at least 3 samples of
>> testpmd, discarding the obvious deviations at start/end (like warming
>> up or waiting for pktgen to start). The result of pktgen is directly
>> c&p from its output.
>>
>> The numbers do not change very much from one stats printing to another
>> of testpmd.
>>
>> [2] Obtained subtracting each accumulated tx-packets from one stats
>> print to the previous one. If we attend testpmd output about Tx-pps,
>> it counts a little bit less performance, but it follows the same
>> pattern:
>>
>> Testpmd pps/bps stats:
>> Baseline: 3510826.25 pps, 1797887912bps = 224735989bytes/sec
>> Batch: 3448515.571pps, 1765640226bps = 220705028.3bytes/sec
>> Buf api: 3794115.333pps, 1942587286bps = 242823410.8bytes/sec
>>
>> [3] This is obtained using the rx-pps/rx-bps report of testpmd.
>>
>> Seems strange to me that the relation between pps/bps is ~336 this
>> time, and between accumulated pkts/accumulated bytes is ~58. Also, the
>> relation between them is not even close to 8.
>>
>> However, testpmd shows a lot of absolute packets received. If we see
>> the received packets in a period subtracting from the previous one,
>> testpmd tells that receive more pps than pktgen tx-pps:
>> Baseline: ~2222668.667pps 128914784.3bps.
>> Batch: 2269260.933pps, 131617134.9bps
>> Buf_api: 2213226.467pps, 128367135.9bp
> How about playing with the batch size? Make it a mod parameter instead
> of the hard coded 64, and measure for all values 1 to 64 ...


Right, according to the test result, 64 seems to be too aggressive in 
the case of TX.

And it might also be worth to check:

1) Whether vhost thread is stressed as 100% CPU utilization, if not, 
there's bottleneck elsewhere
2) For RX test, make sure pktgen kthread is running in the same NUMA 
node with virtio-user

Thanks


>