Virtualization - Jul 2019 - [PATCH] virtio-net: parameterize min ring num_free for virtio receive

On 2019/7/19 22:29, Jiang wrote:
>
> On 2019/7/19 10:36, Jason Wang wrote:
>>
>> On 2019/7/18 ??10:43, Michael S. Tsirkin wrote:
>>> On Thu, Jul 18, 2019 at 10:42:47AM -0400, Michael S. Tsirkin wrote:
>>>> On Thu, Jul 18, 2019 at 10:01:05PM +0800, Jason Wang wrote:
>>>>> On 2019/7/18 ??9:04, Michael S. Tsirkin wrote:
>>>>>> On Thu, Jul 18, 2019 at 12:55:50PM +0000, ? jiang
wrote:
>>>>>>> This change makes ring buffer reclaim threshold
num_free
>>>>>>> configurable
>>>>>>> for better performance, while it's hard coded
as 1/2 * queue now.
>>>>>>> According to our test with qemu + dpdk, packet
dropping happens
>>>>>>> when
>>>>>>> the guest is not able to provide free buffer in
avail ring timely.
>>>>>>> Smaller value of num_free does decrease the number
of packet
>>>>>>> dropping
>>>>>>> during our test as it makes virtio_net reclaim
buffer earlier.
>>>>>>>
>>>>>>> At least, we should leave the value changeable to
user while the
>>>>>>> default value as 1/2 * queue is kept.
>>>>>>>
>>>>>>> Signed-off-by: jiangkidd<jiangkidd at
hotmail.com>
>>>>>> That would be one reason, but I suspect it's not
the
>>>>>> true one. If you need more buffer due to jitter
>>>>>> then just increase the queue size. Would be cleaner.
>>>>>>
>>>>>>
>>>>>> However are you sure this is the reason for
>>>>>> packet drops? Do you see them dropped by dpdk
>>>>>> due to lack of space in the ring? As opposed to
>>>>>> by guest?
>>>>>>
>>>>>>
>>>>> Besides those, this patch depends on the user to choose a
suitable
>>>>> threshold
>>>>> which is not good. You need either a good value with
demonstrated
>>>>> numbers or
>>>>> something smarter.
>>>>>
>>>>> Thanks
>>>> I do however think that we have a problem right now:
try_fill_recv can
>>>> take up a long time during which net stack does not run at all.
>>>> Imagine
>>>> a 1K queue - we are talking 512 packets. That's exceessive.
>>
>>
>> Yes, we will starve a fast host in this case.
>>
>>
>>>> ?? napi poll
>>>> weight solves a similar problem, so it might make sense to cap
this at
>>>> napi_poll_weight.
>>>>
>>>> Which will allow tweaking it through a module parameter as a
>>>> side effect :) Maybe just do NAPI_POLL_WEIGHT.
>>> Or maybe NAPI_POLL_WEIGHT/2 like we do at half the queue ;). Please
>>> experiment, measure performance and let the list know
>>>
>>>> Need to be careful though: queues can also be small and I
don't
>>>> think we
>>>> want to exceed queue size / 2, or maybe queue size -
napi_poll_weight.
>>>> Definitely must not exceed the full queue size.
>>
>>
>> Looking at intel, it uses 16 and i40e uses 32.? It looks to me 
>> NAPI_POLL_WEIGHT/2 is better.
>>
>> Jiang, want to try that and post a new patch?
>>
>> Thanks
>>
>>
>>>>
>>>> -- 
>>>> MST
>
> We did have completed several rounds of test with setting the value to 
> budget (64 as the default value). It does improve a lot with pps is 
> below 400pps for a single stream. Let me consolidate the data and will 
> send it soon. Actually, we are confident that it runs out of free 
> buffer in avail ring when packet dropping happens with below systemtap:
>
> Just a snippet:
>
> probe
module("virtio_ring").function("virtqueue_get_buf")
> {
> ??? x = (@cast($_vq,
"vring_virtqueue")->vring->used->idx)-
> (@cast($_vq, "vring_virtqueue")->last_used_idx) ---> we use
this one
> to verify if the queue is full, which means guest is not able to take 
> buffer from the queue timely
>
> ??? if (x<0 && (x+65535)<4096)
> ??? ??? x = x+65535
>
> ??? if((x==1024) && @cast($_vq,
"vring_virtqueue")->vq->callback ==
> callback_addr)
> ??? ??? netrxcount[x] <<< gettimeofday_s()
> }
>
>
> probe
module("virtio_ring").function("virtqueue_add_inbuf")
> {
> ??? y = (@cast($vq,
"vring_virtqueue")->vring->avail->idx)-
> (@cast($vq, "vring_virtqueue")->vring->used->idx)
---> we use this one
> to verify if we run out of free buffer in avail ring
> ??? if (y<0 && (y+65535)<4096)
> ??? ??? y = y+65535
>
> ??? if(@2=="debugon")
> ??? {
> ??? ??? if(y==0 && @cast($vq,
"vring_virtqueue")->vq->callback ==
> callback_addr)
> ??? ??? {
> ??? ??? ??? netrxfreecount[y] <<< gettimeofday_s()
>
> ??? ??? ??? printf("no avail ring left seen, printing most recent 5 
> num free, vq: %lx, current index: %d\n", $vq, recentfreecount)
> ??? ??? ??? for(i=recentfreecount; i!=((recentfreecount+4) % 5); 
> i=((i+1) % 5))
> ??? ??? ??? {
> ??? ??? ??? ??? printf("index: %d, num free: %d\n", i,
recentfree[$vq,
> i])
> ??? ??? ??? }
>
> ??? ??? ??? printf("index: %d, num free: %d\n", i,
recentfree[$vq, i])
> ??? ??? ??? //exit()
> ??? ??? }
> ??? }
> }
>
>
> probe 
> module("virtio_net").statement("virtnet_receive at
drivers/net/virtio_net.c:732")
> {
> ??? recentfreecount++
> ??? recentfreecount = recentfreecount % 5
> ??? recentfree[$rq->vq, recentfreecount] = $rq->vq->num_free
--->
> record the num_free for the last 5 calls to virtnet_receive, so we can 
> see if lowering the bar helps.
> }
>
>
> Here is the result:
>
> no avail ring left seen, printing most recent 5 num free, vq: 
> ffff9c13c1200000, current index: 1
> index: 1, num free: 561
> index: 2, num free: 305
> index: 3, num free: 369
> index: 4, num free: 433
> index: 0, num free: 497
> no avail ring left seen, printing most recent 5 num free, vq: 
> ffff9c13c1200000, current index: 1
> index: 1, num free: 543
> index: 2, num free: 463
> index: 3, num free: 469
> index: 4, num free: 476
> index: 0, num free: 479
> no avail ring left seen, printing most recent 5 num free, vq: 
> ffff9c13c1200000, current index: 2
> index: 2, num free: 555
> index: 3, num free: 414
> index: 4, num free: 420
> index: 0, num free: 427
> index: 1, num free: 491
>
> You can see in the last 4 calls to virtnet_receive before we run out 
> of free buffer and start to relaim, num_free is quite high. So if we 
> can do the reclaim earlier, it will certainly help.
>
> Meanwhile, the patch I proposed actually keeps the default value as 
> 1/2 * queue. So the default behavior remains and only leave the 
> interface to advanced users, who really understands what they are 
> doing. Also, the best value may vary in different environment. Do you 
> still think hardcoding this is better option?
>
>
> Jiang
>
Here is the snippet from our test result. Test1 was done with default 
driver with the value of 1/2 * queue, while test2 is with my patch and 
min_numfree set to 64 (the default budget value). We can see average 
drop packets do decrease a lot in test2. Let me know if you need the 
full testing data.

test1Time??? avgDropPackets??? test2Time??? avgDropPackets??? pps
> 16:21.0??? 12.295??? 56:50.4??? 0??? 300k
> 17:19.1??? 15.244??? 56:50.4??? 0??? 300k
> 18:17.5??? 18.789??? 56:50.4??? 0??? 300k
> 19:15.1??? 14.208??? 56:50.4??? 0??? 300k
> 20:13.2??? 20.818??? 56:50.4??? 0.267??? 300k
> 21:11.2??? 12.397??? 56:50.4??? 0??? 300k
> 22:09.3??? 12.599??? 56:50.4??? 0??? 300k
> 23:07.3??? 15.531??? 57:48.4??? 0??? 300k
> 24:05.5??? 13.664??? 58:46.5??? 0??? 300k
> 25:03.7??? 13.158??? 59:44.5??? 4.73??? 300k
> 26:01.1??? 2.486??? 00:42.6??? 0??? 300k
> 26:59.1??? 11.241??? 01:40.6??? 0??? 300k
> 27:57.2??? 20.521??? 02:38.6??? 0??? 300k
> 28:55.2??? 30.094??? 03:36.7??? 0??? 300k
> 29:53.3??? 16.828??? 04:34.7??? 0.963??? 300k
> 30:51.3??? 46.916??? 05:32.8??? 0??? 400k
> 31:49.3??? 56.214??? 05:32.8??? 0??? 400k
> 32:47.3??? 58.69??? 05:32.8??? 0??? 400k
> 33:45.3??? 61.486??? 05:32.8??? 0??? 400k
> 34:43.3??? 72.175??? 05:32.8??? 0.598??? 400k
> 35:41.3??? 56.699??? 05:32.8??? 0??? 400k
> 36:39.3??? 61.071??? 05:32.8??? 0??? 400k
> 37:37.3??? 43.355??? 06:30.8??? 0??? 400k
> 38:35.4??? 44.644??? 06:30.8??? 0??? 400k
> 39:33.4??? 72.336??? 06:30.8??? 0??? 400k
> 40:31.4??? 70.676??? 06:30.8??? 0??? 400k
> 41:29.4??? 108.009??? 06:30.8??? 0??? 400k
> 42:27.4??? 65.216??? 06:30.8??? 0??? 400k

Jiang

On 2019/7/20 0:13, Michael S. Tsirkin wrote:
> On Fri, Jul 19, 2019 at 03:31:29PM +0000, ? jiang wrote:
>> On 2019/7/19 22:29, Jiang wrote:
>>> On 2019/7/19 10:36, Jason Wang wrote:
>>>> On 2019/7/18 ??10:43, Michael S. Tsirkin wrote:
>>>>> On Thu, Jul 18, 2019 at 10:42:47AM -0400, Michael S.
Tsirkin wrote:
>>>>>> On Thu, Jul 18, 2019 at 10:01:05PM +0800, Jason Wang
wrote:
>>>>>>> On 2019/7/18 ??9:04, Michael S. Tsirkin wrote:
>>>>>>>> On Thu, Jul 18, 2019 at 12:55:50PM +0000, ?
jiang wrote:
>>>>>>>>> This change makes ring buffer reclaim
threshold num_free
>>>>>>>>> configurable
>>>>>>>>> for better performance, while it's hard
coded as 1/2 * queue now.
>>>>>>>>> According to our test with qemu + dpdk,
packet dropping happens
>>>>>>>>> when
>>>>>>>>> the guest is not able to provide free
buffer in avail ring timely.
>>>>>>>>> Smaller value of num_free does decrease the
number of packet
>>>>>>>>> dropping
>>>>>>>>> during our test as it makes virtio_net
reclaim buffer earlier.
>>>>>>>>>
>>>>>>>>> At least, we should leave the value
changeable to user while the
>>>>>>>>> default value as 1/2 * queue is kept.
>>>>>>>>>
>>>>>>>>> Signed-off-by: jiangkidd<jiangkidd at
hotmail.com>
>>>>>>>> That would be one reason, but I suspect
it's not the
>>>>>>>> true one. If you need more buffer due to jitter
>>>>>>>> then just increase the queue size. Would be
cleaner.
>>>>>>>>
>>>>>>>>
>>>>>>>> However are you sure this is the reason for
>>>>>>>> packet drops? Do you see them dropped by dpdk
>>>>>>>> due to lack of space in the ring? As opposed to
>>>>>>>> by guest?
>>>>>>>>
>>>>>>>>
>>>>>>> Besides those, this patch depends on the user to
choose a suitable
>>>>>>> threshold
>>>>>>> which is not good. You need either a good value
with demonstrated
>>>>>>> numbers or
>>>>>>> something smarter.
>>>>>>>
>>>>>>> Thanks
>>>>>> I do however think that we have a problem right now:
try_fill_recv can
>>>>>> take up a long time during which net stack does not run
at all.
>>>>>> Imagine
>>>>>> a 1K queue - we are talking 512 packets. That's
exceessive.
>>>>
>>>> Yes, we will starve a fast host in this case.
>>>>
>>>>
>>>>>>  ?? napi poll
>>>>>> weight solves a similar problem, so it might make sense
to cap this at
>>>>>> napi_poll_weight.
>>>>>>
>>>>>> Which will allow tweaking it through a module parameter
as a
>>>>>> side effect :) Maybe just do NAPI_POLL_WEIGHT.
>>>>> Or maybe NAPI_POLL_WEIGHT/2 like we do at half the queue
;). Please
>>>>> experiment, measure performance and let the list know
>>>>>
>>>>>> Need to be careful though: queues can also be small and
I don't
>>>>>> think we
>>>>>> want to exceed queue size / 2, or maybe queue size -
napi_poll_weight.
>>>>>> Definitely must not exceed the full queue size.
>>>>
>>>> Looking at intel, it uses 16 and i40e uses 32.? It looks to me
>>>> NAPI_POLL_WEIGHT/2 is better.
>>>>
>>>> Jiang, want to try that and post a new patch?
>>>>
>>>> Thanks
>>>>
>>>>
>>>>>> -- 
>>>>>> MST
>>> We did have completed several rounds of test with setting the value
to
>>> budget (64 as the default value). It does improve a lot with pps is
>>> below 400pps for a single stream. Let me consolidate the data and
will
>>> send it soon. Actually, we are confident that it runs out of free
>>> buffer in avail ring when packet dropping happens with below
systemtap:
>>>
>>> Just a snippet:
>>>
>>> probe
module("virtio_ring").function("virtqueue_get_buf")
>>> {
>>>  ??? x = (@cast($_vq,
"vring_virtqueue")->vring->used->idx)-
>>> (@cast($_vq, "vring_virtqueue")->last_used_idx)
---> we use this one
>>> to verify if the queue is full, which means guest is not able to
take
>>> buffer from the queue timely
>>>
>>>  ??? if (x<0 && (x+65535)<4096)
>>>  ??? ??? x = x+65535
>>>
>>>  ??? if((x==1024) && @cast($_vq,
"vring_virtqueue")->vq->callback =>>> callback_addr)
>>>  ??? ??? netrxcount[x] <<< gettimeofday_s()
>>> }
>>>
>>>
>>> probe
module("virtio_ring").function("virtqueue_add_inbuf")
>>> {
>>>  ??? y = (@cast($vq,
"vring_virtqueue")->vring->avail->idx)-
>>> (@cast($vq,
"vring_virtqueue")->vring->used->idx) ---> we use this one
>>> to verify if we run out of free buffer in avail ring
>>>  ??? if (y<0 && (y+65535)<4096)
>>>  ??? ??? y = y+65535
>>>
>>>  ??? if(@2=="debugon")
>>>  ??? {
>>>  ??? ??? if(y==0 && @cast($vq,
"vring_virtqueue")->vq->callback =>>> callback_addr)
>>>  ??? ??? {
>>>  ??? ??? ??? netrxfreecount[y] <<< gettimeofday_s()
>>>
>>>  ??? ??? ??? printf("no avail ring left seen, printing most
recent 5
>>> num free, vq: %lx, current index: %d\n", $vq, recentfreecount)
>>>  ??? ??? ??? for(i=recentfreecount; i!=((recentfreecount+4) % 5);
>>> i=((i+1) % 5))
>>>  ??? ??? ??? {
>>>  ??? ??? ??? ??? printf("index: %d, num free: %d\n", i,
recentfree[$vq,
>>> i])
>>>  ??? ??? ??? }
>>>
>>>  ??? ??? ??? printf("index: %d, num free: %d\n", i,
recentfree[$vq, i])
>>>  ??? ??? ??? //exit()
>>>  ??? ??? }
>>>  ??? }
>>> }
>>>
>>>
>>> probe
>>> module("virtio_net").statement("virtnet_receive at
drivers/net/virtio_net.c:732")
>>> {
>>>  ??? recentfreecount++
>>>  ??? recentfreecount = recentfreecount % 5
>>>  ??? recentfree[$rq->vq, recentfreecount] =
$rq->vq->num_free --->
>>> record the num_free for the last 5 calls to virtnet_receive, so we
can
>>> see if lowering the bar helps.
>>> }
>>>
>>>
>>> Here is the result:
>>>
>>> no avail ring left seen, printing most recent 5 num free, vq:
>>> ffff9c13c1200000, current index: 1
>>> index: 1, num free: 561
>>> index: 2, num free: 305
>>> index: 3, num free: 369
>>> index: 4, num free: 433
>>> index: 0, num free: 497
>>> no avail ring left seen, printing most recent 5 num free, vq:
>>> ffff9c13c1200000, current index: 1
>>> index: 1, num free: 543
>>> index: 2, num free: 463
>>> index: 3, num free: 469
>>> index: 4, num free: 476
>>> index: 0, num free: 479
>>> no avail ring left seen, printing most recent 5 num free, vq:
>>> ffff9c13c1200000, current index: 2
>>> index: 2, num free: 555
>>> index: 3, num free: 414
>>> index: 4, num free: 420
>>> index: 0, num free: 427
>>> index: 1, num free: 491
>>>
>>> You can see in the last 4 calls to virtnet_receive before we run
out
>>> of free buffer and start to relaim, num_free is quite high. So if
we
>>> can do the reclaim earlier, it will certainly help.
>>>
>>> Meanwhile, the patch I proposed actually keeps the default value as
>>> 1/2 * queue. So the default behavior remains and only leave the
>>> interface to advanced users, who really understands what they are
>>> doing. Also, the best value may vary in different environment. Do
you
>>> still think hardcoding this is better option?
>>>
>>>
>>> Jiang
>>>
>> Here is the snippet from our test result. Test1 was done with default
>> driver with the value of 1/2 * queue, while test2 is with my patch and
>> min_numfree set to 64 (the default budget value). We can see average
>> drop packets do decrease a lot in test2. Let me know if you need the
>> full testing data.
>>
>> test1Time??? avgDropPackets??? test2Time??? avgDropPackets??? pps
>>
>>> 16:21.0??? 12.295??? 56:50.4??? 0??? 300k
>>> 17:19.1??? 15.244??? 56:50.4??? 0??? 300k
>>> 18:17.5??? 18.789??? 56:50.4??? 0??? 300k
>>> 19:15.1??? 14.208??? 56:50.4??? 0??? 300k
>>> 20:13.2??? 20.818??? 56:50.4??? 0.267??? 300k
>>> 21:11.2??? 12.397??? 56:50.4??? 0??? 300k
>>> 22:09.3??? 12.599??? 56:50.4??? 0??? 300k
>>> 23:07.3??? 15.531??? 57:48.4??? 0??? 300k
>>> 24:05.5??? 13.664??? 58:46.5??? 0??? 300k
>>> 25:03.7??? 13.158??? 59:44.5??? 4.73??? 300k
>>> 26:01.1??? 2.486??? 00:42.6??? 0??? 300k
>>> 26:59.1??? 11.241??? 01:40.6??? 0??? 300k
>>> 27:57.2??? 20.521??? 02:38.6??? 0??? 300k
>>> 28:55.2??? 30.094??? 03:36.7??? 0??? 300k
>>> 29:53.3??? 16.828??? 04:34.7??? 0.963??? 300k
>>> 30:51.3??? 46.916??? 05:32.8??? 0??? 400k
>>> 31:49.3??? 56.214??? 05:32.8??? 0??? 400k
>>> 32:47.3??? 58.69??? 05:32.8??? 0??? 400k
>>> 33:45.3??? 61.486??? 05:32.8??? 0??? 400k
>>> 34:43.3??? 72.175??? 05:32.8??? 0.598??? 400k
>>> 35:41.3??? 56.699??? 05:32.8??? 0??? 400k
>>> 36:39.3??? 61.071??? 05:32.8??? 0??? 400k
>>> 37:37.3??? 43.355??? 06:30.8??? 0??? 400k
>>> 38:35.4??? 44.644??? 06:30.8??? 0??? 400k
>>> 39:33.4??? 72.336??? 06:30.8??? 0??? 400k
>>> 40:31.4??? 70.676??? 06:30.8??? 0??? 400k
>>> 41:29.4??? 108.009??? 06:30.8??? 0??? 400k
>>> 42:27.4??? 65.216??? 06:30.8??? 0??? 400k
>>
>> Jiang
>
> OK I find this surprising but I accept what you see.
> I'm inclined not to add a tunable and just select
> a value ourselves.
> I'm also fine with using the napi poll module parameter
> which will give you a bit of tunability.
OK, kindly take a look if you prefer the below code change. I tested 
budget/2 and the result is almost the same as budget when pps below 
400k, but a little better when it goes beyond 400k in my environment.

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c

index 0d4115c9e20b..bc08be7925eb 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -1331,7 +1331,7 @@ static int virtnet_receive(struct receive_queue 
*rq, int budget,
 ??????????????? }
 ??????? }

-?????? if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2)
{
+?????? if (rq->vq->num_free > min((unsigned int)budget, 
virtqueue_get_vring_size(rq->vq)) / 2) {
 ??????????????? if (!try_fill_recv(vi, rq, GFP_ATOMIC))
 ??????????????????????? schedule_delayed_work(&vi->refill, 0);
 ??????? }


Jiang