diff mbox

[dpdk-dev] vhost: fix virtio_net cache sharing of broadcast_rarp

Message ID fb0af1bf-9518-f8f2-d288-f8f2e77a53c8@redhat.com (mailing list archive)
State Not Applicable, archived
Headers show

Checks

Context Check Description
ci/Intel-compilation fail apply patch file failure
ci/checkpatch warning coding style issues

Commit Message

Kevin Traynor March 16, 2017, 10:10 a.m. UTC
On 03/16/2017 06:21 AM, Yuanhan Liu wrote:
> On Wed, Mar 15, 2017 at 07:10:49PM +0000, Kevin Traynor wrote:
>> The virtio_net structure is used in both enqueue and dequeue datapaths.
>> broadcast_rarp is checked with cmpset in the dequeue datapath regardless
>> of whether descriptors are available or not.
>>
>> It is observed in some cases where dequeue and enqueue are performed by
>> different cores and no packets are available on the dequeue datapath
>> (i.e. uni-directional traffic), the frequent checking of broadcast_rarp
>> in dequeue causes performance degradation for the enqueue datapath.
>>
>> In OVS the issue can cause a uni-directional performance drop of up to 15%.
>>
>> Fix that by moving broadcast_rarp to a different cache line in
>> virtio_net struct.
> 
> Thanks, but I'm a bit confused. The drop looks like being caused by
> cache false sharing, but I don't see anything would lead to a false
> sharing. I mean, there is no write in the same cache line where the
> broadcast_rarp belongs. Or, the "volatile" type is the culprit here?
> 

Yes, the cmpset code uses cmpxchg and that performs a write regardless
of the result - it either writes the new value or back the old value.

> Talking about that, I had actually considered to turn "broadcast_rarp"
> to a simple "int" or "uint16_t" type, to make it more light weight.
> The reason I used atomic type is to exactly send one broadcast RARP
> packet once SEND_RARP request is recieved. Otherwise, we may send more
> than one RARP packet when MQ is invovled. But I think we don't have
> to be that accurate: it's tolerable when more RARP are sent. I saw 4
> SEND_RARP requests (aka 4 RARP packets) in the last time I tried
> vhost-user live migration after all. I don't quite remember why
> it was 4 though.
> 
> That said, I think it also would resolve the performance issue if you
> change "rte_atomic16_t" to "uint16_t", without moving the place?
> 

Yes, that should work fine, with the side effect you mentioned of
possibly some more rarps - no big deal.

I tested another solution also - as it is unlikely we would need to send
the broadcast_rarp, you can first read and only do the cmpset if it is
likely to succeed. This resolved the issue too.


I choose changing the struct because the 'read && cmpset' code is
non-obvious and someone might not think to do that in the future. I did
a PVP test with testpmd and didn't see any degradation with the struct
change, so I thought it can be a good solution.

I tested the struct change with several combinations of DPDK
16.11.1/17.02/master combined with OVS 2.6/2.7/master. If you prefer one
of the other solutions, let me know and I'll perform some additional
testing.

Kevin.

> 	--yliu
>

Comments

Yuanhan Liu March 17, 2017, 5:47 a.m. UTC | #1
On Thu, Mar 16, 2017 at 10:10:05AM +0000, Kevin Traynor wrote:
> On 03/16/2017 06:21 AM, Yuanhan Liu wrote:
> > On Wed, Mar 15, 2017 at 07:10:49PM +0000, Kevin Traynor wrote:
> >> The virtio_net structure is used in both enqueue and dequeue datapaths.
> >> broadcast_rarp is checked with cmpset in the dequeue datapath regardless
> >> of whether descriptors are available or not.
> >>
> >> It is observed in some cases where dequeue and enqueue are performed by
> >> different cores and no packets are available on the dequeue datapath
> >> (i.e. uni-directional traffic), the frequent checking of broadcast_rarp
> >> in dequeue causes performance degradation for the enqueue datapath.
> >>
> >> In OVS the issue can cause a uni-directional performance drop of up to 15%.
> >>
> >> Fix that by moving broadcast_rarp to a different cache line in
> >> virtio_net struct.
> > 
> > Thanks, but I'm a bit confused. The drop looks like being caused by
> > cache false sharing, but I don't see anything would lead to a false
> > sharing. I mean, there is no write in the same cache line where the
> > broadcast_rarp belongs. Or, the "volatile" type is the culprit here?
> > 
> 
> Yes, the cmpset code uses cmpxchg and that performs a write regardless
> of the result - it either writes the new value or back the old value.

Oh, right, I missed this part!

> > Talking about that, I had actually considered to turn "broadcast_rarp"
> > to a simple "int" or "uint16_t" type, to make it more light weight.
> > The reason I used atomic type is to exactly send one broadcast RARP
> > packet once SEND_RARP request is recieved. Otherwise, we may send more
> > than one RARP packet when MQ is invovled. But I think we don't have
> > to be that accurate: it's tolerable when more RARP are sent. I saw 4
> > SEND_RARP requests (aka 4 RARP packets) in the last time I tried
> > vhost-user live migration after all. I don't quite remember why
> > it was 4 though.
> > 
> > That said, I think it also would resolve the performance issue if you
> > change "rte_atomic16_t" to "uint16_t", without moving the place?
> > 
> 
> Yes, that should work fine, with the side effect you mentioned of
> possibly some more rarps - no big deal.
> 
> I tested another solution also - as it is unlikely we would need to send
> the broadcast_rarp, you can first read and only do the cmpset if it is
> likely to succeed. This resolved the issue too.
> 
> --- a/lib/librte_vhost/virtio_net.c
> +++ b/lib/librte_vhost/virtio_net.c
> @@ -1057,7 +1057,8 @@ static inline bool __attribute__((always_inline))
>          *
>          * Check user_send_rarp() for more information.
>          */
> -       if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
> +       if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
> +                       rte_atomic16_cmpset((volatile uint16_t *)
>                                          &dev->broadcast_rarp.cnt, 1, 0))) {
>                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
>                 if (rarp_mbuf == NULL) {

I'm okay with this one. It's simple and clean enough, that it could
be picked to a stable release. Later, I'd like to send another patch
to turn it to "uint16_t". Since it changes the behaviour a bit, it
is not a good candidate for stable release.

BTW, would you please include the root cause (false sharing) into
your commit log?

	--yliu
> 
> I choose changing the struct because the 'read && cmpset' code is
> non-obvious and someone might not think to do that in the future. I did
> a PVP test with testpmd and didn't see any degradation with the struct
> change, so I thought it can be a good solution.
> 
> I tested the struct change with several combinations of DPDK
> 16.11.1/17.02/master combined with OVS 2.6/2.7/master. If you prefer one
> of the other solutions, let me know and I'll perform some additional
> testing.
> 
> Kevin.
> 
> > 	--yliu
> >
Maxime Coquelin March 17, 2017, 10:01 a.m. UTC | #2
On 03/17/2017 06:47 AM, Yuanhan Liu wrote:
> On Thu, Mar 16, 2017 at 10:10:05AM +0000, Kevin Traynor wrote:
>> On 03/16/2017 06:21 AM, Yuanhan Liu wrote:
>>> On Wed, Mar 15, 2017 at 07:10:49PM +0000, Kevin Traynor wrote:
>>>> The virtio_net structure is used in both enqueue and dequeue datapaths.
>>>> broadcast_rarp is checked with cmpset in the dequeue datapath regardless
>>>> of whether descriptors are available or not.
>>>>
>>>> It is observed in some cases where dequeue and enqueue are performed by
>>>> different cores and no packets are available on the dequeue datapath
>>>> (i.e. uni-directional traffic), the frequent checking of broadcast_rarp
>>>> in dequeue causes performance degradation for the enqueue datapath.
>>>>
>>>> In OVS the issue can cause a uni-directional performance drop of up to 15%.
>>>>
>>>> Fix that by moving broadcast_rarp to a different cache line in
>>>> virtio_net struct.
>>>
>>> Thanks, but I'm a bit confused. The drop looks like being caused by
>>> cache false sharing, but I don't see anything would lead to a false
>>> sharing. I mean, there is no write in the same cache line where the
>>> broadcast_rarp belongs. Or, the "volatile" type is the culprit here?
>>>
>>
>> Yes, the cmpset code uses cmpxchg and that performs a write regardless
>> of the result - it either writes the new value or back the old value.
>
> Oh, right, I missed this part!
>
>>> Talking about that, I had actually considered to turn "broadcast_rarp"
>>> to a simple "int" or "uint16_t" type, to make it more light weight.
>>> The reason I used atomic type is to exactly send one broadcast RARP
>>> packet once SEND_RARP request is recieved. Otherwise, we may send more
>>> than one RARP packet when MQ is invovled. But I think we don't have
>>> to be that accurate: it's tolerable when more RARP are sent. I saw 4
>>> SEND_RARP requests (aka 4 RARP packets) in the last time I tried
>>> vhost-user live migration after all. I don't quite remember why
>>> it was 4 though.
>>>
>>> That said, I think it also would resolve the performance issue if you
>>> change "rte_atomic16_t" to "uint16_t", without moving the place?
>>>
>>
>> Yes, that should work fine, with the side effect you mentioned of
>> possibly some more rarps - no big deal.
>>
>> I tested another solution also - as it is unlikely we would need to send
>> the broadcast_rarp, you can first read and only do the cmpset if it is
>> likely to succeed. This resolved the issue too.
>>
>> --- a/lib/librte_vhost/virtio_net.c
>> +++ b/lib/librte_vhost/virtio_net.c
>> @@ -1057,7 +1057,8 @@ static inline bool __attribute__((always_inline))
>>          *
>>          * Check user_send_rarp() for more information.
>>          */
>> -       if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
>> +       if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
>> +                       rte_atomic16_cmpset((volatile uint16_t *)
>>                                          &dev->broadcast_rarp.cnt, 1, 0))) {
>>                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
>>                 if (rarp_mbuf == NULL) {
>
> I'm okay with this one. It's simple and clean enough, that it could
> be picked to a stable release. Later, I'd like to send another patch
> to turn it to "uint16_t". Since it changes the behaviour a bit, it
> is not a good candidate for stable release.
>
> BTW, would you please include the root cause (false sharing) into
> your commit log?
And maybe also adds the info to the comment just above?
I will help people wondering why we read before cmpset.

Maxime
Kevin Traynor March 20, 2017, 11:13 a.m. UTC | #3
On 03/17/2017 10:01 AM, Maxime Coquelin wrote:
> 
> 
> On 03/17/2017 06:47 AM, Yuanhan Liu wrote:
>> On Thu, Mar 16, 2017 at 10:10:05AM +0000, Kevin Traynor wrote:
>>> On 03/16/2017 06:21 AM, Yuanhan Liu wrote:
>>>> On Wed, Mar 15, 2017 at 07:10:49PM +0000, Kevin Traynor wrote:
>>>>> The virtio_net structure is used in both enqueue and dequeue
>>>>> datapaths.
>>>>> broadcast_rarp is checked with cmpset in the dequeue datapath
>>>>> regardless
>>>>> of whether descriptors are available or not.
>>>>>
>>>>> It is observed in some cases where dequeue and enqueue are
>>>>> performed by
>>>>> different cores and no packets are available on the dequeue datapath
>>>>> (i.e. uni-directional traffic), the frequent checking of
>>>>> broadcast_rarp
>>>>> in dequeue causes performance degradation for the enqueue datapath.
>>>>>
>>>>> In OVS the issue can cause a uni-directional performance drop of up
>>>>> to 15%.
>>>>>
>>>>> Fix that by moving broadcast_rarp to a different cache line in
>>>>> virtio_net struct.
>>>>
>>>> Thanks, but I'm a bit confused. The drop looks like being caused by
>>>> cache false sharing, but I don't see anything would lead to a false
>>>> sharing. I mean, there is no write in the same cache line where the
>>>> broadcast_rarp belongs. Or, the "volatile" type is the culprit here?
>>>>
>>>
>>> Yes, the cmpset code uses cmpxchg and that performs a write regardless
>>> of the result - it either writes the new value or back the old value.
>>
>> Oh, right, I missed this part!
>>
>>>> Talking about that, I had actually considered to turn "broadcast_rarp"
>>>> to a simple "int" or "uint16_t" type, to make it more light weight.
>>>> The reason I used atomic type is to exactly send one broadcast RARP
>>>> packet once SEND_RARP request is recieved. Otherwise, we may send more
>>>> than one RARP packet when MQ is invovled. But I think we don't have
>>>> to be that accurate: it's tolerable when more RARP are sent. I saw 4
>>>> SEND_RARP requests (aka 4 RARP packets) in the last time I tried
>>>> vhost-user live migration after all. I don't quite remember why
>>>> it was 4 though.
>>>>
>>>> That said, I think it also would resolve the performance issue if you
>>>> change "rte_atomic16_t" to "uint16_t", without moving the place?
>>>>
>>>
>>> Yes, that should work fine, with the side effect you mentioned of
>>> possibly some more rarps - no big deal.
>>>
>>> I tested another solution also - as it is unlikely we would need to send
>>> the broadcast_rarp, you can first read and only do the cmpset if it is
>>> likely to succeed. This resolved the issue too.
>>>
>>> --- a/lib/librte_vhost/virtio_net.c
>>> +++ b/lib/librte_vhost/virtio_net.c
>>> @@ -1057,7 +1057,8 @@ static inline bool __attribute__((always_inline))
>>>          *
>>>          * Check user_send_rarp() for more information.
>>>          */
>>> -       if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
>>> +       if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
>>> +                       rte_atomic16_cmpset((volatile uint16_t *)
>>>                                          &dev->broadcast_rarp.cnt, 1,
>>> 0))) {
>>>                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
>>>                 if (rarp_mbuf == NULL) {
>>
>> I'm okay with this one. It's simple and clean enough, that it could
>> be picked to a stable release. Later, I'd like to send another patch
>> to turn it to "uint16_t". Since it changes the behaviour a bit, it
>> is not a good candidate for stable release.
>>
>> BTW, would you please include the root cause (false sharing) into
>> your commit log?
> And maybe also adds the info to the comment just above?
> I will help people wondering why we read before cmpset.
> 

Sure, I will re-spin, do some testing and submit a v2.

> Maxime
diff mbox

Patch

--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -1057,7 +1057,8 @@  static inline bool __attribute__((always_inline))
         *
         * Check user_send_rarp() for more information.
         */
-       if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
+       if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
+                       rte_atomic16_cmpset((volatile uint16_t *)
                                         &dev->broadcast_rarp.cnt, 1, 0))) {
                rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
                if (rarp_mbuf == NULL) {