Re: [dpdk-dev] [PATCH] ixgbe: prefetch packet headers in vector PMD receive function

[Zoltan Kiss <zoltan.kiss@linaro.org>]

On 19/10/15 19:57, Ananyev, Konstantin wrote:
>
>
>> -----Original Message-----
>> From: Zoltan Kiss [mailto:zoltan.kiss@linaro.org]
>> Sent: Monday, October 19, 2015 5:30 PM
>> To: Ananyev, Konstantin; Richardson, Bruce; dev@dpdk.org
>> Subject: Re: [PATCH] ixgbe: prefetch packet headers in vector PMD receive function
>>
>>
>>
>> On 15/10/15 16:43, Ananyev, Konstantin wrote:
>>>
>>>
>>>> -----Original Message-----
>>>> From: Zoltan Kiss [mailto:zoltan.kiss@linaro.org]
>>>> Sent: Thursday, October 15, 2015 11:33 AM
>>>> To: Ananyev, Konstantin; Richardson, Bruce; dev@dpdk.org
>>>> Subject: Re: [PATCH] ixgbe: prefetch packet headers in vector PMD receive function
>>>>
>>>>
>>>>
>>>> On 15/10/15 00:23, Ananyev, Konstantin wrote:
>>>>>
>>>>>
>>>>>> -----Original Message-----
>>>>>> From: Zoltan Kiss [mailto:zoltan.kiss@linaro.org]
>>>>>> Sent: Wednesday, October 14, 2015 5:11 PM
>>>>>> To: Ananyev, Konstantin; Richardson, Bruce; dev@dpdk.org
>>>>>> Subject: Re: [PATCH] ixgbe: prefetch packet headers in vector PMD receive function
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 28/09/15 00:19, Ananyev, Konstantin wrote:
>>>>>>>
>>>>>>>
>>>>>>>> -----Original Message-----
>>>>>>>> From: Zoltan Kiss [mailto:zoltan.kiss@linaro.org]
>>>>>>>> Sent: Friday, September 25, 2015 7:29 PM
>>>>>>>> To: Richardson, Bruce; dev@dpdk.org
>>>>>>>> Cc: Ananyev, Konstantin
>>>>>>>> Subject: Re: [PATCH] ixgbe: prefetch packet headers in vector PMD receive function
>>>>>>>>
>>>>>>>> On 07/09/15 07:41, Richardson, Bruce wrote:
>>>>>>>>>
>>>>>>>>>> -----Original Message-----
>>>>>>>>>> From: Zoltan Kiss [mailto:zoltan.kiss@linaro.org]
>>>>>>>>>> Sent: Monday, September 7, 2015 3:15 PM
>>>>>>>>>> To: Richardson, Bruce; dev@dpdk.org
>>>>>>>>>> Cc: Ananyev, Konstantin
>>>>>>>>>> Subject: Re: [PATCH] ixgbe: prefetch packet headers in vector PMD receive
>>>>>>>>>> function
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On 07/09/15 13:57, Richardson, Bruce wrote:
>>>>>>>>>>>
>>>>>>>>>>>> -----Original Message-----
>>>>>>>>>>>> From: Zoltan Kiss [mailto:zoltan.kiss@linaro.org]
>>>>>>>>>>>> Sent: Monday, September 7, 2015 1:26 PM
>>>>>>>>>>>> To: dev@dpdk.org
>>>>>>>>>>>> Cc: Ananyev, Konstantin; Richardson, Bruce
>>>>>>>>>>>> Subject: Re: [PATCH] ixgbe: prefetch packet headers in vector PMD
>>>>>>>>>>>> receive function
>>>>>>>>>>>>
>>>>>>>>>>>> Hi,
>>>>>>>>>>>>
>>>>>>>>>>>> I just realized I've missed the "[PATCH]" tag from the subject. Did
>>>>>>>>>>>> anyone had time to review this?
>>>>>>>>>>>>
>>>>>>>>>>> Hi Zoltan,
>>>>>>>>>>>
>>>>>>>>>>> the big thing that concerns me with this is the addition of new
>>>>>>>>>>> instructions for each packet in the fast path. Ideally, this
>>>>>>>>>>> prefetching would be better handled in the application itself, as for
>>>>>>>>>>> some apps, e.g. those using pipelining, the core doing the RX from the
>>>>>>>>>>> NIC may not touch the packet data at all, and the prefetches will
>>>>>>>>>> instead cause a performance slowdown.
>>>>>>>>>>> Is it possible to get the same performance increase - or something
>>>>>>>>>>> close to it - by making changes in OVS?
>>>>>>>>>> OVS already does a prefetch when it's processing the previous packet, but
>>>>>>>>>> apparently it's not early enough. At least for my test scenario, where I'm
>>>>>>>>>> forwarding UDP packets with the least possible overhead. I guess in tests
>>>>>>>>>> where OVS does more complex processing it should be fine.
>>>>>>>>>> I'll try to move the prefetch earlier in OVS codebase, but I'm not sure if
>>>>>>>>>> it'll help.
>>>>>>>>> I would suggest trying to prefetch more than one packet ahead. Prefetching 4 or
>>>>>>>>> 8 ahead might work better, depending on the processing being done.
>>>>>>>>
>>>>>>>> I've moved the prefetch earlier, and it seems to work:
>>>>>>>>
>>>>>>>> https://patchwork.ozlabs.org/patch/519017/
>>>>>>>>
>>>>>>>> However it raises the question: should we remove header prefetch from
>>>>>>>> all the other drivers, and the CONFIG_RTE_PMD_PACKET_PREFETCH option?
>>>>>>>
>>>>>>> My vote would be for that.
>>>>>>> Konstantin
>>>>>>
>>>>>> After some further thinking I would rather support the
>>>>>> rte_packet_prefetch() macro (prefetch the header in the driver, and
>>>>>> configure it through CONFIG_RTE_PMD_PACKET_PREFETCH)
>>>>>>
>>>>>> - the prefetch can happen earlier, so if an application needs the header
>>>>>> right away, this is the fastest
>>>>>> - if the application doesn't need header prefetch, it can turn it off
>>>>>> compile time. Same as if we wouldn't have this option.
>>>>>> - if the application has mixed needs (sometimes it needs the header
>>>>>> right away, sometimes it doesn't), it can turn it off and do what it
>>>>>> needs. Same as if we wouldn't have this option.
>>>>>>
>>>>>> A harder decision would be whether it should be turned on or off by
>>>>>> default. Currently it's on, and half of the receive functions don't use it.
>>>>>
>>>>> Yep,  it is sort of a mess right now.
>>>>> Another question if we'd like to keep it and standardise it:
>>>>> at what moment to prefetch: as soon as we realize that HW is done with that buffer,
>>>>> or as late inside rx_burst() as possible?
>>>>> I suppose there is no clear answer for that.
>>>> I think if the application needs the driver start the prefetching, it
>>>> does it because it's already too late when rte_eth_rx_burst() returns.
>>>> So I think it's best to do it as soon as we learn that the HW is done.
>>>
>>> Probably, but it might be situations when it would be more plausible to do it later too.
>> Could you elaborate?
>> If the application wants prefetch to start earlier than could be done by
>> itself (right after rte_eth_rx_burst() returns), earlier is better. So
>> it will have the best chances to have the data in cache.
>> If it would need the data later, then it could do the prefetch by itself.
>>
>>> Again it might depend on particular HW and your memory access pattern.
>>>
>>>>
>>>>> That's why my thought was to just get rid of it.
>>>>> Though if it would be implemented in some standardized way and disabled by default -
>>>>> that's probably ok too.
>>>>> BTW, couldn't that be implemented just via rte_ethdev rx callbacks mechanism?
>>>>> Then we can have the code one place and don't need compile option at all -
>>>>> could be ebabled/disabled dynamically on a per nic basis.
>>>>> Or would it be too high overhead for that?
>>>>
>>>> I think if we go that way, it's better to pass an option to
>>>> rte_eth_rx_burst() and tell if you want the header prefetched by the
>>>> driver. That would be the most flexible.
>>>
>>> That would mean ABI breakage for rx_burst()...
>>> Probably then better a new parameter in rx_conf or something.
>>> Though it still means that each PMD has to implement it on its own.
>> That would be the case in any way, as we are are talking about
>> prefetching in the receive function.
>>
>>> And again there might be PMDs that would just ignore that parameter.
>> If the PMD has a good reason to do that (e.g. prefetch has undesirable
>> effects), I think it's fine.
>>
>>> While for callback it would be all in one and known place.
>> Who and when would call that callback? If the application after
>> rte_eth_rx_burst() returned, it wouldn't have too much use, it could
>> just call prefetch by itself.
>
> I am talking about the callbacks called by rx_burst() itself:

I don't think it would give us any advantage: the app can just prefetch 
the headers after rte_eth_rx_burst(). If you look at the original patch 
I've sent in about this topic:

http://dpdk.org/ml/archives/dev/2015-September/023144.html

It will start prefetching immediately when an iteration of packets is 
finished. If RTE_IXGBE_DESCS_PER_LOOP is 4 and the burst is 32, by the 
time the receive function returns the first packet's headers could be 
almost there in the cache. And actually you can do it earlier than in 
that patch: right when the pointers become available at "B.2 copy 2 mbuf 
point into rx_pkts"


>
> static inline uint16_t
> rte_eth_rx_burst(uint8_t port_id, uint16_t queue_id,
>                   struct rte_mbuf **rx_pkts, const uint16_t nb_pkts)
> {
>          struct rte_eth_dev *dev;
>
>          dev = &rte_eth_devices[port_id];
>
>          int16_t nb_rx = (*dev->rx_pkt_burst)(dev->data->rx_queues[queue_id],
>                          rx_pkts, nb_pkts);
>
> #ifdef RTE_ETHDEV_RXTX_CALLBACKS
>          struct rte_eth_rxtx_callback *cb = dev->post_rx_burst_cbs[queue_id];
>
>          if (unlikely(cb != NULL)) {
>                  do {
>                          nb_rx = cb->fn.rx(port_id, queue_id, rx_pkts, nb_rx,
>                                                  nb_pkts, cb->param);
>                          cb = cb->next;
>                  } while (cb != NULL);
> #endif
>
>          return nb_rx;
> }
>
> Konstantin
>
>>
>>> But again, I didn't measure it, so not sure what will be an impact of the callback itself.
>>> Might be it not worth it.
>>> Konstantin
>>>
>>>>
>>>>> Konstantin
>>>>>
>>>>>
>>>>>
>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>>>
>>>>>>>>> /Bruce
>>>>>>>