> -----邮件原件-----
> 发件人: Feifei Wang
> 发送时间: Wednesday, July 6, 2022 4:53 PM
> 收件人: Konstantin Ananyev <konstantin.v.anan...@yandex.ru>; Yuying
> Zhang <yuying.zh...@intel.com>; Beilei Xing <beilei.x...@intel.com>;
> Ruifeng Wang <ruifeng.w...@arm.com>
> 抄送: dev@dpdk.org; nd <n...@arm.com>; Honnappa Nagarahalli
> <honnappa.nagaraha...@arm.com>; nd <n...@arm.com>
> 主题: 回复: [RFC PATCH v1] net/i40e: put mempool cache out of API
>
>
>
> > -----邮件原件-----
> > 发件人: Konstantin Ananyev <konstantin.v.anan...@yandex.ru>
> > 发送时间: Sunday, July 3, 2022 8:20 PM
> > 收件人: Feifei Wang <feifei.wa...@arm.com>; Yuying Zhang
> > <yuying.zh...@intel.com>; Beilei Xing <beilei.x...@intel.com>; Ruifeng
> > Wang <ruifeng.w...@arm.com>
> > 抄送: dev@dpdk.org; nd <n...@arm.com>; Honnappa Nagarahalli
> > <honnappa.nagaraha...@arm.com>
> > 主题: Re: [RFC PATCH v1] net/i40e: put mempool cache out of API
> >
> >
> > > Refer to "i40e_tx_free_bufs_avx512", this patch puts mempool cache
> > > out of API to free buffers directly. There are two changes different
> > > with previous version:
> > > 1. change txep from "i40e_entry" to "i40e_vec_entry"
> > > 2. put cache out of "mempool_bulk" API to copy buffers into it
> > > directly
> > >
> > > Performance Test with l3fwd neon path:
> > > with this patch
> > > n1sdp: no perforamnce change
> > > amper-altra: +4.0%
> > >
> >
> >
> Thanks for your detailed comments.
>
> > Thanks for RFC, appreciate your effort.
> > So, as I understand - bypassing mempool put/get itself gives about
> > 7-10% speedup for RX/TX on ARM platforms, correct?
> [Feifei] Yes.
>
> >
> > About direct-rearm RX approach you propose:
> > After another thought, probably it is possible to re-arrange it in a
> > way that would help avoid related negatives.
> > The basic idea as follows:
> >
> > 1. Make RXQ sw_ring visible and accessible by 'attached' TX queues.
> > Also make sw_ring de-coupled from RXQ itself, i.e:
> > when RXQ is stopped or even destroyed, related sw_ring may still
> > exist (probably ref-counter or RCU would be sufficient here).
> > All that means we need a common layout/api for rxq_sw_ring
> > and PMDs that would like to support direct-rearming will have to
> > use/obey it.
> [Feifei] de-coupled sw-ring and RXQ may cause dangerous case due to RXQ is
> stopped but elements of it (sw-ring) is still kept and we may forget to free
> this sw-ring in the end.
> Furthermore, if we apply this, we need to separate operation when closing
> RXQ and add Rx sw-ring free operation when closing TXQ. This will be
> complex and it is not conducive to subsequent maintenance if maintainer
> does not understand direct-rearm mode very well.
>
> >
> > 2. Make RXQ sw_ring 'direct' rearming driven by TXQ itself, i.e:
> > at txq_free_bufs() try to store released mbufs inside attached
> > sw_ring directly. If there is no attached sw_ring, or not enough
> > free space in it - continue with mempool_put() as usual.
> > Note that actual arming of HW RXDs still remains responsibility
> > of RX code-path:
> > rxq_rearm(rxq) {
> > ...
> > - check are there are N already filled entries inside rxq_sw_ring.
> > if not, populate them from mempool (usual mempool_get()).
> > - arm related RXDs and mark these sw_ring entries as managed by HW.
> > ...
> > }
> >
> [Feifei] We try to create two modes, one is direct-rearm and the other is
> direct-free like above.
> And by performance comparison, we select direct-rearm which improve
> performance by 7% - 14% compared with direct-free by 3.6% - 7% in n1sdp.
> Furthermore, I think put direct mode in Tx or Rx is equivalent. For direct-
> rearm, if there is no Tx sw-ring, Rx will get mbufs from mempool. For direct-
> fee, if there is no Rx sw-ring, Tx will put mbufs into mempool. At last, what
> affects our decision-making is the improvement of performance.
>
> >
> > So rxq_sw_ring will serve two purposes:
> > - track mbufs that are managed by HW (that what it does now)
> > - private (per RXQ) mbuf cache
> >
> > Now, if TXQ is stopped while RXQ is running - no extra synchronization
> > is required, RXQ would just use
> > mempool_get() to rearm its sw_ring itself.
> >
> > If RXQ is stopped while TXQ is still running - TXQ can still continue
> > to populate related sw_ring till it gets full.
> > Then it will continue with mempool_put() as usual.
> > Of-course it means that user who wants to use this feature should
> > probably account some extra mbufs for such case, or might be
> > rxq_sw_ring can have enable/disable flag to mitigate such situation.
> >
> [Feifei] For direct-rearm, the key point should be the communication
> between TXQ and RXQ when TXQ is stopped. De-coupled sw-ring is complex,
> maybe we can simplify this and assign this to the application. My thought is
> that if direct-rearm is enabled, when users want to close TX port, they must
> firstly close mapped RX port and disable direct-rearm feature. Then they can
> restart RX port.
[Feifei] I think here we can refer to "hairpin queue" in '
rte_eth_dev_tx_queue_stop ':
''
if (rte_eth_dev_is_tx_hairpin_queue(dev, tx_queue_id)) {
RTE_ETHDEV_LOG(INFO,
"Can't stop Tx hairpin queue %"PRIu16" of device with
port_id=%"PRIu16"\n",
tx_queue_id, port_id);
return -EINVAL;
}
"
>
> > As another benefit here - such approach makes possible to use several
> > TXQs (even from different devices) to rearm same RXQ.
> [Feifei] Actually, for direct-rearm, it can use several RXQs to rearm same
> TXQ,
> so this is equivalent for direct-rearm and direct-free. Furthermore, If use
> multiple cores, I think we need to consider synchronization of variables, and
> lock is necessary.
>
> >
> > Have to say, that I am still not sure that 10% RX/TX improvement is
> > worth bypassing mempool completely and introducing all this extra
> > complexity in RX/TX path.
> [Feifei] Thus maybe we can avoid this complexity as much as possible.
> We should not increase the complexity of the bottom layer for convenience,
> but leave it to the user to decide. If user wants performance, he needs to
> consider and operate more.
>
> > But, if we'll still decide to go ahead with direct-rearming, this
> > re-arrangement, I think, should help to keep things clear and avoid
> > introducing new limitations in existing functionality.
> >
> > WDYT?
> >
> > Konstantin
> >
> >
> >
> >
> >
> >
> >
