[dpdk-dev] [RFC] lib/librte_ether: consistent PMD batching behavior
konstantin.ananyev at intel.com
Fri Jan 20 12:24:40 CET 2017
> From: Andrew Rybchenko [mailto:arybchenko at solarflare.com]
> Sent: Friday, January 20, 2017 10:26 AM
> To: Yang, Zhiyong <zhiyong.yang at intel.com>; dev at dpdk.org
> Cc: thomas.monjalon at 6wind.com; Richardson, Bruce <bruce.richardson at intel.com>; Ananyev, Konstantin
> <konstantin.ananyev at intel.com>
> Subject: Re: [dpdk-dev] [RFC] lib/librte_ether: consistent PMD batching behavior
> On 01/20/2017 12:51 PM, Zhiyong Yang wrote:
> The rte_eth_tx_burst() function in the file Rte_ethdev.h is invoked to
> transmit output packets on the output queue for DPDK applications as
> static inline uint16_t
> rte_eth_tx_burst(uint8_t port_id, uint16_t queue_id,
> struct rte_mbuf **tx_pkts, uint16_t nb_pkts);
> Note: The fourth parameter nb_pkts: The number of packets to transmit.
> The rte_eth_tx_burst() function returns the number of packets it actually
> sent. The return value equal to *nb_pkts* means that all packets have been
> sent, and this is likely to signify that other output packets could be
> immediately transmitted again. Applications that implement a "send as many
> packets to transmit as possible" policy can check this specific case and
> keep invoking the rte_eth_tx_burst() function until a value less than
> *nb_pkts* is returned.
> When you call TX only once in rte_eth_tx_burst, you may get different
> behaviors from different PMDs. One problem that every DPDK user has to
> face is that they need to take the policy into consideration at the app-
> lication level when using any specific PMD to send the packets whether or
> not it is necessary, which brings usage complexities and makes DPDK users
> easily confused since they have to learn the details on TX function limit
> of specific PMDs and have to handle the different return value: the number
> of packets transmitted successfully for various PMDs. Some PMDs Tx func-
> tions have a limit of sending at most 32 packets for every invoking, some
> PMDs have another limit of at most 64 packets once, another ones have imp-
> lemented to send as many packets to transmit as possible, etc. This will
> easily cause wrong usage for DPDK users.
> This patch proposes to implement the above policy in DPDK lib in order to
> simplify the application implementation and avoid the incorrect invoking
> as well. So, DPDK Users don't need to consider the implementation policy
> and to write duplicated code at the application level again when sending
> packets. In addition to it, the users don't need to know the difference of
> specific PMD TX and can transmit the arbitrary number of packets as they
> expect when invoking TX API rte_eth_tx_burst, then check the return value
> to get the number of packets actually sent.
> How to implement the policy in DPDK lib? Two solutions are proposed below.
> Solution 1:
> Implement the wrapper functions to remove some limits for each specific
> PMDs as i40e_xmit_pkts_simple and ixgbe_xmit_pkts_simple do like that.
> > IMHO, the solution is a bit better since it:
> > 1. Does not affect other PMDs at all
> > 2. Could be a bit faster for the PMDs which require it since has no indirect
> > function call on each iteration
> > 3. No ABI change
I also would prefer solution number 1 for the reasons outlined by Andrew above.
Also, IMO current limitation for number of packets to TX in some Intel PMD TX routines
are sort of artificial:
- they are not caused by any real HW limitations
- avoiding them at PMD level shouldn't cause any performance or functional degradation.
So I don't see any good reason why instead of fixing these limitations in
our own PMDs we are trying to push them to the upper (rte_ethdev) layer.
> Solution 2:
> Implement the policy in the function rte_eth_tx_burst() at the ethdev lay-
> er in a more consistent batching way. Make best effort to send *nb_pkts*
> packets with bursts of no more than 32 by default since many DPDK TX PMDs
> are using this max TX burst size(32). In addition, one data member which
> defines the max TX burst size such as "uint16_t max_tx_burst_pkts;"will be
> added to rte_eth_dev_data, which drivers can override if they work with
> bursts of 64 or other NB(thanks for Bruce <bruce.richardson at intel.com>'s
> suggestion). This can reduce the performance impacting to the lowest limit.
> > I see no noticeable difference in performance, so don't mind if this is finally choosen.
> > Just be sure that you update all PMDs to set reasonable default values, or may be
> > even better, set UINT16_MAX in generic place - 0 is a bad default here.
> > (Lost few seconds wondering why nothing is sent and cannot stop)
> I prefer the latter between the 2 solutions because it makes DPDK code more
> consistent and easier and avoids to write too much duplicate logic in DPDK
> source code. In addition, I think no or a little performance drop is
> brought by solution 2. But ABI change will be introduced.
> In fact, the current rte_eth_rx_burst() function is using the similar
> mechanism and faces the same problem as rte_eth_tx_burst().
> 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);
> Applications are responsible of implementing the policy "retrieve as many
> received packets as possible", and check this specific case and keep
> invoking the rte_eth_rx_burst() function until a value less than *nb_pkts*
> is returned.
> The patch proposes to apply the above method to rte_eth_rx_burst() as well.
> In summary, The purpose of the RFC makes the job easier and more simple for
> driver writers and avoids to write too much duplicate code at the applica-
> tion level.
More information about the dev