* [RFC] Define well known packet burst sizes
@ 2024-10-12 12:35 Morten Brørup
2024-10-12 13:55 ` Pavan Nikhilesh Bhagavatula
0 siblings, 1 reply; 3+ messages in thread
From: Morten Brørup @ 2024-10-12 12:35 UTC (permalink / raw)
To: dev
We should define some "well known" packet burst sizes in rte_config.h.
Especially the default packet burst size is interesting;
if known at compile time, various drivers and libraries can optimize for it (i.e. special handling for nb_pkts == RTE_PKT_BURST_DEFAULT).
It should also be used in DPDK examples and apps, instead of defining MAX_PKT_BURST in each and everyone.
Specifically:
/**
* Default packet burst size.
*
* Also intended for optimizing packet processing (e.g. by loop unrolling).
*/
#define RTE_PKT_BURST_DEFAULT 32
/**
* Largest packet burst size guaranteed to be supported throughout DPDK.
*
* Also intended for sizing large temporary arrays of mbufs, e.g. in rte_pktmbuf_free_bulk().
*/
#define RTE_PKT_BURST_MAX 512
#define RTE_MEMPOOL_CACHE_MAX_SIZE RTE_PKT_BURST_MAX
/**
* Smallest packet burst size recommended for latency sensitive applications, when throughput still matters.
*
* Also intended for sizing small staging arrays of mbufs, e.g. in drivers.
*
* Note: Corresponds to one CPU cache line of object pointers.
* - 8 on most 64 bit architectures, 16 on POWER architecture (ppc_64).
* - 16 on all 32 bit architectures.
*/
#define RTE_PKT_BURST_SMALL (RTE_CACHE_LINE_SIZE / sizeof(void *))
^ permalink raw reply [flat|nested] 3+ messages in thread
* RE: [RFC] Define well known packet burst sizes
2024-10-12 12:35 [RFC] Define well known packet burst sizes Morten Brørup
@ 2024-10-12 13:55 ` Pavan Nikhilesh Bhagavatula
2024-10-12 15:07 ` Morten Brørup
0 siblings, 1 reply; 3+ messages in thread
From: Pavan Nikhilesh Bhagavatula @ 2024-10-12 13:55 UTC (permalink / raw)
To: Morten Brørup, dev
> We should define some "well known" packet burst sizes in rte_config.h.
>
A while back we had the same idea, except that it should be platform specific
Ex., for CNXK optimal burst size across workloads is 64.
Instead of rte_config.h maybe we should have it as a meson option in
meson_options.txt with default as 32.
That way platforms can also modify it in config/<arch>/meson.build
> Especially the default packet burst size is interesting;
> if known at compile time, various drivers and libraries can optimize for it (i.e.
> special handling for nb_pkts == RTE_PKT_BURST_DEFAULT).
>
> It should also be used in DPDK examples and apps, instead of defining
> MAX_PKT_BURST in each and everyone.
>
>
> Specifically:
>
> /**
> * Default packet burst size.
> *
> * Also intended for optimizing packet processing (e.g. by loop unrolling).
> */
> #define RTE_PKT_BURST_DEFAULT 32
>
> /**
> * Largest packet burst size guaranteed to be supported throughout DPDK.
> *
> * Also intended for sizing large temporary arrays of mbufs, e.g. in
> rte_pktmbuf_free_bulk().
> */
> #define RTE_PKT_BURST_MAX 512
> #define RTE_MEMPOOL_CACHE_MAX_SIZE RTE_PKT_BURST_MAX
>
> /**
> * Smallest packet burst size recommended for latency sensitive applications,
> when throughput still matters.
> *
> * Also intended for sizing small staging arrays of mbufs, e.g. in drivers.
> *
> * Note: Corresponds to one CPU cache line of object pointers.
> * - 8 on most 64 bit architectures, 16 on POWER architecture (ppc_64).
> * - 16 on all 32 bit architectures.
> */
> #define RTE_PKT_BURST_SMALL (RTE_CACHE_LINE_SIZE / sizeof(void *))
^ permalink raw reply [flat|nested] 3+ messages in thread
* RE: [RFC] Define well known packet burst sizes
2024-10-12 13:55 ` Pavan Nikhilesh Bhagavatula
@ 2024-10-12 15:07 ` Morten Brørup
0 siblings, 0 replies; 3+ messages in thread
From: Morten Brørup @ 2024-10-12 15:07 UTC (permalink / raw)
To: Pavan Nikhilesh Bhagavatula, dev
> From: Pavan Nikhilesh Bhagavatula [mailto:pbhagavatula@marvell.com]
> Sent: Saturday, 12 October 2024 15.55
>
> > We should define some "well known" packet burst sizes in
> rte_config.h.
> >
>
> A while back we had the same idea, except that it should be platform
> specific
> Ex., for CNXK optimal burst size across workloads is 64.
>
> Instead of rte_config.h maybe we should have it as a meson option in
> meson_options.txt with default as 32.
>
> That way platforms can also modify it in config/<arch>/meson.build
Agree. The default burst size should be platform specific for optimal performance. The "generic" platform value can be 32.
And applications with different needs can use a meson option to override the platform specific default.
The other two suggested "well known" (intended to be commonly used) burst sizes can reside in rte_config.h.
>
> > Especially the default packet burst size is interesting;
> > if known at compile time, various drivers and libraries can optimize
> for it (i.e.
> > special handling for nb_pkts == RTE_PKT_BURST_DEFAULT).
> >
> > It should also be used in DPDK examples and apps, instead of defining
> > MAX_PKT_BURST in each and everyone.
> >
> >
> > Specifically:
> >
> > /**
> > * Default packet burst size.
> > *
> > * Also intended for optimizing packet processing (e.g. by loop
> unrolling).
> > */
> > #define RTE_PKT_BURST_DEFAULT 32
> >
> > /**
> > * Largest packet burst size guaranteed to be supported throughout
> DPDK.
> > *
> > * Also intended for sizing large temporary arrays of mbufs, e.g. in
> > rte_pktmbuf_free_bulk().
> > */
> > #define RTE_PKT_BURST_MAX 512
> > #define RTE_MEMPOOL_CACHE_MAX_SIZE RTE_PKT_BURST_MAX
> >
> > /**
> > * Smallest packet burst size recommended for latency sensitive
> applications,
> > when throughput still matters.
> > *
> > * Also intended for sizing small staging arrays of mbufs, e.g. in
> drivers.
> > *
> > * Note: Corresponds to one CPU cache line of object pointers.
> > * - 8 on most 64 bit architectures, 16 on POWER architecture
> (ppc_64).
> > * - 16 on all 32 bit architectures.
> > */
> > #define RTE_PKT_BURST_SMALL (RTE_CACHE_LINE_SIZE / sizeof(void *))
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2024-10-12 15:07 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2024-10-12 12:35 [RFC] Define well known packet burst sizes Morten Brørup
2024-10-12 13:55 ` Pavan Nikhilesh Bhagavatula
2024-10-12 15:07 ` Morten Brørup
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).