[dpdk-dev] [PATCH] app/testpmd: add packet data prefetch in macswap loop

[Ananyev, Konstantin]

> -----Original Message-----
> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Bruce Richardson
> Sent: Tuesday, May 03, 2016 11:20 AM
> To: Ivan Boule
> Cc: Jerin Jacob; dev at dpdk.org; De Lara Guarch, Pablo
> Subject: Re: [dpdk-dev] [PATCH] app/testpmd: add packet data prefetch in macswap loop
> 
> On Tue, May 03, 2016 at 11:50:31AM +0200, Ivan Boule wrote:
> > On 05/03/2016 11:45 AM, Bruce Richardson wrote:
> > >On Mon, May 02, 2016 at 05:29:37PM +0530, Jerin Jacob wrote:
> > >>prefetch the next packet data address in advance in macswap loop
> > >>for performance improvement.
> > >>
> > >>...
> > >>  	for (i = 0; i < nb_rx; i++) {
> > >>+		if (likely(i < nb_rx - 1))
> > >>+			rte_prefetch0(rte_pktmbuf_mtod(pkts_burst[i + 1],
> > >>+						       void *));
> > >
> > >At least on IA platforms, there is no issue with prefetching beyond the end of
> > >the array, since it's only a hint to the cpu. If this is true for other platforms,
> > >then I suggest we just drop the conditional and just always prefetch.
> >
> > This is an interesting point.
> > Bruce, are you suggesting that prefetching at an invalid [virtual] address
> > won't trigger a CPU exception?
> >
> 
> Yep. For example, adding "rte_prefetch0(NULL)" at the start of main in testpmd
> causes no ill effects when running the app.
> 

One correction - while on IA prefetch(inval_addr) wouldn't cause any functional problems,
it still might cause DTLB miss and can be a source of noticeable performance degradation.
So it is better to avoid such constructions for performance critical code.
Konstantin