[dpdk-dev,2/5] ixgbe: add prefetch to improve slow-path tx perf

  Make a small improvement to slow path TX performance by adding in a
prefetch for the second mbuf cache line.
Also move assignment of l2/l3 length values only when needed.

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
---
 lib/librte_pmd_ixgbe/ixgbe_rxtx.c | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)
  

On Wed, Sep 17, 2014 at 11:01:39AM +0100, Bruce Richardson wrote:
> Make a small improvement to slow path TX performance by adding in a
> prefetch for the second mbuf cache line.
> Also move assignment of l2/l3 length values only when needed.
> 
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> ---
>  lib/librte_pmd_ixgbe/ixgbe_rxtx.c | 12 +++++++-----
>  1 file changed, 7 insertions(+), 5 deletions(-)
> 
> diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> index 6f702b3..c0bb49f 100644
> --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> @@ -565,25 +565,26 @@ ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
>  		ixgbe_xmit_cleanup(txq);
>  	}
>  
> +	rte_prefetch0(&txe->mbuf->pool);
> +

Can you explain what all of these prefetches are doing?  It looks to me like
they're just fetching the first caheline of the mempool structure, which it
appears amounts to the pools name.  I don't see that having any use here.

>  	/* TX loop */
>  	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
>  		new_ctx = 0;
>  		tx_pkt = *tx_pkts++;
>  		pkt_len = tx_pkt->pkt_len;
>  
> -		RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
> -
>  		/*
>  		 * Determine how many (if any) context descriptors
>  		 * are needed for offload functionality.
>  		 */
>  		ol_flags = tx_pkt->ol_flags;
> -		vlan_macip_lens.f.vlan_tci = tx_pkt->vlan_tci;
> -		vlan_macip_lens.f.l2_l3_len = tx_pkt->l2_l3_len;
>  
>  		/* If hardware offload required */
>  		tx_ol_req = ol_flags & PKT_TX_OFFLOAD_MASK;
>  		if (tx_ol_req) {
> +			vlan_macip_lens.f.vlan_tci = tx_pkt->vlan_tci;
> +			vlan_macip_lens.f.l2_l3_len = tx_pkt->l2_l3_len;
> +
>  			/* If new context need be built or reuse the exist ctx. */
>  			ctx = what_advctx_update(txq, tx_ol_req,
>  				vlan_macip_lens.data);
> @@ -720,7 +721,7 @@ ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
>  				    &txr[tx_id];
>  
>  				txn = &sw_ring[txe->next_id];
> -				RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
> +				rte_prefetch0(&txn->mbuf->pool);
>  
>  				if (txe->mbuf != NULL) {
>  					rte_pktmbuf_free_seg(txe->mbuf);
> @@ -749,6 +750,7 @@ ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
>  		do {
>  			txd = &txr[tx_id];
>  			txn = &sw_ring[txe->next_id];
> +			rte_prefetch0(&txn->mbuf->pool);
>  
>  			if (txe->mbuf != NULL)
>  				rte_pktmbuf_free_seg(txe->mbuf);
> -- 
> 1.9.3
> 
>
  

> -----Original Message-----
> From: Neil Horman [mailto:nhorman@tuxdriver.com]
> Sent: Wednesday, September 17, 2014 4:21 PM
> To: Richardson, Bruce
> Cc: dev@dpdk.org
> Subject: Re: [dpdk-dev] [PATCH 2/5] ixgbe: add prefetch to improve slow-path tx
> perf
> 
> On Wed, Sep 17, 2014 at 11:01:39AM +0100, Bruce Richardson wrote:
> > Make a small improvement to slow path TX performance by adding in a
> > prefetch for the second mbuf cache line.
> > Also move assignment of l2/l3 length values only when needed.
> >
> > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > ---
> >  lib/librte_pmd_ixgbe/ixgbe_rxtx.c | 12 +++++++-----
> >  1 file changed, 7 insertions(+), 5 deletions(-)
> >
> > diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > index 6f702b3..c0bb49f 100644
> > --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > @@ -565,25 +565,26 @@ ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf
> **tx_pkts,
> >  		ixgbe_xmit_cleanup(txq);
> >  	}
> >
> > +	rte_prefetch0(&txe->mbuf->pool);
> > +
> 
> Can you explain what all of these prefetches are doing?  It looks to me like
> they're just fetching the first caheline of the mempool structure, which it
> appears amounts to the pools name.  I don't see that having any use here.
> 
This does make a decent enough performance difference in my tests (the amount varies depending on the RX path being used by testpmd). 

What I've done with the prefetches is two-fold:
1) changed it from prefetching the mbuf (first cache line) to prefetching the mbuf pool pointer (second cache line) so that when we go to access the pool pointer to free transmitted mbufs we don't get a cache miss. When clearing the ring and freeing mbufs, the pool pointer is the only mbuf field used, so we don't need that first cache line.
2) changed the code to prefetch earlier - in effect to prefetch one mbuf ahead. The original code prefetched the mbuf to be freed as soon as it started processing the mbuf to replace it. Instead now, every time we calculate what the next mbuf position is going to be we prefetch the mbuf in that position (i.e. the mbuf pool pointer we are going to free the mbuf to), even while we are still updating the previous mbuf slot on the ring. This gives the prefetch much more time to resolve and get the data we need in the cache before we need it.

Hope this clarifies things.

/Bruce
  

On Wed, Sep 17, 2014 at 03:35:19PM +0000, Richardson, Bruce wrote:
> 
> > -----Original Message-----
> > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > Sent: Wednesday, September 17, 2014 4:21 PM
> > To: Richardson, Bruce
> > Cc: dev@dpdk.org
> > Subject: Re: [dpdk-dev] [PATCH 2/5] ixgbe: add prefetch to improve slow-path tx
> > perf
> > 
> > On Wed, Sep 17, 2014 at 11:01:39AM +0100, Bruce Richardson wrote:
> > > Make a small improvement to slow path TX performance by adding in a
> > > prefetch for the second mbuf cache line.
> > > Also move assignment of l2/l3 length values only when needed.
> > >
> > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > > ---
> > >  lib/librte_pmd_ixgbe/ixgbe_rxtx.c | 12 +++++++-----
> > >  1 file changed, 7 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > index 6f702b3..c0bb49f 100644
> > > --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > @@ -565,25 +565,26 @@ ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf
> > **tx_pkts,
> > >  		ixgbe_xmit_cleanup(txq);
> > >  	}
> > >
> > > +	rte_prefetch0(&txe->mbuf->pool);
> > > +
> > 
> > Can you explain what all of these prefetches are doing?  It looks to me like
> > they're just fetching the first caheline of the mempool structure, which it
> > appears amounts to the pools name.  I don't see that having any use here.
> > 
> This does make a decent enough performance difference in my tests (the amount varies depending on the RX path being used by testpmd). 
> 
> What I've done with the prefetches is two-fold:
> 1) changed it from prefetching the mbuf (first cache line) to prefetching the mbuf pool pointer (second cache line) so that when we go to access the pool pointer to free transmitted mbufs we don't get a cache miss. When clearing the ring and freeing mbufs, the pool pointer is the only mbuf field used, so we don't need that first cache line.
ok, this makes some sense, but you're not guaranteed to either have that
prefetch be needed, nor are you certain it will still be in cache by the time
you get to the free call.  Seems like it might be preferable to prefecth the
data pointed to by tx_pkt, as you're sure to use that every loop iteration.

> 2) changed the code to prefetch earlier - in effect to prefetch one mbuf ahead. The original code prefetched the mbuf to be freed as soon as it started processing the mbuf to replace it. Instead now, every time we calculate what the next mbuf position is going to be we prefetch the mbuf in that position (i.e. the mbuf pool pointer we are going to free the mbuf to), even while we are still updating the previous mbuf slot on the ring. This gives the prefetch much more time to resolve and get the data we need in the cache before we need it.
> 
Again, early isn't necessecarily better, as it just means more time for the data
in cache to get victimized. It seems like it would be better to prefetch the
tx_pkts data a few cache lines ahead.

Neil

> Hope this clarifies things.
> 
> /Bruce
>
  

On Wed, Sep 17, 2014 at 01:59:36PM -0400, Neil Horman wrote:
> On Wed, Sep 17, 2014 at 03:35:19PM +0000, Richardson, Bruce wrote:
> > 
> > > -----Original Message-----
> > > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > > Sent: Wednesday, September 17, 2014 4:21 PM
> > > To: Richardson, Bruce
> > > Cc: dev@dpdk.org
> > > Subject: Re: [dpdk-dev] [PATCH 2/5] ixgbe: add prefetch to improve slow-path tx
> > > perf
> > > 
> > > On Wed, Sep 17, 2014 at 11:01:39AM +0100, Bruce Richardson wrote:
> > > > Make a small improvement to slow path TX performance by adding in a
> > > > prefetch for the second mbuf cache line.
> > > > Also move assignment of l2/l3 length values only when needed.
> > > >
> > > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > > > ---
> > > >  lib/librte_pmd_ixgbe/ixgbe_rxtx.c | 12 +++++++-----
> > > >  1 file changed, 7 insertions(+), 5 deletions(-)
> > > >
> > > > diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > index 6f702b3..c0bb49f 100644
> > > > --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > @@ -565,25 +565,26 @@ ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf
> > > **tx_pkts,
> > > >  		ixgbe_xmit_cleanup(txq);
> > > >  	}
> > > >
> > > > +	rte_prefetch0(&txe->mbuf->pool);
> > > > +
> > > 
> > > Can you explain what all of these prefetches are doing?  It looks to me like
> > > they're just fetching the first caheline of the mempool structure, which it
> > > appears amounts to the pools name.  I don't see that having any use here.
> > > 
> > This does make a decent enough performance difference in my tests (the amount varies depending on the RX path being used by testpmd). 
> > 
> > What I've done with the prefetches is two-fold:
> > 1) changed it from prefetching the mbuf (first cache line) to prefetching the mbuf pool pointer (second cache line) so that when we go to access the pool pointer to free transmitted mbufs we don't get a cache miss. When clearing the ring and freeing mbufs, the pool pointer is the only mbuf field used, so we don't need that first cache line.
> ok, this makes some sense, but you're not guaranteed to either have that
> prefetch be needed, nor are you certain it will still be in cache by the time
> you get to the free call.  Seems like it might be preferable to prefecth the
> data pointed to by tx_pkt, as you're sure to use that every loop iteration.

The vast majority of the times the prefetch is necessary, and it does help 
performance doing things this way. If the prefetch is not necessary, it's 
just one extra instruction, while, if it is needed, having the prefetch 
occur 20 cycles before access (picking an arbitrary value) means that we 
have cut down the time it takes to pull the data from cache when it is 
needed by 20 cycles. As for the value pointed to by tx_pkt, since this is a 
packet the app has just been working on, it's almost certainly already in 
l1/l2 cache.  

> 
> > 2) changed the code to prefetch earlier - in effect to prefetch one mbuf ahead. The original code prefetched the mbuf to be freed as soon as it started processing the mbuf to replace it. Instead now, every time we calculate what the next mbuf position is going to be we prefetch the mbuf in that position (i.e. the mbuf pool pointer we are going to free the mbuf to), even while we are still updating the previous mbuf slot on the ring. This gives the prefetch much more time to resolve and get the data we need in the cache before we need it.
> > 
> Again, early isn't necessecarily better, as it just means more time for the data
> in cache to get victimized. It seems like it would be better to prefetch the
> tx_pkts data a few cache lines ahead.
> 
> Neil

Basically it all comes down to measured performance - working with 
prefetches is not an exactly science, sadly. I've just re-run a quick sanity 
test on this patch in the sequence. Running with testpmd on a single core, 
40G of small packet input, I see considerable performance increases. What 
I've run is:
* testpmd with a single forwarding core, defaults - which means slow path RX 
+ slow path TX (i.e. this code): Performance with this patch increases by 
almost 8%
* testpmd with a single forwarding core, defaults + rxfreet=32 - which means 
vector RX path +  slow path TX (again, this code path): Performance 
increases by over 18%.

Given these numbers, the prefetching seems better this way.  Perhaps you 
could run some tests yourself and see if you see a similar performance delta 
(or perhaps there are other scenarios I'm missing here)?

Regards,
/Bruce
  

On Thu, Sep 18, 2014 at 02:36:13PM +0100, Bruce Richardson wrote:
> On Wed, Sep 17, 2014 at 01:59:36PM -0400, Neil Horman wrote:
> > On Wed, Sep 17, 2014 at 03:35:19PM +0000, Richardson, Bruce wrote:
> > > 
> > > > -----Original Message-----
> > > > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > > > Sent: Wednesday, September 17, 2014 4:21 PM
> > > > To: Richardson, Bruce
> > > > Cc: dev@dpdk.org
> > > > Subject: Re: [dpdk-dev] [PATCH 2/5] ixgbe: add prefetch to improve slow-path tx
> > > > perf
> > > > 
> > > > On Wed, Sep 17, 2014 at 11:01:39AM +0100, Bruce Richardson wrote:
> > > > > Make a small improvement to slow path TX performance by adding in a
> > > > > prefetch for the second mbuf cache line.
> > > > > Also move assignment of l2/l3 length values only when needed.
> > > > >
> > > > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > > > > ---
> > > > >  lib/librte_pmd_ixgbe/ixgbe_rxtx.c | 12 +++++++-----
> > > > >  1 file changed, 7 insertions(+), 5 deletions(-)
> > > > >
> > > > > diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > index 6f702b3..c0bb49f 100644
> > > > > --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > @@ -565,25 +565,26 @@ ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf
> > > > **tx_pkts,
> > > > >  		ixgbe_xmit_cleanup(txq);
> > > > >  	}
> > > > >
> > > > > +	rte_prefetch0(&txe->mbuf->pool);
> > > > > +
> > > > 
> > > > Can you explain what all of these prefetches are doing?  It looks to me like
> > > > they're just fetching the first caheline of the mempool structure, which it
> > > > appears amounts to the pools name.  I don't see that having any use here.
> > > > 
> > > This does make a decent enough performance difference in my tests (the amount varies depending on the RX path being used by testpmd). 
> > > 
> > > What I've done with the prefetches is two-fold:
> > > 1) changed it from prefetching the mbuf (first cache line) to prefetching the mbuf pool pointer (second cache line) so that when we go to access the pool pointer to free transmitted mbufs we don't get a cache miss. When clearing the ring and freeing mbufs, the pool pointer is the only mbuf field used, so we don't need that first cache line.
> > ok, this makes some sense, but you're not guaranteed to either have that
> > prefetch be needed, nor are you certain it will still be in cache by the time
> > you get to the free call.  Seems like it might be preferable to prefecth the
> > data pointed to by tx_pkt, as you're sure to use that every loop iteration.
> 
> The vast majority of the times the prefetch is necessary, and it does help 
> performance doing things this way. If the prefetch is not necessary, it's 
> just one extra instruction, while, if it is needed, having the prefetch 
> occur 20 cycles before access (picking an arbitrary value) means that we 
> have cut down the time it takes to pull the data from cache when it is 
> needed by 20 cycles.
I understand how prefetch works. What I'm concerned about is its overuse, and
its tendency to frequently need re-calibration (though I admit I missed the &
operator in the patch, and thought you were prefetching the contents of the
struct, not the pointer value itself).  As you say, if the pool pointer is
almost certain to be used, then it may well make sense to prefetch the data, but
in doing so, you potentially evict something that you were about to use, so
you're not doing yourself any favors.  I understand that you've validated this
experimentally, and so it works, right now.  I just like to be very careful
about how prefetch happens, as it can easily (and sliently) start hurting far
more than it helps.

> As for the value pointed to by tx_pkt, since this is a 
> packet the app has just been working on, it's almost certainly already in 
> l1/l2 cache.  
> 
Not sure I follow you here.  tx_pkts is an array of mbufs passed to the pmd from
rte_eth_tx_burts, which in turn is called by the application.  I don't see any
reasonable guarantee that any of those packets have been touch in sufficiently
recent history that they are likely to be in cache.  It seems like, if you do
want to do prefetching, interrotagting nb_tx and doing a prefetch of an
approriate stride to fill multiple cachelines with successive mbuf headers might
provide superior performance.
Neil
  

On Thu, Sep 18, 2014 at 11:29:30AM -0400, Neil Horman wrote:
> On Thu, Sep 18, 2014 at 02:36:13PM +0100, Bruce Richardson wrote:
> > On Wed, Sep 17, 2014 at 01:59:36PM -0400, Neil Horman wrote:
> > > On Wed, Sep 17, 2014 at 03:35:19PM +0000, Richardson, Bruce wrote:
> > > > 
> > > > > -----Original Message-----
> > > > > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > > > > Sent: Wednesday, September 17, 2014 4:21 PM
> > > > > To: Richardson, Bruce
> > > > > Cc: dev@dpdk.org
> > > > > Subject: Re: [dpdk-dev] [PATCH 2/5] ixgbe: add prefetch to improve slow-path tx
> > > > > perf
> > > > > 
> > > > > On Wed, Sep 17, 2014 at 11:01:39AM +0100, Bruce Richardson wrote:
> > > > > > Make a small improvement to slow path TX performance by adding in a
> > > > > > prefetch for the second mbuf cache line.
> > > > > > Also move assignment of l2/l3 length values only when needed.
> > > > > >
> > > > > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > > > > > ---
> > > > > >  lib/librte_pmd_ixgbe/ixgbe_rxtx.c | 12 +++++++-----
> > > > > >  1 file changed, 7 insertions(+), 5 deletions(-)
> > > > > >
> > > > > > diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > index 6f702b3..c0bb49f 100644
> > > > > > --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > @@ -565,25 +565,26 @@ ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf
> > > > > **tx_pkts,
> > > > > >  		ixgbe_xmit_cleanup(txq);
> > > > > >  	}
> > > > > >
> > > > > > +	rte_prefetch0(&txe->mbuf->pool);
> > > > > > +
> > > > > 
> > > > > Can you explain what all of these prefetches are doing?  It looks to me like
> > > > > they're just fetching the first caheline of the mempool structure, which it
> > > > > appears amounts to the pools name.  I don't see that having any use here.
> > > > > 
> > > > This does make a decent enough performance difference in my tests (the amount varies depending on the RX path being used by testpmd). 
> > > > 
> > > > What I've done with the prefetches is two-fold:
> > > > 1) changed it from prefetching the mbuf (first cache line) to prefetching the mbuf pool pointer (second cache line) so that when we go to access the pool pointer to free transmitted mbufs we don't get a cache miss. When clearing the ring and freeing mbufs, the pool pointer is the only mbuf field used, so we don't need that first cache line.
> > > ok, this makes some sense, but you're not guaranteed to either have that
> > > prefetch be needed, nor are you certain it will still be in cache by the time
> > > you get to the free call.  Seems like it might be preferable to prefecth the
> > > data pointed to by tx_pkt, as you're sure to use that every loop iteration.
> > 
> > The vast majority of the times the prefetch is necessary, and it does help 
> > performance doing things this way. If the prefetch is not necessary, it's 
> > just one extra instruction, while, if it is needed, having the prefetch 
> > occur 20 cycles before access (picking an arbitrary value) means that we 
> > have cut down the time it takes to pull the data from cache when it is 
> > needed by 20 cycles.
> I understand how prefetch works. What I'm concerned about is its overuse, and
> its tendency to frequently need re-calibration (though I admit I missed the &
> operator in the patch, and thought you were prefetching the contents of the
> struct, not the pointer value itself).  As you say, if the pool pointer is
> almost certain to be used, then it may well make sense to prefetch the data, but
> in doing so, you potentially evict something that you were about to use, so
> you're not doing yourself any favors.  I understand that you've validated this
> experimentally, and so it works, right now.  I just like to be very careful
> about how prefetch happens, as it can easily (and sliently) start hurting far
> more than it helps.
> 
> > As for the value pointed to by tx_pkt, since this is a 
> > packet the app has just been working on, it's almost certainly already in 
> > l1/l2 cache.  
> > 
> Not sure I follow you here.  tx_pkts is an array of mbufs passed to the pmd from
> rte_eth_tx_burts, which in turn is called by the application.  I don't see any
> reasonable guarantee that any of those packets have been touch in sufficiently
> recent history that they are likely to be in cache.  It seems like, if you do
> want to do prefetching, interrotagting nb_tx and doing a prefetch of an
> approriate stride to fill multiple cachelines with successive mbuf headers might
> provide superior performance.
> Neil
>
Prefetching the mbuf is probably best left to the application. For all our 
sample applications used for benchmarking, and almost certainly the vast 
majority of all our example applications, the packet being transmitted is 
already in cache on the core itself. Adding a prefetch to the tx function I 
would expect to see a performance decrease in both testpmd and l3fwd apps.  
I would be useful for apps where the packets are passed from one core to 
another core which does no processing of them before transmitting them - but 
in that case, it's better to have the TX thread of the app do the prefetch 
rather than forcing it in the driver and reduce the performance of those 
apps that have the packets already in cache.

The prefetch added by the patch under discussion doesn't suffer from this 
issue as the data being prefetched is for the mbuf that was previously 
transmitted some time previously, and the tx function has fully looped back 
around the TX ring to get to it again.

/Bruce
  

On Thu, Sep 18, 2014 at 04:42:36PM +0100, Bruce Richardson wrote:
> On Thu, Sep 18, 2014 at 11:29:30AM -0400, Neil Horman wrote:
> > On Thu, Sep 18, 2014 at 02:36:13PM +0100, Bruce Richardson wrote:
> > > On Wed, Sep 17, 2014 at 01:59:36PM -0400, Neil Horman wrote:
> > > > On Wed, Sep 17, 2014 at 03:35:19PM +0000, Richardson, Bruce wrote:
> > > > > 
> > > > > > -----Original Message-----
> > > > > > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > > > > > Sent: Wednesday, September 17, 2014 4:21 PM
> > > > > > To: Richardson, Bruce
> > > > > > Cc: dev@dpdk.org
> > > > > > Subject: Re: [dpdk-dev] [PATCH 2/5] ixgbe: add prefetch to improve slow-path tx
> > > > > > perf
> > > > > > 
> > > > > > On Wed, Sep 17, 2014 at 11:01:39AM +0100, Bruce Richardson wrote:
> > > > > > > Make a small improvement to slow path TX performance by adding in a
> > > > > > > prefetch for the second mbuf cache line.
> > > > > > > Also move assignment of l2/l3 length values only when needed.
> > > > > > >
> > > > > > > Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > > > > > > ---
> > > > > > >  lib/librte_pmd_ixgbe/ixgbe_rxtx.c | 12 +++++++-----
> > > > > > >  1 file changed, 7 insertions(+), 5 deletions(-)
> > > > > > >
> > > > > > > diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > > index 6f702b3..c0bb49f 100644
> > > > > > > --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > > +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > > @@ -565,25 +565,26 @@ ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf
> > > > > > **tx_pkts,
> > > > > > >  		ixgbe_xmit_cleanup(txq);
> > > > > > >  	}
> > > > > > >
> > > > > > > +	rte_prefetch0(&txe->mbuf->pool);
> > > > > > > +
> > > > > > 
> > > > > > Can you explain what all of these prefetches are doing?  It looks to me like
> > > > > > they're just fetching the first caheline of the mempool structure, which it
> > > > > > appears amounts to the pools name.  I don't see that having any use here.
> > > > > > 
> > > > > This does make a decent enough performance difference in my tests (the amount varies depending on the RX path being used by testpmd). 
> > > > > 
> > > > > What I've done with the prefetches is two-fold:
> > > > > 1) changed it from prefetching the mbuf (first cache line) to prefetching the mbuf pool pointer (second cache line) so that when we go to access the pool pointer to free transmitted mbufs we don't get a cache miss. When clearing the ring and freeing mbufs, the pool pointer is the only mbuf field used, so we don't need that first cache line.
> > > > ok, this makes some sense, but you're not guaranteed to either have that
> > > > prefetch be needed, nor are you certain it will still be in cache by the time
> > > > you get to the free call.  Seems like it might be preferable to prefecth the
> > > > data pointed to by tx_pkt, as you're sure to use that every loop iteration.
> > > 
> > > The vast majority of the times the prefetch is necessary, and it does help 
> > > performance doing things this way. If the prefetch is not necessary, it's 
> > > just one extra instruction, while, if it is needed, having the prefetch 
> > > occur 20 cycles before access (picking an arbitrary value) means that we 
> > > have cut down the time it takes to pull the data from cache when it is 
> > > needed by 20 cycles.
> > I understand how prefetch works. What I'm concerned about is its overuse, and
> > its tendency to frequently need re-calibration (though I admit I missed the &
> > operator in the patch, and thought you were prefetching the contents of the
> > struct, not the pointer value itself).  As you say, if the pool pointer is
> > almost certain to be used, then it may well make sense to prefetch the data, but
> > in doing so, you potentially evict something that you were about to use, so
> > you're not doing yourself any favors.  I understand that you've validated this
> > experimentally, and so it works, right now.  I just like to be very careful
> > about how prefetch happens, as it can easily (and sliently) start hurting far
> > more than it helps.
> > 
> > > As for the value pointed to by tx_pkt, since this is a 
> > > packet the app has just been working on, it's almost certainly already in 
> > > l1/l2 cache.  
> > > 
> > Not sure I follow you here.  tx_pkts is an array of mbufs passed to the pmd from
> > rte_eth_tx_burts, which in turn is called by the application.  I don't see any
> > reasonable guarantee that any of those packets have been touch in sufficiently
> > recent history that they are likely to be in cache.  It seems like, if you do
> > want to do prefetching, interrotagting nb_tx and doing a prefetch of an
> > approriate stride to fill multiple cachelines with successive mbuf headers might
> > provide superior performance.
> > Neil
> >
> Prefetching the mbuf is probably best left to the application. For all our 
> sample applications used for benchmarking, and almost certainly the vast 
> majority of all our example applications, the packet being transmitted is 
> already in cache on the core itself. Adding a prefetch to the tx function I 
> would expect to see a performance decrease in both testpmd and l3fwd apps.  
> I would be useful for apps where the packets are passed from one core to 
> another core which does no processing of them before transmitting them - but 
> in that case, it's better to have the TX thread of the app do the prefetch 
> rather than forcing it in the driver and reduce the performance of those 
> apps that have the packets already in cache.
> 

Regarding the performance decrease, I think you're trying to have it both ways
here.  Above you indicate that if the prefetch of the pool pointer isn't needed
its just an extra instruction, which I think is true.  But now you are saying
that if the tx buffers are in cache, the extra instructions will have an impact.
Granted its potentially nb_tx prefetches, not one, but none of them stall the
cpu pipeline as far as Im aware, so I can't imagine 1 prefetch vs several will
have a significant impact on performance.

Regarding where to do prefecth.  Leaving prefetch in the hands of the application is a
bad idea, because the application has no visibility into the code path once you
enter the DPDK. It doesn't know if the buffers are going to be accessed in 20
cycles or 20,000 cycles, which will be all the difference between a useful and
harmful prefetch.  Sure you can calibrate your application to correspond to a
given version of the dpdk and optimize such a prefetch, but that will be
completely obsoleted the first time the dpdk transmit path changes.

As for the use of prefetching tx buffers at all, I think theres several cases
where you might find that those buffers are vicimized in cache.  consider the
situation where a receive interrupt triggers on a cpu right before rte_eth_trans
is called.  For a heavily loaded system, the receive buffers may frequently push
the soon-to-be-transmitted buffers out of cache.

> The prefetch added by the patch under discussion doesn't suffer from this 
> issue as the data being prefetched is for the mbuf that was previously 
> transmitted some time previously, and the tx function has fully looped back 
> around the TX ring to get to it again.
> 

I get what you're saying here, that after the first prefetch the data stays hot
in cache because it is continually re-accessed.  Thats fine.  But that would
happen after the first fetch anyway, without the prefetch.

You know what would put this argument to rest?  If you could run whatever
benchmark you were running under the perf utility so we could see the L1 cache
misses from the baseline dpdk, the variant where you prefetch the pool pointer,
and a variant in which you prefetch the next tx buf at the top of the loop.

Neil



> /Bruce
> 
>