Skip to contentSkip to conversation

skb align patch

RFC In review Stephen Hemminger 5 participants 8 messages Sep 20, 2009 – Sep 25, 2009

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

RFC 1/1 skb align patch

Based on the Intel suggestion that PCI-express overhead is a significant cost.

Would people doing performance please measure the impact of changing SKB alignment (64 bit only).

arch/x86/include/asm/system.h +10 -0
@@ -455,4 +455,14 @@ static inline void rdtsc_barrier(void)
 	alternative(ASM_NOP3, "lfence", X86_FEATURE_LFENCE_RDTSC);
 }
 
+#ifndef CONFIG_X86_32
+/*
+ * DMA to unaligned address is more expensive than the the
+ * overhead of unaligned CPU access.
+ */
+#define NET_IP_ALIGN	0
+#define NET_SKB_PAD	L1_CACHE_BYTES
+#endif
+
+
 #endif /* _ASM_X86_SYSTEM_H */
#1 lore raw
Quoted 5 lines · Stephen Hemminger

Based on the Intel suggestion that PCI-express overhead is a significant cost.

Would people doing performance please measure the impact of changing SKB alignment (64 bit only).

I had this idea some time ago when I hit a limit on bnx2 adapter (Giga bit link, BCM5708S), with small packets. pktgen was able to send ~500 Mbps 'only', or 700kps if I remember well. So I tried to align the pktgen build packet to a cache line, it gave no difference at all, but it was on a 32 bit kernel. (Thus my patch was for pktgen only, not a generic one as yours)

Could you elaborate why this change could be useful on 64bit ?

Thanks

Quoted 20 lines
arch/x86/include/asm/system.h +10 -1
@@ -455,4 +455,14 @@ static inline void rdtsc_barrier(void)
 	alternative(ASM_NOP3, "lfence", X86_FEATURE_LFENCE_RDTSC);
 }
 
+#ifndef CONFIG_X86_32
+/*
+ * DMA to unaligned address is more expensive than the the
+ * overhead of unaligned CPU access.
+ */
+#define NET_IP_ALIGN	0
+#define NET_SKB_PAD	L1_CACHE_BYTES
+#endif
+
+
 #endif /* _ASM_X86_SYSTEM_H */
--
#2 lore raw
Quoted 7 lines · On Mon, 21 Sep 2009 08:13:20 +0200 Eric Dumazet <[email protected]> wrote:
Quoted 5 lines · Stephen Hemminger

Based on the Intel suggestion that PCI-express overhead is a significant cost.

Would people doing performance please measure the impact of changing SKB alignment (64 bit only).

Quoted 9 lines

I had this idea some time ago when I hit a limit on bnx2 adapter (Giga bit link, BCM5708S), with small packets. pktgen was able to send ~500 Mbps 'only', or 700kps if I remember well. So I tried to align the pktgen build packet to a cache line, it gave no difference at all, but it was on a 32 bit kernel. (Thus my patch was for pktgen only, not a generic one as yours)

Could you elaborate why this change could be useful on 64bit ?

It is useful on all architecture where unaligned CPU access is relatively cheap.

The issue is that a unaligned DMA requires a read/modify/write cache line access versus just a write access. I am not a bus expert, but writes are probably more pipelined as well.

#3 lore raw
Quoted 19 lines · Stephen Hemminger
Quoted 6 lines · On Mon, 21 Sep 2009 08:13:20 +0200 Eric Dumazet <[email protected]> wrote:
Quoted 5 lines · Stephen Hemminger

Based on the Intel suggestion that PCI-express overhead is a significant cost.

Would people doing performance please measure the impact of changing SKB alignment (64 bit only).

Quoted 9 lines

I had this idea some time ago when I hit a limit on bnx2 adapter (Giga bit link, BCM5708S), with small packets. pktgen was able to send ~500 Mbps 'only', or 700kps if I remember well. So I tried to align the pktgen build packet to a cache line, it gave no difference at all, but it was on a 32 bit kernel. (Thus my patch was for pktgen only, not a generic one as yours)

Could you elaborate why this change could be useful on 64bit ?

Quoted 7 lines

It is useful on all architecture where unaligned CPU access is relatively cheap.

The issue is that a unaligned DMA requires a read/modify/write cache line access versus just a write access. I am not a bus expert, but writes are probably more pipelined as well.

Oh I see, you want to optimize the rx (NIC has to do a DMA to write packet into host memory and this DMA could be a read /modify/write if address is not aligned, instead of a pure write),
while I tried to align skb to optimize the pktgen tx
(NIC has to do a DMA to read packet from host), and align the skb had no effect.

Maybe we should separate the rx/tx, and try your idea only for skb allocated for rx.

Also/Or we might try
__builtin_prefetch (addr, 0, 0);
to instruct cpu to commit to memory cache lines that are going to be modified by NIC.

#4 lore raw
Quoted 35 lines · On Tue, 22 Sep 2009 05:20:53 +0200 Eric Dumazet <[email protected]> wrote:
Quoted 19 lines · Stephen Hemminger
Quoted 6 lines · On Mon, 21 Sep 2009 08:13:20 +0200 Eric Dumazet <[email protected]> wrote:
Quoted 5 lines · Stephen Hemminger

Based on the Intel suggestion that PCI-express overhead is a significant cost.

Would people doing performance please measure the impact of changing SKB alignment (64 bit only).

Quoted 9 lines

I had this idea some time ago when I hit a limit on bnx2 adapter (Giga bit link, BCM5708S), with small packets. pktgen was able to send ~500 Mbps 'only', or 700kps if I remember well. So I tried to align the pktgen build packet to a cache line, it gave no difference at all, but it was on a 32 bit kernel. (Thus my patch was for pktgen only, not a generic one as yours)

Could you elaborate why this change could be useful on 64bit ?

Quoted 7 lines

It is useful on all architecture where unaligned CPU access is relatively cheap.

The issue is that a unaligned DMA requires a read/modify/write cache line access versus just a write access. I am not a bus expert, but writes are probably more pipelined as well.

Oh I see, you want to optimize the rx (NIC has to do a DMA to write packet into host memory and this DMA could be a read /modify/write if address is not aligned, instead of a pure write),
while I tried to align skb to optimize the pktgen tx
(NIC has to do a DMA to read packet from host), and align the skb had no effect.

Quoted 7 lines

Maybe we should separate the rx/tx, and try your idea only for skb allocated for rx.

Also/Or we might try
__builtin_prefetch (addr, 0, 0);
to instruct cpu to commit to memory cache lines that are going to be modified by NIC.

Don't think it matters whether RX buffer has to read/modify/write from cpu cache or memory on modern cache snooping architecures. The cost is the PCI traffic.

maintainer
#5 lore raw
In reply to Eric Dumazet

From: Eric Dumazet <[email protected]>
Date: Tue, 22 Sep 2009 05:20:53 +0200

Quoted 5 lines

Oh I see, you want to optimize the rx (NIC has to do a DMA to write packet into host memory and this DMA could be a read /modify/write if address is not aligned, instead of a pure write), while I tried to align skb to optimize the pktgen tx (NIC has to do a DMA to read packet from host), and align the skb had no effect.

This is a problem with these kinds of changes.

This patch from Stephen came out of a presentation and discussion at netconf where the Intel folks showed that if they did a combination of things it improved NUMA forwarding numbers a lot.

So you couldn't just do NUMA spreading of RX queue memory, or just do this ALIGN patch, or just eliminate the false sharing from statistics updates.

You had to do all three to start seeing forwarding rates go up.

So don't worry, this is getting us somewhere to where improvement shows, but individually each change won't trigger it.

The alignment in this patch is a real big deal for 64 byte forwarding tests, where the entire packet is a whole PCI-E cacheline. But not if it isn't aligned properly.

#6 lore raw
Quoted 3 lines · David Miller

The alignment in this patch is a real big deal for 64 byte forwarding tests, where the entire packet is a whole PCI-E cacheline. But not if it isn't aligned properly.

As I pointed out to Herbert already, this alignment change may actually make things worse or even break things as long as compare_ether_header() used in __napi_gro_receive() expects the IP header to be aligned to 4 bytes. That can be fixed of course, just wanted to mention it.

#7 lore raw
Quoted 5 lines · Thomas Graf

As I pointed out to Herbert already, this alignment change may actually make things worse or even break things as long as compare_ether_header() used in __napi_gro_receive() expects the IP header to be aligned to 4 bytes. That can be fixed of course, just wanted to mention it.

Well this particular patch doesn't because x86-64 is fine with unaligned access :)

But yeah we need to fix up GRO to not do that for other arches. It's on my list.

Thanks,

Signature · 4 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt