2.6.9 Use skb_padto() in drivers/net/8390.c]
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 8 lines
From: Alan Cox <[email protected]>
This was done because it benchmarked materially faster than the
skb_padtoversion you just reverted. Its only 64 bytes on the stack and its cached.
Quoted 10 lines
ie the old 8390 code was quite intentional and done because it commonly occurs on very old machines where clock count matters. Because of its commonness I actually hand optimised this one when I did the original fixes to avoid doing extra memory allocations.
Summary: Please revert.
Alan
Is it possible that skb_padto has since got its act together? Reason I
ask is that I just dusted off a crusty 386dx40 (doesn't get much older than that) with a wd8013. As a basic test, I did ttcp Tx tests with small
packets and they came out to all intents and purposes, identical. Kernel
was 2.6.10, with stack vs skb_padto, each size test ran 3 times, even tested packets bigger than ETH_ZLEN as a (hopefully) invariant. I've attached the edited down results below.
Paul.
--- Stock 2.6.10 kernel with 8390 doing the padding ---
ttcp-t: buflen=1, nbuf=32768, align=16384/0, port=5001 tcp -> g 1: ttcp-t: 32768 bytes in 32.81 real seconds = 0.98 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.79
ttcp-t: 3.5user 29.2sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
2: ttcp-t: 32768 bytes in 32.80 real seconds = 0.98 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.99
ttcp-t: 3.7user 29.0sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
3: ttcp-t: 32768 bytes in 32.80 real seconds = 0.98 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.02, calls/sec = 999.10
ttcp-t: 0.2user 32.5sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
ttcp-t: buflen=10, nbuf=32768, align=16384/0, port=5001 tcp -> g 1: ttcp-t: 327680 bytes in 32.80 real seconds = 9.76 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.02, calls/sec = 999.06
ttcp-t: 1.3user 31.4sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
2: ttcp-t: 327680 bytes in 32.80 real seconds = 9.76 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 999.02
ttcp-t: 1.5user 31.2sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
3: ttcp-t: 327680 bytes in 32.80 real seconds = 9.76 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.02, calls/sec = 999.07
ttcp-t: 2.3user 30.3sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
ttcp-t: buflen=40, nbuf=32768, align=16384/0, port=5001 tcp -> g 1: ttcp-t: 1310720 bytes in 32.80 real seconds = 39.02 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.96
ttcp-t: 14.2user 18.5sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
2: ttcp-t: 1310720 bytes in 32.82 real seconds = 39.00 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.37
ttcp-t: 14.6user 18.1sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+24csw
3: ttcp-t: 1310720 bytes in 32.80 real seconds = 39.02 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.91
ttcp-t: 15.1user 17.5sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
ttcp-t: buflen=100, nbuf=32768, align=16384/0, port=5001 tcp -> g 1: ttcp-t: 3276800 bytes in 32.86 real seconds = 97.37 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 997.05
ttcp-t: 0.1user 32.7sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
2: ttcp-t: 3276800 bytes in 32.93 real seconds = 97.19 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 995.22
ttcp-t: 0.1user 32.7sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+24csw
3: ttcp-t: 3276800 bytes in 32.83 real seconds = 97.47 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.12
ttcp-t: 0.0user 32.7sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+22csw
--- 2.6.10 Kernel with 8390 calling kernel's skb_padto ---
ttcp-t: buflen=1, nbuf=32768, align=16384/0, port=5001 tcp -> g 1: ttcp-t: 32768 bytes in 32.81 real seconds = 0.98 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.68
ttcp-t: 1.5user 31.2sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
2: ttcp-t: 32768 bytes in 32.79 real seconds = 0.98 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.02, calls/sec = 999.36
ttcp-t: 4.3user 28.3sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+24csw
3: ttcp-t: 32768 bytes in 32.80 real seconds = 0.98 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.99
ttcp-t: 9.3user 23.3sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+24csw
ttcp-t: buflen=10, nbuf=32768, align=16384/0, port=5001 tcp -> g 1: ttcp-t: 327680 bytes in 32.80 real seconds = 9.76 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.02, calls/sec = 999.14
ttcp-t: 12.5user 20.2sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
2: ttcp-t: 327680 bytes in 32.78 real seconds = 9.76 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.02, calls/sec = 999.52
ttcp-t: 1.2user 31.5sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
3: ttcp-t: 327680 bytes in 32.78 real seconds = 9.76 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.02, calls/sec = 999.60
ttcp-t: 5.6user 27.1sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+24csw
ttcp-t: buflen=40, nbuf=32768, align=16384/0, port=5001 tcp -> g 1: ttcp-t: 1310720 bytes in 32.80 real seconds = 39.03 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.02, calls/sec = 999.09
ttcp-t: 17.3user 15.4sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+22csw
2: ttcp-t: 1310720 bytes in 32.80 real seconds = 39.02 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 999.00
ttcp-t: 18.2user 14.5sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+24csw
3: ttcp-t: 1310720 bytes in 32.80 real seconds = 39.03 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.02, calls/sec = 999.06
ttcp-t: 17.8user 14.9sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+22csw
ttcp-t: buflen=100, nbuf=32768, align=16384/0, port=5001 tcp -> g 1: ttcp-t: 3276800 bytes in 32.81 real seconds = 97.54 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.77
ttcp-t: 0.0user 32.7sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
2: ttcp-t: 3276800 bytes in 32.82 real seconds = 97.51 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 998.54
ttcp-t: 0.0user 32.7sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
3: ttcp-t: 3276800 bytes in 32.83 real seconds = 97.46 KB/sec +++
ttcp-t: 32768 I/O calls, msec/call = 1.03, calls/sec = 997.99
ttcp-t: 0.0user 32.7sys 0:32real 99% 0i+0d 0maxrss 0+0pf 0+23csw
Quoted 2 lines · Paul Gortmaker
Is it possible that
skb_padtohas since got its act together? Reason I
ask is that I just dusted off a crusty 386dx40 (doesn't get much older
Could be - kmalloc has probably improved but the skbuffs have got far more complex
Quoted 5 lines
than that) with a wd8013. As a basic test, I did ttcp Tx tests with small
packets and they came out to all intents and purposes, identical. Kernel
was 2.6.10, with stack vsskb_padto, each size test ran 3 times, even tested packets bigger thanETH_ZLENas a (hopefully) invariant. I've attached the edited down results below.
What are you testing ? I don't see the relationship between network throughput and efficiency on this device.
Drop it on a pentium or late 486 and use the tsc to compare the two code paths. One is much much more efficienct.
Re: [PATCH] 2.6.9 Use skb_padto() in drivers/net/8390.c
Quoted 18 lines · Alan Cox
Quoted 3 lines · Paul Gortmaker
Is it possible that
skb_padtohas since got its act together? Reason I
ask is that I just dusted off a crusty 386dx40 (doesn't get much olderCould be - kmalloc has probably improved but the skbuffs have got far more complex
Quoted 6 lines
than that) with a wd8013. As a basic test, I did ttcp Tx tests with small
packets and they came out to all intents and purposes, identical. Kernel
was 2.6.10, with stack vsskb_padto, each size test ran 3 times, even tested packets bigger thanETH_ZLENas a (hopefully) invariant. I've attached the edited down results below.
Quoted 2 lines
What are you testing ? I don't see the relationship between network throughput and efficiency on this device.
I was thinking that for a sufficiently slow CPU running at 100% (hence the lowly 386), the throughput would be influenced by how efficiently we can get in, get the Tx set up and get out.
Quoted 2 lines
Drop it on a pentium or late 486 and use the tsc to compare the two code paths. One is much much more efficienct.
Using rdtscl over the area affected by the patch on the two variants for a sample of 234 small packets, I see an average of 4141 for using the
existing stack scratch area, and 4162 for using skb_padto. That is a
difference of about 0.5%, which is significantly less than the typical spread in the samples themselves. To help give a relevant scale, feeding it a larger 1400 byte packet comes in at around 60,000 cycles on this
particular box. Am I being optimistic to see this as good news -- meaning
that there is no longer a need for driver specific padto implementations, whereas it looks like there was back when you did your tests?
Paul.
Re: [PATCH] 2.6.9 Use skb_padto() in drivers/net/8390.c
Quoted 9 lines · Paul Gortmaker
Using rdtscl over the area affected by the patch on the two variants for a sample of 234 small packets, I see an average of 4141 for using the
existing stack scratch area, and 4162 for usingskb_padto. That is a
difference of about 0.5%, which is significantly less than the typical spread in the samples themselves. To help give a relevant scale, feeding it a larger 1400 byte packet comes in at around 60,000 cycles on this
particular box. Am I being optimistic to see this as good news -- meaning
that there is no longer a need for driver specific padto implementations, whereas it looks like there was back when you did your tests?
It means that padto has improved a lot (or the underlying allocators). It also still means the patch makes the code slower and introduces changes that have no benefit into the kernel, so while its good to see its not relevant its still a pointless change.
Re: [PATCH] 2.6.9 Use skb_padto() in drivers/net/8390.c
Quoted 13 lines · Alan Cox
Quoted 9 lines · Paul Gortmaker
Using rdtscl over the area affected by the patch on the two variants for a sample of 234 small packets, I see an average of 4141 for using the
existing stack scratch area, and 4162 for usingskb_padto. That is a
difference of about 0.5%, which is significantly less than the typical spread in the samples themselves. To help give a relevant scale, feeding it a larger 1400 byte packet comes in at around 60,000 cycles on this
particular box. Am I being optimistic to see this as good news -- meaning
that there is no longer a need for driver specific padto implementations, whereas it looks like there was back when you did your tests?
Quoted 4 lines
It means that padto has improved a lot (or the underlying allocators). It also still means the patch makes the code slower and introduces changes that have no benefit into the kernel, so while its good to see its not relevant its still a pointless change.
So the verdict is to revert?
Jeff
Re: [PATCH] 2.6.9 Use skb_padto() in drivers/net/8390.c
Quoted 6 lines · Jeff Garzik
Quoted 4 lines
It means that padto has improved a lot (or the underlying allocators). It also still means the patch makes the code slower and introduces changes that have no benefit into the kernel, so while its good to see its not relevant its still a pointless change.
So the verdict is to revert?
Not sure. The old code is known to work and a fraction faster, the new code makes the driver use the same logic as the others. Having seen the numbers from Paul I personally don't care either way.