25% gain - This sounds cool !! Any chance you can send the patch, so
that I can try it out in scaled up setup?

It would be awesome if quad-loop gets applied to the most common set of
graph nodes involved in IP forwarding, as you say.

I also get the point that it is not always obvious which VPP nodes gain
from having the quad-loop, but with my limited understanding of VPP code
and CPU prefetch and load-store operations it’s hard for me to say which
ones are go and which ones are no-go nodes for quad-loop code.

Dave, from your note I glean that ip4-input-no-checksum and
ip4-rewrite-transit are the obvious nodes to go after. Same I read would
apply to ip6 equivalents. What about L2? Others?


On 17 Oct 2016, at 18:35, Dave Barach (dbarach) 
<dbar...@cisco.com<mailto:dbar...@cisco.com>> wrote:


I finally decided to quad-loop ip4-lookup, just to see what would happen. 
Various folks have suggested giving it a try, etc.

Note the 12 clocks/pkt improvement on a single core, with an offered load of 
around ~15mpps. See below for CPU details. That’s approximately a 25% 
performance gain. Quad-looping [only] ip4-lookup increases the ip4 racetrack 
forwarding NDR by something like 800 kpps.

It’s not worth retrofitting every node - even though the work is mechanical - 
but the ip speed path nodes are obvious candidates.

In case enquiring minds want to know... D.

vpp# sh ru
Time 3.8, average vectors/node 251.55, last 128 main loops 12.00 per node 256.00
  vector rates in 1.4877e7, out 1.4877e7, drop 0.0000e0, punt 0.0000e0
             Name              Clocks       Vectors/Call
FortyGigabitEthernet84/0/1-out      7.56e0          251.55
FortyGigabitEthernet84/0/1-tx       4.79e1          251.55
dpdk-input                          5.85e1          251.55
ip4-input-no-checksum               2.94e1          251.55
ip4-lookup                          4.54e1          251.55
ip4-rewrite-transit                 2.42e1          251.55

vpp# sh ru
Time 3.2, average vectors/node 149.42, last 128 main loops 3.09 per node 66.00
  vector rates in 1.5601e7, out 1.5601e7, drop 0.0000e0, punt 0.0000e0
             Name              Clocks       Vectors/Call
FortyGigabitEthernet84/0/1-out      7.68e0          149.42
FortyGigabitEthernet84/0/1-tx       4.74e1          149.42
dpdk-input                          5.96e1          149.42
ip4-input-no-checksum               3.04e1          149.42
ip4-lookup                          3.34e1          149.42
ip4-rewrite-transit                 2.41e1          149.42

Thanks… Dave

$ cat /proc/cpuinfo

processor       : 31
vendor_id       : GenuineIntel
cpu family      : 6
model           : 63
model name      : Intel(R) Xeon(R) CPU E5-2667 v3 @ 3.20GHz
stepping        : 2
microcode       : 0x27
cpu MHz         : 1201.000
cache size      : 20480 KB
physical id     : 1
siblings        : 16
core id         : 7
cpu cores       : 8
apicid          : 31
initial apicid  : 31
fpu             : yes
fpu_exception   : yes
cpuid level     : 15
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov 
pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb 
rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology 
nonstop_tsc aperfmperf eagerfpu pni pclmulqdq dtes64 monitor ds_cpl vmx smx est 
tm2 ssse3 fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt 
tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm ida arat epb xsaveopt 
pln pts dtherm tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 
avx2 smep bmi2 erms invpcid
bogomips        : 6401.16
clflush size    : 64
cache_alignment : 64
address sizes   : 46 bits physical, 48 bits virtual
power management:
vpp-dev mailing list

vpp-dev mailing list

Reply via email to