Skip to contentSkip to conversation

SMP even/odd cores latency

DISCUSSION DIXLOR 2 participants 2 messages Feb 3, 2016

lore
Share & tools
lore

DIXLOR

author
#0 lore raw
# ./cyclictest -sm -n -p99 --policy=fifo -S -D60
# /dev/cpu_dma_latency set to 0us
policy: fifo: loadavg: 2.89 1.37 0.64 1/506 15232

T: 0 (15179) P:99 I:1000 C: 59999 Min: 14 Act: 56 Avg: 51 Max: 99
T: 1 (15180) P:99 I:1500 C: 39999 Min: 3 Act: 13 Avg: 29 Max: 100
T: 2 (15181) P:99 I:2000 C: 29999 Min: 12 Act: 50 Avg: 48 Max: 101
T: 3 (15182) P:99 I:2500 C: 23999 Min: 4 Act: 13 Avg: 29 Max: 101
T: 4 (15183) P:99 I:3000 C: 19999 Min: 16 Act: 60 Avg: 51 Max: 104
T: 5 (15184) P:99 I:3500 C: 17142 Min: 4 Act: 60 Avg: 30 Max: 102
T: 6 (15185) P:99 I:4000 C: 14999 Min: 14 Act: 29 Avg: 51 Max: 106
T: 7 (15186) P:99 I:4500 C: 13333 Min: 4 Act: 11 Avg: 32 Max: 106
T: 8 (15187) P:99 I:5000 C: 11999 Min: 17 Act: 42 Avg: 47 Max: 90
T: 9 (15188) P:99 I:5500 C: 10908 Min: 5 Act: 40 Avg: 29 Max: 89

T:10 (15189) P:99 I:6000 C:   9999 Min:     18 Act:   43 Avg:   49 Max:      99
T:11 (15190) P:99 I:6500 C:   9230 Min:      5 Act:   28 Avg:   29 Max:      99
T:12 (15191) P:99 I:7000 C:   8571 Min:     14 Act:   56 Avg:   47 Max:      98
T:13 (15192) P:99 I:7500 C:   7999 Min:      5 Act:   11 Avg:   31 Max:      97
T:14 (15193) P:99 I:8000 C:   7499 Min:     16 Act:   64 Avg:   49 Max:     100
T:15 (15194) P:99 I:8500 C:   7058 Min:      5 Act:   42 Avg:   30 Max:      96

Minimum latency on even cores is approximately equal to 15, on the odd - 5 Аverage latency on the even - 50, on the odd cores - 30 :)

- Linux 4.4.1-rt4+ x86_64
- gcc version 4.9.2 (Debian 4.9.2-10)
- 2 x AMD Opteron(tm) Processor 4386
- Mb: TYAN S8225

#1 lore raw
Quoted 8 lines · DIXLOR
Quoted 1 line

Minimum latency on even cores is approximately equal to 15, on the

odd - 5
Аverage latency on the even - 50, on the odd cores - 30 :)

Quoted 1 line

- Linux 4.4.1-rt4+ x86_64

- gcc version 4.9.2 (Debian 4.9.2-10)
- 2 x AMD Opteron(tm) Processor 4386
- Mb: TYAN S8225
You could confirm by checking /proc/cpuinfo, but it looks like the AMD

cpu enumeration matches that of the last few Intel iterations, where the even cores are one NUMA node, and the odd the other.