Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/rouming/ccont
Tool burns CPUs on different NUMA nodes and measures execution time
https://github.com/rouming/ccont
Last synced: 5 days ago
JSON representation
Tool burns CPUs on different NUMA nodes and measures execution time
- Host: GitHub
- URL: https://github.com/rouming/ccont
- Owner: rouming
- Created: 2016-03-21T12:09:03.000Z (over 8 years ago)
- Default Branch: master
- Last Pushed: 2024-06-19T11:13:11.000Z (5 months ago)
- Last Synced: 2024-06-19T21:20:29.022Z (5 months ago)
- Language: C
- Size: 2.93 KB
- Stars: 1
- Watchers: 3
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README
Awesome Lists containing this project
README
ccont: Tool burns CPUs on different NUMA nodes and measures execution time.
Description:
The goal is to measure cache contention on different NUMA nodes,
burn different CPUs, execute different instructions with different
load patterns, e.g. the following is the list of three load patterns
which were executed on machine with 2 NUMA nodes and 8 CPUs:
o cpu-increase - on each iteration number of CPU is increased:
# ./ccont --load cpu-increase --op cmpxchg
Nodes N0 N1 CPUs operation min max avg stdev
CPUs *--- ---- 1 cmpxchg 8.938 8.938 8.938 0.000
CPUs **-- ---- 2 cmpxchg 36.114 36.119 36.117 0.004
CPUs ***- ---- 3 cmpxchg 54.270 54.272 54.271 0.001
CPUs **** ---- 4 cmpxchg 72.292 72.321 72.313 0.013
CPUs **** *--- 5 cmpxchg 61.691 108.060 98.782 20.735
CPUs **** **-- 6 cmpxchg 101.316 136.923 125.059 18.369
CPUs **** ***- 7 cmpxchg 151.639 169.218 161.702 9.358
CPUs **** **** 8 cmpxchg 192.281 196.250 194.281 2.098
o node-cascade - on each iteration CPUs from each node are burned:
# ./ccont --load node-cascade --op cmpxchg
Nodes N0 N1 CPUs operation min max avg stdev
CPUs **** ---- 4 cmpxchg 72.287 72.322 72.310 0.016
CPUs ---- **** 4 cmpxchg 72.327 72.333 72.330 0.003
o cpu-rollover - on each iteration executor thread rolls to another CPU on
the next node, keeping the same amount of CPUs burning:
# ./ccont --load cpu-rollover --op cmpxcgh
Nodes N0 N1 CPUs operation min max avg stdev
CPUs **** ---- 4 cmpxchg 48.769 48.774 48.772 0.002
CPUs ***- *--- 4 cmpxchg 85.506 97.754 94.683 6.118
CPUs **-- **-- 4 cmpxchg 116.803 121.450 119.108 2.658
CPUs *--- ***- 4 cmpxchg 91.312 103.877 100.721 6.273
CPUs ---- **** 4 cmpxchg 48.288 48.368 48.323 0.038
Memory chunk for each load is always allocated on the node#0.
Results show, that scattered tasks over NUMA nodes show bad performance for
cmpxchg instruction (cpu-rollover pattern), but execution on remote node
is not so bad, because of the L3 cache (node-cascade pattern). Increase of
the CPUs number can degrade performance by factor of 24 because of the cache
line contention (cpu-increase pattern).
The following burning operations are supported:
o "idle" - idle loop:
used just for calibrating.
while (spins--)
;
o "memset64" - memset glibc call:
memsets 64 bytes (usual cache line size).
o "memset128" - memset glibc call:
memsets 128 bytes.
o "memset256" - memset glibc call:
memsets 256 bytes.
o "test_bit" - btl:
testing a bit, used for test_bit() in Linux kernel.
var | (1 << bit)
o "set_bit" - bts:
test and set bit, used for test_and_set_bit() in Linux kernel.
"test_bit" - name in test results.
res = var | (1 << bit)
var |= (1 << bit)
o "inc" - lock inc:
increment, used for atomic_inc() in Linux kernel.
var += 1
o "xadd" - lock xadd:
exchanges operands, used for __sync_fetch_and_add()
and similar gcc atomic builtins.
tmp = src + dst;
src = dst;
dst = tmp;
o "cmpxchg" - lock cmpxchg:
exchanges operangs, used for cmpxchg() for all sorts of atomic
exchanges in Linux kernel.
res = var
if (res == old)
var = new
o "mfence" - mfence:
memory barrier for load and store, used for smp_mb() in Linux kernel.
o "sfence" - sfence:
memory barrier for store, used for smp_wmb() in Linux kernel.
o "lfence" - lfence:
memory barrier for load, used for smp_rmb() in Linux kernel.