Re: arm ldxr/stxr vs cas

Some testing of the effects of backoff on contention.
$ lscpu -e
CPU NODE SOCKET CORE L1d:L1i:L2:L3 ONLINE    MAXMHZ   MINMHZ       MHZ
   0    0      0    0 0:0:0:0          yes 3400.0000 400.0000  799.9840
   1    0      0    1 1:1:1:0          yes 3400.0000 400.0000 2296.7949
   2    0      0    2 2:2:2:0          yes 3400.0000 400.0000  712.2160
   3    0      0    3 3:3:3:0          yes 3400.0000 400.0000 1574.7550
   4    0      0    0 0:0:0:0          yes 3400.0000 400.0000  799.9760
   5    0      0    1 1:1:1:0          yes 3400.0000 400.0000 2600.7561
   6    0      0    2 2:2:2:0          yes 3400.0000 400.0000 2599.2959
   7    0      0    3 3:3:3:0          yes 3400.0000 400.0000  400.0000
so 4 cores, 2 threads / core
The options
  -n #enqueues per producer
  -p #producers
  -c #consumers
  -b backoff if > 0
  -f 0 linear backoff 1 exponential backoff
  -t queue type mpmc,spsc, spmc, mpsc   default mpmc
taskset used to specify which cpus to run on.
There is a lot more stats being outputed but I'm
just showing the overall rate.
$ taskset -c 0-7 test/lfrbtest -n 4000000 -p6 -c2 -f 0 -b 0
   overall rate = 5,988,679.5093 /sec
$ taskset -c 0-3 test/lfrbtest -n 4000000 -p6 -c2 -f 0 -b 0
   overall rate = 7,206,319.9145 /sec
$ taskset -c 0-7 test/lfrbtest -n 4000000 -p6 -c2 -f 1 -b 12
   overall rate = 7,243,427.7786 /sec
$ taskset -c 0-3 test/lfrbtest -n 4000000 -p6 -c2 -f 1 -b 12
   overall rate = 8,722,397.3449 /sec
$ taskset -c 3 test/lfrbtest -n 4000000 -p6 -c2 -f 1 -b 12
   overall rate = 19,950,919.8062 /sec
Some more threads than cpus timings ...
$ taskset -c 0-7 test/lfrbtest -n 1000000 -p24 -c8 -f 1 -b 0
   overall rate = 6,454,866.9162 /sec
$ taskset -c 0-7 test/lfrbtest -n 1000000 -p24 -c8 -f 1 -b 12
   overall rate = 7,350,906.2839 /sec
Some 1 producer 1 consumer timings ... (no contention so backoff setting
has no effect)
$ taskset -c 3,4 test/lfrbtest -n 10000000 -p1 -c1 -f 1 -b 0
   overall rate = 15,329,272.1370 /sec
$ taskset -c 3,4 test/lfrbtest -n 10000000 -p1 -c1 -f 1 -b 12
   overall rate = 15,063,074.6140 /sec
$ taskset -c 3 test/lfrbtest -n 10000000 -p1 -c1 -f 1 -b 0
   overall rate = 19,983,735.6372 /sec
$ taskset -c 3 test/lfrbtest -n 10000000 -p1 -c1 -f 1 -b 12
   overall rate = 19,935,213.3078 /sec
$ taskset -c 3 test/lfrbtest -n 10000000 -p1 -c1 -f 1 -b 12 -x n
   overall rate = 32,921,077.1692 /sec
$ taskset -c 2,3 test/lfrbtest -n 10000000 -p1 -c1 -f 1 -b 12 -t spsc
   overall rate = 38,620,404.6845 /sec
  $ taskset -c 3 test/lfrbtest -n 10000000 -p1 -c1 -f 1 -b 12 -t spsc
    overall rate = 109,891,021.6233 /sec
This is producer and consumer running on same thread so no thread
switching overhead
$ taskset -c 3 test/lfrbtest -n 10000000 -p1 -c1 -f 1 -b 12 -x n -t spsc
   overall rate = 146,941,972.2478 /sec
Joe Seigh