2020-11-28
numbers is a quick-and-dirty code that prints some of the
"latency numbers every programmer should know" 1,
2, 3, 4.
n.b. it's for linux on x86-64 with c++17 compiler only (gcc 9, 10; clang 10, 11).
$ git clone https://github.com/jaeheum/numbers.git
$ cd numbers
# a quick-and-dirty run:
$ make all && ./build/numbers
g++ -Wall -Wextra -Wpedantic -Ofast --std=c++17 -Iinclude -c src/nanobench.cc -o build/nanobench.o
g++ -Wall -Wextra -Wpedantic -Ofast --std=c++17 -Iinclude -c src/numbers.cc -o build/numbers.o
g++ -o build/numbers build/*.o -lpthread
Warning, results might be unstable:
* CPU governor is 'schedutil' but should be 'performance'
Recommendations
* Use 'pyperf system tune' before benchmarking. See https://github.com/psf/pyperf
| ns/op | op/s | err% | ins/op | cyc/op | IPC | bra/op | miss% | total | benchmark
|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
| 16.50 | 60,604,416.47 | 0.0% | 63.01 | 56.08 | 1.124 | 18.00 | 0.0% | 0.00 | `mutex_access`
| 5,217.33 | 191,668.80 | 0.6% | 12,301.00 | 17,736.67 | 0.694 | 4,097.00 | 0.0% | 0.00 | `L1_random_access`
| 250,328.00 | 3,994.76 | 0.2% | 98,329.00 | 850,748.00 | 0.116 | 32,769.00 | 0.0% | 0.00 | `L2_random_access`
| 52,028,319.00 | 19.22 | 1.5% | 5,898,281.00 | 176,516,236.00 | 0.033 | 1,966,097.00 | 0.0% | 0.60 | `L3_random_access`
| 2,796,066,612.00 | 0.36 | 0.0% | 100,664,178.00 |9,487,319,166.00 | 0.011 | 33,555,290.00 | 0.0% | 2.80 | `memory_random_access`
| 8,813,073.00 | 113.47 | 1.7% | 110,100,499.00 | 29,845,812.00 | 3.689 | 31,457,284.00 | 0.0% | 0.10 | `sorted_memory_branch_mispredictions`
| 57,787,781.00 | 17.30 | 0.1% | 110,100,513.00 | 196,153,140.00 | 0.561 | 31,457,298.00 | 25.0% | 0.64 | `unsorted_memory_branch_mispredictions`
| 28,025,098.00 | 35.68 | 0.1% | 23,985.00 | 94,954,316.00 | 0.000 | 4,964.00 | 0.3% | 0.31 | `memory_copy_1MiB`
| 156,300,394.00 | 6.40 | 0.0% | 4,350.00 | 53,176.00 | 0.082 | 796.00 | 21.6% | 0.16 | `fwrite_1MiB_to_disk`
| 388,215.00 | 2,575.89 | 0.0% | 81,101.00 | 77,384.00 | 1.048 | 18,564.00 | 0.6% | 0.00 | `fseek_from_disk`
| 34,513,692.00 | 28.97 | 0.0% | 67,884.00 | 396,406.00 | 0.171 | 15,174.00 | 2.5% | 0.03 | `fread_1MiB_from_disk`
L1_random_access 1.3 ns 4.3 cycles
L2_random_access 7.6 ns 26.0 cycles
L3_random_access 26.5 ns 89.8 cycles
memory_random_access 83.3 ns 282.7 cycles
branch_miss_penalty 6.2 ns 21.2 cycles
mutex_access 16.5 ns 56.1 cycles
fseek_from_disk 1516.5 ns 302.3 cycles
memory_copy_1MiB 109473.0 ns 370915.3 cycles
fread_1MiB_from_disk 134819.1 ns 1548.5 cycles
fwrite_1MiB_to_disk 610548.4 ns 207.7 cycles
numbers's output comes from simplistic, best-effort, low-cost, fast
measurements.
For a more careful measurement, run the following commands:
$ sudo pip install pyperf # so that 'sudo pyperf system tune' can run
## On Arch linux, pyperf is available as a package
$ ./print-numbers
n.b. Read notes.md for
benchmarking tips and more sophisticated
tools.
notes.md also contains information about internals of numbers.
There are also known issues.
Depending on Linux configuration, numbers may not be able to access
hardware performance counters. In this case, numbers does not print
branch_miss_penalty row and cycles column in the output.
MIT License
numbers benefits from
- code derived from https://github.com/afborchert/pointer-chasing (MIT)
- nanobench for microbenchmarking suite (MIT)
- argh for command line handling (BSD-3)
- a stackflow post on sorted vs unsorted array