Newer
Older
Apart from the micro-benchmark functionality this is also a blueprint for other micro-benchmark applications.
It contains C modules for:
* Aligned data allocation
* Query and control affinity settings
* Accurate timing
Moreover the benchmark showcases a simple generic Makefile that can be used in other projects.
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
## Overview
The benchmark is heavily inspired by John McCalpin's https://www.cs.virginia.edu/stream/ benchmark.
It contains the following streaming kernels with corresponding data access pattern (Notation: S - store, L - load, WA - write allocate). All variables are vectors, s is a scalar:
* init (S1, WA): Initilize an array: `a = s`. Store only.
* sum (L1): Vector reduction: `s += a`. Load only.
* copy (L1, S1, WA): Classic memcopy: `a = b`.
* update (L1, S1): Update vector: `a = a * scalar`. Also load + store but without write allocate.
* triad (L2, S1, WA): Stream triad: `a = b + c * scalar`.
* daxpy (L2, S1): Daxpy: `a = a + b * scalar`.
* striad (L3, S1, WA): Schoenauer triad: `a = b + c * d`.
* sdaxpy (L3, S1): Schoenauer triad without write allocate: `a = a + b * c`.
As added benefit the code is a blueprint for a minimal benchmarking application with a generic makefile and modules for aligned array allocation, accurate timing and affinity settings. Those components can be used standalone in your own project.
## Build
1. Configure the toolchain and additional options in `config.mk`:
```
# Supported: GCC, CLANG, ICC
TAG ?= GCC
ENABLE_OPENMP ?= false
ENABLE_LIKWID ?= false
OPTIONS = -DSIZE=40000000ull
OPTIONS += -DNTIMES=10
OPTIONS += -DARRAY_ALIGNMENT=64
#OPTIONS += -DVERBOSE_AFFINITY
#OPTIONS += -DVERBOSE_DATASIZE
#OPTIONS += -DVERBOSE_TIMER
```
The verbosity options enable detailed output about affinity settings, allocation sizes and timer resolution.
2. Build with:
```
make
```
You can build multiple toolchains in the same directory, but notice that the Makefile is only acting on the one currently set. Intermediate build results are located in the `<TOOLCHAIN>` directory.
To output the executed commands use:
```
make Q=
```
3. Clean up with:
```
make clean
```
to clean intermediate build results.
```
make distclean
```
to clean intermediate build results and binary.
4. (Optional) Generate assembler:
```
make asm
```
The assembler files will also be located in the `<TOOLCHAIN>` directory.
## Usage
To run the benchmark call:
```
./bwBench-<TOOLCHAIN>
```
The benchmark will output the results similar to the stream benchmark. Results are validated.
For threaded execution it is recommended to control thread affinity.
We recommend to use likwid-pin for benchmarking:
```
likwid-pin -c 0-3 ./bwbench-GCC
```
Example output for threaded execution:
```
-------------------------------------------------------------
[pthread wrapper]
[pthread wrapper] MAIN -> 0
[pthread wrapper] PIN_MASK: 0->1 1->2 2->3
[pthread wrapper] SKIP MASK: 0x0
threadid 140271463495424 -> core 1 - OK
threadid 140271455102720 -> core 2 - OK
threadid 140271446710016 -> core 3 - OK
OpenMP enabled, running with 4 threads
----------------------------------------------------------------------------
Function Rate(MB/s) Rate(MFlop/s) Avg time Min time Max time
Init: 22111.53 - 0.0148 0.0145 0.0165
Sum: 46808.59 46808.59 0.0077 0.0068 0.0140
Copy: 30983.06 - 0.0207 0.0207 0.0208
Update: 43778.69 21889.34 0.0147 0.0146 0.0148
Triad: 34476.64 22984.43 0.0282 0.0278 0.0305
Daxpy: 45908.82 30605.88 0.0214 0.0209 0.0242
STriad: 37502.37 18751.18 0.0349 0.0341 0.0388
SDaxpy: 46822.63 23411.32 0.0281 0.0273 0.0325
----------------------------------------------------------------------------
Solution Validates
```
## Benchmarking skript
A perl wrapper script (bench.pl) is also provided to scan ranges of thread counts and determine the absolute highest sustained main memory bandwidth. In order to use it `likwid-pin` has to be in your path. The script has three required and one optional command line arguments:
```
$./bench.pl <executable> <thread count range> <repetitions> [<SMT setting>]
```
Example usage:
```
$./bench.pl ./bwbench-GCC 2-8 6
```
The script will always use physical cores only, where two SMT threads is the default. For different SMT thread counts use the 4th command line argument. Example for a processor without SMT:
```
$./bench.pl ./bwbench-GCC 14-24 10 1
```