1
2
3
4
5
6
7
8
9
10
11
12
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
|
.TH membench 1 "February 9 2026"
.SH NAME
membench \- modern memory bandwidth and latency benchmarks
.SH SYNOPSIS
.B membench
.RI [ options ]
.br
.SH DESCRIPTION
This manual page documents briefly the
.B membench
command.
.PP
\fBmembench\fP is a portable, multi-platform memory bandwidth benchmark designed for
comprehensive system analysis.
.br
Latency measurement uses linked list traversal with random node order
to defeat prefetchers. Statistical validity ensured via multiple samples
until coefficient of variation < 5% or max samples reached.
.TP
Memory model: each thread gets its own buffer.
.TP
Total memory = size_kb × threads (×2 for copy: src + dst).
.SH OPTIONS
A summary of options is included below.
For a complete description, see the Info files.
.TP
.B \-h
Show this help
.TP
.B \-V
Print version and exit
.TP
.B \-v
Verbose output (use -vv for more detail)
.TP
.B \-s SIZE_KB
Test only this buffer size (in KB), e.g. -s 1024 for 1MB
.TP
.B \-f
Full sweep (test all sizes up to memory limit)
.br
Default: test up to 512 MB per thread
.TP
.B \-p THREADS
Use exactly this many threads (default: num_cpus)
.TP
.B \-a
Auto-scaling: try different thread counts to find best
.br
(slower but finds optimal thread count per buffer size)
.TP
.B \-t SECONDS
Maximum runtime, 0 = unlimited (default: unlimited)
.TP
.B \-r TRIES
Repeat each test N times, report best (default: 3)
.TP
.B \-o OP
Run only this operation: read, write, copy, or latency
.br
Can be specified multiple times (default: all)
.TP
.B \-H
Enable huge pages for large buffers (>= 4MB)
.br
Uses THP (no setup needed) or explicit 2MB pages
.br
Automatically skipped for small buffers
.TP
.B \-R
Human-readable output with summary (default: CSV)
.SH ENVIRONMENT VARIABLES
OpenMP Thread Affinity (environment variables):
OMP_PROC_BIND=spread Spread threads across NUMA nodes (default)
OMP_PLACES=cores One thread per physical core
OMP_NUM_THREADS=N Override thread count
.SH CSV OUTPUT
Output: CSV to stdout with columns:
size_kb - Per-thread buffer size (KB)
operation - read, write, copy, or latency
bandwidth_mb_s - Aggregate bandwidth in MB/s (0 for latency)
latency_ns - Median memory latency in ns (0 for bandwidth)
latency_stddev_ns - Latency standard deviation in ns (0 for bandwidth)
latency_samples - Number of samples for latency measurement
threads - Thread count used
iterations - Iterations performed
elapsed_s - Elapsed time in seconds
.SH SEE ALSO
.BR tinymembench (1),
.BR memtester (8).
.br
|
