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
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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
|
[comment {-*- tcl -*- doctools manpage}]
[manpage_begin bench_lang_intro n 1.0]
[see_also bench_intro]
[see_also bench_lang_spec]
[keywords {bench language}]
[keywords benchmark]
[keywords examples]
[keywords performance]
[keywords testing]
[copyright {2007 Andreas Kupries <andreas_kupries@users.sourceforge.net>}]
[moddesc {Benchmarking/Performance tools}]
[titledesc {bench language introduction}]
[category {Benchmark tools}]
[description]
[para]
This document is an informal introduction to version 1 of the bench
language based on a multitude of examples. After reading this a
benchmark writer should be ready to understand the formal
[term {bench language specification}].
[subsection Fundamentals]
In the broadest terms possible the [term {bench language}] is
essentially Tcl, plus a number of commands to support the declaration
of benchmarks.
A document written in this language is a Tcl script and has the same
syntax.
[para]
[subsection {Basics}]
One of the most simplest benchmarks which can be written in bench is
[example_begin]
bench -desc LABEL -body {
set a b
}
[example_end]
This code declares a benchmark named [const LABEL] which measures the
time it takes to assign a value to a variable. The Tcl code doing this
assignment is the [option -body] of the benchmark.
[subsection {Pre- and postprocessing}]
Our next example demonstrates how to declare [term initialization] and
[term cleanup] code, i.e. code computing information for the use of
the [option -body], and for releasing such resources after the
measurement is done.
They are the [option -pre]- and the [option -post]-body, respectively.
[para]
In our example, directly drawn from the benchmark suite of Tcllib's
[package aes] package, the concrete initialization code constructs the
key schedule used by the encryption command whose speed we measure,
and the cleanup code releases any resources bound to that schedule.
[example_begin]
bench -desc "AES-${len} ECB encryption core" [option -pre] {
set key [lb]aes::Init ecb $k $i[rb]
} -body {
aes::Encrypt $key $p
} [option -post] {
aes::Final $key
}
[example_end]
[subsection {Advanced pre- and postprocessing}]
Our last example again deals with initialization and cleanup code. To
see the difference to the regular initialization and cleanup discussed
in the last section it is necessary to know a bit more about how bench
actually measures the speed of the the [option -body].
[para]
Instead of running the [option -body] just once the system actually
executes the [option -body] several hundred times and then returns the
average of the found execution times. This is done to remove
environmental effects like machine load from the result as much as
possible, with outliers canceling each other out in the average.
[para]
The drawback of doing things this way is that when we measure
operations which are not idempotent we will most likely not measure
the time for the operation we want, but of the state(s) the system is
in after the first iteration, a mixture of things we have no interest
in.
[para]
Should we wish, for example, to measure the time it takes to include
an element into a set, with the element not yet in the set, and the
set having specific properties like being a shared Tcl_Obj, then the
first iteration will measure the time for this. [emph However] all
subsequent iterations will measure the time to include an element
which is already in the set, and the Tcl_Obj holding the set will not
be shared anymore either. In the end the timings taken for the several
hundred iterations of this state will overwhelm the time taken from
the first iteration, the only one which actually measured what we
wanted.
[para]
The advanced initialization and cleanup codes, [option -ipre]- and the
[option -ipost]-body respectively, are present to solve this very
problem. While the regular initialization and cleanup codes are
executed before and after the whole series of iterations the advanced
codes are executed before and after each iteration of the body,
without being measured themselves. This allows them to bring the
system into the exact state the body wishes to measure.
[para]
Our example, directly drawn from the benchmark suite of Tcllib's
[package struct::set] package, is for exactly the example we used
above to demonstrate the necessity for the advanced initialization and
cleanup. Its concrete initialization code constructs a variable
refering to a set with specific properties (The set has a string
representation, which is shared) affecting the speed of the inclusion
command, and the cleanup code releases the temporary variables created
by this initialization.
[example_begin]
bench -desc "set include, missing <SC> x$times $n" [option -ipre] {
set A $sx($times,$n)
set B $A
} -body {
struct::set include A x
} [option -ipost] {
unset A B
}
[example_end]
[section {FURTHER READING}]
Now that this document has been digested the reader, assumed to be a
[term writer] of benchmarks, he should be fortified enough to be able
to understand the formal [term {bench language specfication}]. It will
also serve as the detailed specification and cheat sheet for all
available commands and their syntax.
[para]
[vset CATEGORY bench]
[include ../doctools2base/include/feedback.inc]
[manpage_end]
|