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
|
* Add more test cases. Categories we'd like to cover (with reasonably
real-world tests, preferably not microbenchmarks) include:
(X marks the ones that are fairly well covered now).
X math (general)
X bitops
X 3-d (the math bits)
- crypto / encoding
X string processing
- regexps
- date processing
- array processing
- control flow
- function calls / recursion
- object access (unclear if it is possible to make a realistic
benchmark that isolates this)
I'd specifically like to add all the computer language shootout
tests that Mozilla is using.
* Normalize tests. Most of the test cases available have a repeat
count of some sort, so the time they take can be tuned. The tests
should be tuned so that each category contributes about the same
total, and so each test in each category contributes about the same
amount. The question is, what implementation should be the baseline?
My current thought is to either pick some specific browser on a
specific platform (IE 7 or Firefox 2 perhaps), or try to target the
average that some set of same-generation release browsers get on
each test. The latter is more work. IE7 is probably a reasonable
normalization target since it is the latest version of the most
popular browser, so results on this benchmark will tell you how much
you have to gain or lose by using a different browser.
* Instead of using the standard error, the correct way to calculate
a 95% confidence interval for a small sample is the t-test.
<http://en.wikipedia.org/wiki/Student%27s_t-test>. Basically this involves
using values from a 2-tailed t-distribution table instead of 1.96 to
multiply by the error function, a table is available at
<http://www.medcalc.be/manual/t-distribution.php>
* Add support to compare two different engines (or two builds of the
same engine) interleaved.
* Add support to compare two existing sets of saved results.
* Allow repeat count to be controlled from the browser-hosted version
and the WebKitTools wrapper script.
* Add support to run only a subset of the tests (both command-line and
web versions).
* Add a profile mode for the command-line version that runs the tests
repeatedly in the same command-line interpreter instance, for ease
of profiling.
* Make the browser-hosted version prettier, both in general design and
maybe using bar graphs for the output.
* Make it possible to track change over time and generate a graph per
result showing result and error bar for each version.
* Hook up to automated testing / buildbot infrastructure.
* Possibly... add the ability to download iBench from its original
server, pull out the JS test content, preprocess it, and add it as a
category to the benchmark.
* Profit.
|
