589833 - Vector slower than array in "some" test

Status: RESOLVED WORKSFORME

(Tamarin Graveyard :: Virtual Machine, defect, P3)

Description

[Chris Peyer]

~5% slower in mixed/jit mode, ~3% slower in interp:

mixed:
some	77.1	72.8	-5.6%
interp:
some   176.4	171.7	-2.7%
jit:
some    76.4	73.1	-4.3%

Test is located in
asteam/hg/qe/projects/performance/tests/language/vector/some.abc

Comment 1

[Steven Johnson]

Hmm, this is very peculiar; Vector uses the same implementation as Array, so the difference is presumably in helper functions or memory overhead

Comment 2

[Dan Smith]

Lars, transferring from Steven to you for consideration in the Vector speedup work.

Comment 3

[Lars T Hansen]

The test case is arguably invalid as it stands.  It compares the case of a Vector.<int> with an Array.  The latter holds atoms, the former holds ints.  The some() method takes a callback whose API accepts atoms.  Thus there will be a song and dance on every invocation of the callback to convert a native int argument to the atom accepted by the method.  This adds significant overhead.

With a current Release 32-bit Mac build on MacPro 2x2.93GHz Xeon, we have with the benchmark as written using Vector.<int>:

vector some        : avg: 62.10 dev: 1.66 values: 64,62,63,60,60,65,61,63,62,61
array some         : avg: 55.90 dev: 0.32 values: 56,56,55,56,56,56,56,56,56,56

and with the Vector case modified ot use Vector.<*> instead of Vector.<int>:

vector some        : avg: 57.80 dev: 0.63 values: 58,57,59,58,57,58,58,57,58,58
array some         : avg: 56.00 dev: 0.47 values: 56,55,56,57,56,56,56,56,56,56

There's some variability here, the array case is still faster on most runs so there's a small slowdown somewhere, but the difference is smaller than initially thought.

Comment 4

[Lars T Hansen]

On a MacBook Pro, 2.6GHz Core 2 Duo (2008 model), the results are reversed with the modified benchmark, this run is typical:

vector some        : avg: 82.60 dev: 0.97 values: 85,83,83,82,82,83,82,82,82,82
array some         : avg: 86.10 dev: 0.88 values: 84,86,86,86,87,86,86,87,87,86

ie, the rest is noise.