[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Comparing the performance of the index-mapping implementations

This page is part of the web mail archives of SRFI 25 from before July 7th, 2015. The new archives for SRFI 25 contain all messages, not just those from before July 7th, 2015.

Here is a comparison of the performance of the three index-mapping
schemes provided with the reference implementation of SRFI-25.
All tests were done on a PII/350Mhz with 256M RAM under Linux 2.2.17.
Implementations tested were Gambit-C 3.0 and Chicken 1024.
The array-representation used is `as-procedure.scm'. The test itself
consists just of 

cat [gambit-prelude.scm] as-procedure.scm array.scm op-XXXX.scm \
  ix-XXXX.scm arlib.scm test.scm list.scm >bench.scm


  (declare (standard-bindings) (extended-bindings) (fixnum) (not interrupts-enabled))
  -O3 -fomit-frame-pointer -D___SINGLE_HOST

  -optimize-level 2 -fixnum-arithmetic -disable-interrupts
  -O3 -fomit-frame-pointer -fstrict-aliasing

  (declare (standard-bindings) (extended-bindings) (fixnum) (not interrupts-enabled) (not safe) (block))
  -O3 -fomit-frame-pointer -D___SINGLE_HOST

  -O3 -fomit-frame-pointer -fstrict-aliasing

             Gambit   Chicken   Gambit(unsafe)  Chicken(unsafe)

ctor          1.60     0.46        0.29            0.24
mbda          1.44     0.51        0.36            0.32
tter          1.43     0.50        0.33            0.28

(I tried Stalin and mzc also, but they both failed either while
compiling (Stalin 0.9) or when executing (Mzc 103))

It seems that with all optimizations the ctor indexing is
the fastest - most of the work done in `op-ctor.scm' involves
primitive vector- and math-operations that can be optimized
nicely by the compiler if generic arithmetic is switched off.

Of course much of the code can be tuned to a specific implementation,
and this will change the timings accordingly. This is just an attempt
to find a general idea about performance of the reference implementation(s).
(Multiple values under Gambit have been simulated with simple lists, so
there is a lot of room for improvement).

All files (makefile, SRFI-25 ref.impl. with some minor bugs fixed
and preludes for Gambit) can be obtained at