2006-02-01

Comparison of speed of different CPUs and variation of 
client/server JVM and 32 or 64 bit matlab on the
64-bit machine.

NOTE:  don't pay TOO much attention to the means: note
that there can be considerable differences between 
different algoithms' speed on different processors.

Computers:

`box':  
	amd64 3500+,  Venice core, 2.2GHz, 512KB L2
	1G RAM (dualch DDR400)
	linux-2.6.15 (gentoo system)
	
`one':  
	p4, HT, 3.0GHz, 512KB L2
	1G RAM
	linux-2.6.13 (gentoo system)

Using Matlab Version 7.1.0.183 (R14) Service Pack 3, with
variation of JVM and architecture (64-bit amd, 32-bit intel).
Matlab was started with the -nodektop option, in a `konsole'.

Tests are as in the matlab `bench' command in the above 
version.  Only the non-graphical ones are performed.
A parameter can be used to change the size of the data
worked upon.

Results given are means of several (3-5) consecutive runs, and
are not given unless every test had the same time to within
about 2% between runs.

The `server' Java virtual machine is used unless otherwise 
stated.  It is expected that this makes no difference as no
GUI or plotting is performed.

==========================================================
1:  parameter of 1 (same data-size as matlab bench.m)

LU    FFT   ODE   Sparse TOTAL

0.40  0.48  0.20  0.56   1.65   (box, 32-bit)
0.29  0.38  0.14  0.50   1.31   (box, 64-bit)

0.23  0.53  0.32  0.57   1.66   (one, client-jvm)
0.23  0.54  0.32  0.57   1.66   (one)


==========================================================
2:  parameter of 1.15 (1150x1150 matrix for LU, FFT on 
    2^23 instead of 2^20, longer ODE time...)

LU    FFT   ODE   Sparse TOTAL

0.59  6.21  0.24  0.96   8.00   (box, 32-bit)
0.42  6.33  0.17  0.83   7.75   (box, 64-bit)

0.43  8.55  0.37  0.95  10.29   (one)


==========================================================
3:  parameter of 0.9

LU    FFT   ODE   Sparse TOTAL

0.31  0.10  0.18  0.38   0.97   (box, 32-bit)
0.22  0.09  0.13  0.34   0.78   (box, 64-bit)

0.18  0.12  0.29  0.40   0.99   (one)


==========================================================


Conclusions?

Somehow, the 64bit version on the same processor does
have a considerable advantage over the 32bit version, 
of some tens of percent for LU factorisation.

For the benchmark mixture of 4 tests of similar time
consumption (at parameter of 1) the 3.0GHz P4 does
overall about the same as the 2.2GHz AMD64 in 32bit
mode.

The Pentium 4 is rather faster even than the AMD64
in 64bit mode for LU factorisation:  at paramter 
of 1, for example, it's ~25% longer time for the AMD.

The AMD64 is much faster then the Pentium 4 at ODE
solving:  in 64bit mode it's about twice as fast!