Re: java performance on machines
Per Sandberg wrote:
Hi
I would like to do a performance comparison of different machines and
architectures. And would like some suggestions on actual java tests to
program.
Primarily the goal is to compare cpu/memory power on different
operations, integer and float. The premise, is to test sparc agains
opteron, so we need numbers to be able to compare.
The idea is to write a series of tests in java, both integer and float
based, which I then run on the machines to compare their performance. I
need to do the same in C++ as well, since some programs run in Java and
some run in C++, so it might be sensible to reuse some of the test
algorithms in C++ as well.
I was thinking of writing tests based on float operations and on integer
operations.
The integer operations would be tests of typical java tasks:
1- parsing input strings and building objects
2- manipulating a collection
3- matrix multiplication with integers
Other tests I have come up with are (which I dont know if is
1- matrix multiplication with float
2- matrix inversion
3- fibonacci sequence with internal multiplication
and signum test on roedys page (integer or float?)
any other suggestion of representative tests?
regards
Per
If synthetic benchmarks are ok then you could look at the SPEC benchmarks
(www.spec.org). The main benchmark is now CPU2006 which incorporates both CINT
and CFP which are designed to measure integer and floating point performance.
There are already published results for many systems which might save you some
work. If these benchmarks are the same as the previous SPECs then they are
essentially single-threaded, designed to test only a single core. The RATE
tests also indicate how well the system handles multiple simultaneous load by
running several single-threaded benchmarks at the same time. When I was
comparing Opteron vs. Sparc 3 years ago Sparc didn't come out too well,
certainly not in the price/performance stakes.
There is also a new JVM benchmark, SPECjvm2008 but so far there is only 1 result
submitted by Sun. I have no idea what this might test in terms of
cores/threads.
Of course, the only really valid benchmark is the actual code you intend to
run...
--
Nigel Wade