Re: java performance on machines
Per Sandberg wrote:
Nigel Wade wrote:
Of course, the only really valid benchmark is the actual code you intend to
run...
Thats one aspect of it, but the code base is so big, and unknown to me,
that I cant spend time searching the code for good pieces to extract and
use in the test.
Normally you would use the entire application. After all, that's what you'll be
doing on the real system. To be realistic the benchmark should represent your
normal workload as closely as possible. Pure CPU performance on bits of code
won't tell you the whole story. The bottleneck might turn out to be the network
or disk I/O and your high performance CPUs would end up sitting idle.
However, it's not usual for you to be able to actually run benchmarks on a real
system. Unless you are spending large sums manufacturers or systems integrators
won't normally loan you a system or provide access to theirs. Typically you
have to rely on synthetic benchmarks such as those produced by SPEC.
So I will add a couple of tests which covers usual
programming techniques and some little bit more synthetic tests, such as
matrix manipulation etc.
Is there any point in developing your own synthetic benchmark when SPEC have
developed one which has evolved over 20 years? There are results for many
system configurations on the SPEC website.
--
Nigel Wade