OPTIMIZING THE JAVA VIRTUAL MACHINE INSTRUCTION SET BY ...

80
by simply applying it to a larger problem. As a result, the overall list of specialized
bytecodes can be skewed to be similar to the list for any one benchmark by simply
increasing the problem size of that benchmark, unless steps are taken to ensure that
each benchmark contributes equally to the overall list.
This issue was addressed in the following manner. When the lists of specialized
bytecodes were generated for each benchmark, the frequency of execution was expressed
as a percentage of the total number of bytecodes executed by the benchmark.
Using this representation eliminated the opportunity for the distribution to be modified
by increasing the size of the problem solved by the benchmark because increasing
the problem size increases both the frequency with which each bytecode executed
and the total number of bytecodes executed by the application. Consequently, the
percentage value reported for each bytecode will not change as a result of increasing
the size of the problem being solved, successfully preventing one benchmark from
skewing the overall list of specialized bytecodes when the amount of work performed
is changed.
Once the normalized list of specialized bytecodes is generated for each benchmark,
the overall list of specialized bytecodes is generated. The frequency of each specialized
bytecode is the sum of the normalized execution frequency for each benchmark,
divided by the total number of benchmarks considered in this study. Thus the value
presented as the frequency of execution in the overall list is not the percentage of the
total bytecodes executed by all of the benchmarks. Instead, it is the average percentage
of the bytecodes executed across all benchmarks. The result of this profiling
effort is shown in Table 5.1.
5.2 Execution Environments
A total of 8 distinct hardware architecture, operating system and virtual machine
combinations were tested in this study. These include virtual machines developed by
both Sun and IBM, operating systems including SunOS, Linux and Windows running
on both Sun SPARC and Intel processors. Testing such a variety of platforms ensures
that any results observed will be useful across many of the platforms on which Java
applications are executed rather than presenting results that are an artifact of a
particular platform. The specific combinations tested are described more thoroughly
in Table 5.2.
It is worth noting that the virtual machines considered in this portion of the study
also cover three different techniques for executing Java applications despite being