In this project, we try to compare usage of class member variables to compacted arrays for primitive/simply-assigned mutable state as it may appear in classes generated by Spark. We consider two classes and two variants that each perform a set of trivial operations involving setting and accessing variable values: a purely inline-variable based class, and a purely array-based class, each holding 4 ints and 4 Strings, and each with a "large" variant, holding 5000 ints and 5000 Strings.
Compare the .jasm bytecode generated with javap for the VarObject and ArrObject classes. We note that the array access operations during setting and access each take about 2-3 more assembly instructions than setting and accessing inlined variables. However, directly comparing the VarObjectLarge and ArrObjectLarge bytecode files shows that the added declaration and initialization of inlined variables goes on to dominate bytecode generation, whereas the arrays are declared and initialized very concisely. The final length of the VarObjectLarge bytecode file is 80,052 lines, and the ArrObjectLarge bytecode file is 50,121 lines. The init method for the ArrObjectLarge class is represented in some 30 instructions, while the init method for the VarObjectLarge class requires some 30,000 instructions.
Benchmarking the code in milliseconds also seems to show the array based variable storage approach executes at least an order of magnitude more quickly than the inlined-variable object. An example execution of the main method:
Variable execution time: 2587
Array execution time: 56
We may conclude that differences in overhead between the two approaches is negligible, or perhaps even that for classes of a many variables, compaction in arrays is more efficient than inlining all variables.