StackThreads/MP: version 0.77 User's Guide

11.8: Why is my program slow on a single processor?

Q: When I run a program compiled by stgcc on a single processor, it is slower than its serial program compiled by gcc. Why?

A: Although StackThreads/MP tries to minimize the overhead of running parallel programs, there are several sources.

• ST_THREAD_CREATE adds something like ten instructions in addition to a normal procedure call.
• ST_POLLING costs something like ten instructions.
• Making your program runnable on multiprocessors, you must perform synchronization (join_counter, mutex, etc.). They roughly cost a procedure call even when they always succeed.

These overheads are insignificant when thread and synchronization granularity are large enough (several hundreds instructions between two events are large enough). Moreover, these overheads are fairly easy to reason about, because they all appear in the source program. There are other sources that are not directly visible in C/C++ sources.

• stgcc by default prohibits inlining. This may slowdown performance of C++ programs considerably. See Possible Slowdown by Disabling Inline for details and workarounds.
• stgcc prohibits omitting frame pointers that are commonly adopted on Mips and Alpha. This may add a few instructions to a procedure call overhead.
• By inserting a few instructions at the end of each procedure, each procedure call now costs a few extra instructions at the epilogue.

The last two items are not noticeable unless the program calls a very small function (that perform, say, only 10 instructions in its body) too often. Such procedures are most likely to be inline-expanded if you attach inline keyword. This eliminates the procedure call overhead all together.