An Allocation Strategy Using Shadow-processors and Simulation Technique

• Efficient performance tuning of parallel programs for multiprocessors is often hard. When it comes to assigning threads to processors there is not much support from commercial operating systems, like the Solaris operating system. The only known value is, in best case, the total execution time of each thread. The developer is left to the bin packing algorithm with no knowledge about the interactions and dependencies between the threads. The bin packing algorithm assigns, in the worst case, the threads to the processors such that the program will have the longest possible execution time. A simple example of such a program is shown. We present here a way of retrieving more information and a test mechanism that makes it possible to compare two different assignments of threads on processors also with regard to the interactions and dependencies between the threads. Also an algorithm is proposed that gives the best assignment of threads to processors in the case above where the bin packing algorithm gave the worst possible assignment. The algorithm uses shadow-processors and requires more processors than on the target machine during some allocation steps. Thus, a simulation tool like the one presented here must be used.

Ämnesord

NATURVETENSKAP  -- Data- och informationsvetenskap -- Datavetenskap (hsv//swe)

NATURAL SCIENCES  -- Computer and Information Sciences -- Computer Sciences (hsv//eng)

Nyckelord

bin packing

multi-threading

multiprocessing systems

parallel programming

resource allocation

software performance evaluation

virtual machines

Publikations- och innehållstyp

ref (ämneskategori)

kon (ämneskategori)