Abstract: Variable Instruction Issue for Efficient MIMD Interpretation on SIMD Machines.
Multiple execution threads can be simulated on Single Instruction Multiple
Data (SIMD) machines by interpretation.
The control unit would continuously issue the signals necessary for
interpreting the instruction set. A processing element (PE) would only
participate in the part of the loop which necessary for interpreting its
instruction, while remaining idle in the parts which are not.
In this thesis, a model of the interpretation cycle used in
MIMD interpretation on SIMD machines is formulated. This model is then used to
study the problem of Variable Instruction Issue (VII), where different subsets
of the instruction set being interpreted, rather than the complete
instruction set, are issued in order to enhance the performance of
the interpretation process. The problem of choosing these subsets has been
shown to be NP-Complete. Heuristics that provide near optimal solutions are
developed and analyzed. These heuristics have shown considerable speedup
over non-variable interpreter loops.
Finally, various extensions to the model, and their effects on the analysis
are also presented and studied. These include extensions to handle pipelining
and functional unit reuse.
Back to Nael's homepage , or view