Performance Evaluation of a Microarchitecture With Multiple Flows of Control
Resumo
This work presents a new approach to exploit the Instruction-Level Parallelism. In the MULFLUX microarchitecture each branch instruction spawns two different flows, corresponding to the two possible branch paths. Instructions from both flows are executed speculatively in parallel. Cycles consumed by discarded instructions are not visible, because the correct instructions also have been executed. The concept of multiple flows of control as exploited in this research work has provided promising results, with performance gains of up to 109% for a configuration supporting up to 16 aclive flows. Among the configurations considered here, the configuration with a maximum of 4 aclive flows exhibited the best compromise between performance and resource replication: it provided a performance gain of 105% and full dispatch width utilization of up to 41%.
Referências
Patterson, D., Hennessy, J., Computer Architecture: A Quantitative Approach, 2nd. Ed., Morgan-Kaufmann, San Francisco, CA, 1997.
Lam, M. S., R. P. Wilson, Limits of Control Flow on Parallelism, Proc. 19th International Symposium on Computer Architecture, 1992, pp. 46-57.
Wallace, S., Brad Calder, Threaded Multiple Path Execution. Proc. of the 25th International Symposium on Computer Architecture, 1998, pp. 238-249.
Tullsen, D. M., S. J. Eggers, Simultaneous Multithreading: Maximizing On-Chip Parallelism, Proc. of the 22nd International Symposium on Computer Architecture, 1995, pp. 392-403.
Klauser, A., A. Paithankar, Selective Eager Execution on the PolyPath Architecture, Proc. of the 25nd International Symposium on Computer Architecture, 1998, pp. 250-259.
Mulflux: Superscalar Architectures with Multiple Instructions Flows. Research Project Proposal submitted to the Brazilian National Research Council, 1995.
MIPS Technologies Inc., MIPS R10000 Microprocessor User's Manual, Mountain View, CA., 1995.
MULFLUX: A Superscalar Architecture with Multiple Flows of Control, Proceedings of the ProTem-CC III Project Evaluation Workshop, May 1999. (also available at http://www.cos.ufrj.br/mulflux)
Sun Microsystems, Tire SPARC Architecture Manual, Mountain View, CA. 1987.
IBM Corp., PowerPC 604 RISC Microprocessor Technical Summary, IBM Order Number MPR604TSU-01.
Diep, T. A., J. P. Shen, Performance Evaluation of the PowerPC 620 Microarchitecture, Proc. of the 22nd International Symposium on Computer Architecture, 1995, pp. 163-175.
