DEE Simulation Data

Here is data on Disjoint Eager Execution: An Optimal Form of Speculative Execution . A summary of the data is also in: Branch Effect Reduction Techniques.

The marketing slide:

The same data, in
detailed form:

The vertical axis is a model's speedup factor over sequential execution, e.g., "30" = x30 = an improvement in performance of 3,000%. The horizontal axis is the number of resources allowed for the model, in terms of branch paths (a branch path is the dynamic code between two branches; its resources are the functional units necessary to fully exploit the ILP [Instruction Level Parallelism] in the path; this is very roughly equivalent to about 1-2 functional units per path).

Methodology:

SPECint92 benchmarks w/out sc
up to 100 million instructions simulated per benchmark
speedups shown are harmonic means over the five benchmarks' individual speedups
all instructions execute in one cycle

The hardware models simulated are combinations of the following:

"DEE" is Disjoint Eager Execution.
"SP" is Single Path speculative execution, the kind of speculation commonly used in today's microprocessors.
"EE" is Eager Execution.
"CD-MF" is equivalent to Minimal Control Dependencies (MCD); see: A Theory of Reduced and Minimal Procedural Dependencies . Branches execute in parallel.
"CD" is MCD, but branches must execute sequentially.

April 25, 1997 | Gus Uht | uht@ele.uri.edu