Microarchitecture Choices and Tradeoffs for Maximizing Processing Efficiency.

Abstract

This thesis is concerned with hardware approaches for maximizing the number of independent instructions in the execution core and thereby maximizing the processing efficiency for a given amount of compute bandwidth. Compute bandwidth is the number of parallel execution units multiplied by the pipelining of those units in the processor. Keeping those computing elements busy is key to maximize processing efficiency and therefore power efficiency.

While some applications have many independent instructions that can be issued in parallel without inefficiencies due to branch behavior, cache behavior, or instruction dependencies, most applications have limited parallelism and plenty of stalling conditions.

This thesis presents two approaches to this problem, which in combination greatly increases the efficiency of the processor utilization of resources. The first approach addresses the problem of small basic blocks that arise when code has frequent branches. We introduce algorithms and mechanisms to predict multiple branches simultaneously and to fetch multiple non-continuous basic blocks every cycle along a predicted branch path. This makes what was previously an inherently serial process into a parallelized instruction fetch approach. For integer applications, the result is an increase in useful instruction fetch capacity of 40% when two basic blocks are fetched per cycle and 63% for three blocks per cycle. For floating point benchmarks, the associated improvement is 27% and 59%.

The second approach addresses increasing the number of independent instructions to the execution core through simultaneous multi-threading (SMT). We compare to another multithreading approach, Switch-on-Event multithreading, and show that SMT is far superior. Intel Pentium 4 SMT microarchitecture algorithms are analyzed, and we look at the impact of SMT on power efficiency of the Pentium 4 Processor. A new metric, the SMT Energy Benefit is defined. Not only do we show that the SMT Energy Benefit for a given workload with SMT can be quite significant, we also generalize the results and build a model for what other future processors’ SMT Energy Benefit would be. We conclude that while SMT will continue to be an energy-efficient feature, as processors get more energy-efficient in general the relative SMT Energy Benefit may be reduced.