| dc.description |
The purpose of this thesis is to explore dependency speculation in Dynamic Simultaneous Multi-Threading (DSMT). DSMT is a microprocessor architecture which attempts to extract Thread Level Parallelism (TLP) from single-threaded programs at run-time. This is accomplished by running multiple iterations of program loops in parallel. The DSMT architecture was originally developed by Dr. Daniel Ortiz-Arroyo and Dr. Ben Lee at Oregon State University.
To extract TLP from loops successfully, inter-thread dependencies must be resolved by either speculation or stalling. To maximize performance both stalling and misspeculation must be minimized. To this end, two techniques are presented which attempt to improve stride speculation and dynamic inter-thread dependency resolution. To study these proposed changes, a detailed, cycle-accurate simulation environment for DSMT with extensive statistics gathering capabilities was developed. Results generated by the simulator not only show the performance of the proposed changes but also the capabilities of the new simulator. |
|