下載電子全文宣告This thesis is authorized to indicate not accessible
You can not download at the moment.
Your IP address is 220.127.116.11
The defense date of the thesis is 2009-08-26
The current date is 2019-03-23
URN etd-0821109-012912 Statistics This thesis had been viewed 1444 times. Download 4 times. Author Sheng-Zhe Huang Author's Email Address No Public. Department Computer Science and Enginerring Year 2008 Semester 2 Degree Master Type of Document Master's Thesis Language English Page Count 51 Title Speculative Thread-Level Parallelism For Multi-Core System Based On Stride Value Prediction Keyword data dependence value predictor thread parallism value predict value predict thread parallism value predictor data dependence Abstract Computer can greatly improve their computational performance by exploiting inherent parallelism. In order to improve parallelism, we present a new approach termed value prediction for improving the parallelism of a sequential program at run-time. The new approach suggests predicting outcome values of true-data dependent operations before they are executed. As a result, the processor can extract instruction-level parallelism beyond the limits of the program's dataflow graph and speculatively execute the true-data dependent operations in parallel. We revealed new patterns to predict the outcome values of instructions at run-time.
We propose a new modified stride value predictor, which attempts to predict value without waiting for value when data dependence occurs. To decrease the number of misprediction, the highly predictable instructions are determined with program traces statically. According to experiment result, we find that the new predictor improves from 4% to 29% in spec95 benchmark.
Advisor Committee Chia-Ying Tseng - advisor
Liang-Teh Lee - co-chair
Trong-Yen Lee - co-chair
Files Date of Defense 2009-06-29 Date of Submission 2009-08-26