Characterization and Linear Decomposition of Benchmark Workloads


Lev Trubov

Oral Defence Date: 

Friday, May 12, 2017 - 14:30


TH 434


Profs. Jozo Dujmovic & Bill Hsu


Although vast amounts of work have been done in development of benchmark suites for evaluation of computer systems, reliable performance prediction of arbitrary programs is elusive. This project attempts to ll this gap by applying decomposition and predicting performance of programs in terms of weighted compositions of other programs whose performance is known. It rst tests the ability to make such predictions in the rst place by comparing direct results of timing programs (black-box timing based decomposition), and then presents a technique to actually create good decomposition from the dynamic distribution of LLVM IR instructions in the program (glass-box analysis). It demonstrates the failure of both aspects, and explains the inability of accurate performance prediction by the inability to generate decompositions close to their targets in program space


Workload Characterization, Linear Decomposition, Benchmarking, In- struction Mix, LLVM IR


Lev Trubov