|Home | People | Curriculum | Projects | Resources | Media|
This work differs from other work on improving cache performance for regular scientific programs because of its focus on finding and exploiting scalable locality (locality that can be made to grow with the size of the problem). This means that our transformation will provide greater and greater speedups if microprocessor speed continues to grow faster than memory speed. We have demonstrated that time skewing can be beneficial for current architectures, and that it can compensate for machines in which non-local references are extremely expensive (for example, we have obtained speedups ranging from a factor of 27 to a factor of 490 for programs that use data sets too large to fit in RAM). For more details, see the time skewing web page.
I have also developed a version of time skewing for multiprocessors, which was published at IPDPS '00