Prof. Mahmut Taylan Kandemir
The research at Penn State focuses on application modernization targeting emerging many-core systems. In particular, we investigate custom application mapping and optimization strategies that take into account unique architectural features of Intel® Xeon Phi™ coprocessors. Our research spans code and data optimizations (both application level and compiler level), latency hiding techniques such as prefetching, studying the tradeoffs between native and offload operation modes when running HPC applications, application restructuring techniques that target not only cache misses but also miss latencies, and investigation of novel architectural features that can be critical for future manycore systems (beyond Intel Xeon Phi coprocessors).
Our research will lead to:
- Identifying required steps for extracting maximum performance from Intel Xeon Phi coprocessors during application mapping and optimization.
- Understanding whether conventional compiler optimizations (as well as our new optimizations) are effective when targeting large HPC codes, and if not, why.
- Finding additional problems when mapping applications to a system with multiple Intel Xeon Phi coprocessors.
- Comparing our results against those that will be collected using available commercial GPU systems.
- Diana Guttman, Mahmut Taylan Kandemir, Meenakshi Arunachalam, and Vlad Calina, 3/30/2015, "Performance and Energy Evaluation of Data Prefetching on Intel Xeon Phi", Performance Analysis of Systems and Software (ISPASS), 2015 IEEE International (Conference)