PerfExpert and MACPO: Which code segments should (not) be ported to MIC?

Accelerators like Graphics Processing Units (GPUs) or coprocessors like Intel’s MIC (Many Integrated Core) provide means to exploit large-scale SIMT/SIMD parallelism in applications. Tools for converting CPU code to code for accelerators/coprocessors are available. Application developers could quickly exploit these accelerators/coprocessors with modest effort if they could identify the code segments that are suited for effective and efficient execution on these architectures without algorithmic changes. However, such an analysis requires knowledge of models of parallelism and execution behavior of the code segments on multicore CPUs. In this work, we present a semi-automated, tool-supported process of identifying the code segments that will run efficiently on accelerator chips such as GPUs and coprocessors such as Intel MICs. Results on production codes demonstrate reliable separation between code segments that can be directly converted to execute efficiently on the MIC versus those that cannot. Keywords-performance; accelerators; GPU; MIC