A PAPI Implementation for BlueGene

摘要

The IBM BlueGene/L (BG/L) super-computer holds 3 of the top 10 rankings on the 26th TOP500 list of LINPACK performance. The system is novel in its design in many aspects when compared to other more traditional high-performance computing systems. When developing system libraries as well as when tuning application code for BG/L it is essential to be able to measure the impact of code modifications and algorithmic choices. PAPI is a platform neutral user level library to accommodate programmers' need to access on-chip performance counters. This paper describes the implementation of the low-level kernel interface for hardware performance counter access on BG/L and the accompanying PAPI implementation.