Memory Organisation Cartography & Analysis
David Beniamine, Youenn Corre, Damien Dosimont, Guillaume Huard · HAL (Le Centre pour la Communication Scientifique Directe) · 2015
Although performance analysis is one of the most important phase of High Perfor-mance Computing application development, analysis tools are complex to use. Most of the time,they rely on performance counters which are hard to understand for the final user. Therefore onlyfew advanced users are able to do an efficient and precise analysis.Moreover these counters focus on the processor while most of the performance issues are due to abad memory usage.In this report, we present MOCA a new kind of analysis tool which provides an overview of thememory access over time. We present MOCA's trace visualization through Ocelotl, a tool designedto provided an aggregated view of data while loosing as few information as possible. Finally, weexplain how using these tools, the user can identify memory usage patterns, execution phases andhow to interpret them.