We provide a detailed description of the parts of the toolkit we believe are relevant to users of the HPCx system. The descriptions are tailored to the HPCx system, including the path names in order to make using the toolkit as easy as possible. We include various hints on the difficulties we faced when we started using this tool. The document has four main sections.
To read the counters for the entire application one can use the tool HPMCOUNT. HPMCOUNT is very easy to set up and does not require any modification of the source code. Its usage is even possible if you have access to an executable only. The disadvantages of using a global tool are also obvious. The output does not provide any hint on which part of the application contains the bottlenecks and all the setup and initialisation routines are included in the measurements.
To overcome these problems the code can be instrumented using LIBHPM. Obviously this requires access to the source to include the required calls. With LIBHPM it is possible to measure the performance of individual code segments.
In the following section, which is relevant to HPMCOUNT and LIBHPM, we give a comprehensive description of the different counters and metrics. This commented listing of the different numbers and their meaning was one of our key motivations for writing this document.
In the last section we discuss the visualisation tool HPMVIZ, which is useful to analyse the output of LIBHPM.
This document describes version 2.5.2 of the Hardware Performance Monitor Toolkit.