Section: New Results

Topology Aware Performance Monitoring

While system's scale is growing exponentially, memory hierarchy is getting larger, at various levels. Hence optimizing applications to reach an optimal usage of a machine may involve a large spectrum of performance metrics interacting at different level of the system's hierarchy. Memory bound applications showing irregular patterns lead to locality issues. Addressing those issues and getting a good schedule on complex systems is a NP hard problem and can therefore only be solved with heuristics. Although powerful algorithms using the most intuitive heuristics such as communications path reduction and/or cache contention reduction may show good results on some cases, there are still room for improvements in this direction so much the configuration of applications, systems, software stack vary and impact the execution time.

In order to step in this direction we developped a highly extendible tool to gather asynchronously performance data from different sources. This information is then aggregated into different topology objects (cache, node, processing unit, ...) in order to give a synthetic and topology aware information to drive optimization.

In brief the tool works this way: The user provide a description file with arithmetic expression of performance counters(defined into performance data plugins), and topology objects where to map the expression. A pair (expression,object) defines a monitor which will sample performance data and stored them into an history. Then others monitors can be defined as a combination of the previous. For instance we can attach a process and record on each core its L3 cache miss counter, and then add each of those monitor into an upper monitor located on the L3 cache. Several aggregation functions are already available but we aim to provide several statistical function to extend the possibility of data interpretation. Such functions allow to aggregate results in a meaningful way. Then we add a locality insight using lstopo tool from hwloc to draw the results on a topology. This has been published in [12]