Toward the Mixing of Monitoring and Profiling Operations on HPC Platforms
Résumé
System and job monitoring are two established way of measuring the utilization of HPC systems. Due to the scale and complexity of modern HPC systems, users also require to profile their application using the actual system. However, the usage of many tools usually requires more effort in support and can also lead to significant under-performance. We introduce Colmet, an existing system modified to combine monitoring and profiling operations. Users can dynamically reconfigure Colmet at runtime and set a different sampling period for each metric. We present the results of the performance analysis conducted.
Origine : Fichiers produits par l'(les) auteur(s)