Jules Risse (Storm)

The next STORM seminar will be presented by Jules Risse, PhD student at the Institut Polytechnique de Paris.

Speaker

Jules Risse

When

February, 27th at 10 a.m.

Where

room Grace Hopper 2

Title

Fine grain energy consumption in a task based runtime system

Abstract

The power consumption of supercomputers is and will be a major concern. As a matter of fact, El Capitan, the fastest supercomputer in the world, consumes around 29,5 MW of power. As a consequence, reducing the power consumption of HPC applications is mandatory. A first step is to monitor programs’ power consumption: servers usually contain wattmeters able to measure the power consumption of the CPU, the memory, or the GPU, with a typical measurement period of dozens of milliseconds. During this period of time, an application may execute hundreds of functions. As a result, identifying the power-hungry functions that are typically executed in a few microseconds is tedious.

For this work, we propose leveraging the repetitive nature of task-based runtime systems in order to estimate the tasks’ power consumption at the microsecond scale. We study the StarPU task-based runtime system for heterogeneous machines.

To achieve this, we trace the execution of the application while regularly measuring the energy consumption of the machine through various vendor's software interfaces. The resulting trace also contains information about the performed tasks. After the execution of the application, we identify the executed tasks for each energy measurement interval. We then use this information to obtain an overdetermined linear system linking tasks and energy measurements. Subsequently, solving the system allows us to estimate the power consumption of each task and thus infer energy values for all tasks independently of their actual duration.