Large 000001000

Performance Isolation on Modern Multi-Socket Systems

Sandhya Dwarkadas

Recorded 11 December 2017 in Lausanne, Vaud, Switzerland

Event: IC Colloquia - EPFL IC School Colloquia


Recognizing that applications are rarely executed in isolation today, I will discuss some practical challenges in making best use of available hardware and our approach to addressing these challenges. I will describe two independent and complementary control mechanisms using low-overhead hardware performance counters that we have developed: a sharing- and resource-aware mapper (SAM) to effect task placement with the goal of localizing shared data communication and minimizing resource contention based on the offered load; and an application parallelism manager (MAP) that controls the offered load with the goal of improving system parallel efficiency. Our results emphasize the need for low-overhead monitoring of application behavior under changing environmental conditions in order to adapt to environment and application behavior changes. If time permits, I will also outline additional work on memory management design that eliminates address translation redundancy via appropriate sharing.

Watched 191 times.