Nornir: A Power-Aware Runtime Support for Parallel Applications
Presenter
Event Type
Doctoral Showcase
Applications
Architectures
Heterogeneous Systems
Power
Runtime Systems
TimeTuesday, November 14th10am - 5:15pm
LocationFour Seasons Ballroom
DescriptionPower consumption is becoming a key factor in designing applications and computing systems.
In this thesis we propose Nornir, a runtime system for parallel applications, which can be used by application users to enforce power consumption or performance constraints on their applications, by operating on several knobs like: number of used cores, clock frequency of the cores, etc... Nornir is fully customizable and can be used by autonomic strategies designers to plug in new reconfiguration strategies and to validate them against existing ones.
In this thesis we also designed different reconfiguration algorithms: i) an heuristic to enforce performance constraints; ii) an online learning algorithms to enforce both performance and power consumption constraints; iii) a "concurrency control" algorithm to optimize performance and power consumption of single-producer, single-consumer queues.
All these algorithms have been implemented and validated by using Nornir.
In this thesis we propose Nornir, a runtime system for parallel applications, which can be used by application users to enforce power consumption or performance constraints on their applications, by operating on several knobs like: number of used cores, clock frequency of the cores, etc... Nornir is fully customizable and can be used by autonomic strategies designers to plug in new reconfiguration strategies and to validate them against existing ones.
In this thesis we also designed different reconfiguration algorithms: i) an heuristic to enforce performance constraints; ii) an online learning algorithms to enforce both performance and power consumption constraints; iii) a "concurrency control" algorithm to optimize performance and power consumption of single-producer, single-consumer queues.
All these algorithms have been implemented and validated by using Nornir.
Presenter
