Self-Aware Thermal Management for High-Performance Computing Processors

Abstract

Processors for high performance computing and server workload are today thermally constrained. To preserve a safe working temperature, state-of-the-art processors for this market segment integrates many cores on the same die and feature fine-grain power management and thermal management feedback loops implemented in hardware. However, to keep the control policy simple, these controllers fail in taking advantage on the underlining thermal heterogeneity, long thermal transients and specific user mode. In this paper, we present a self-aware thermal management framework for making each chip self-aware of its thermal and workload peculiarities and use it to maximize performance and thermal sustainability.

The ANTAREX domain specific language for high performance computing

“…To do so, we have developed a new power capping run-time based on a set of user space APIs which can be used to define a relative priority for the given task currently in execution on a given core. Thanks to this priority, the run-time is capable of allocating more power to the higher priority process [29,30]. In ANTAREX, these APIs can be inserted by LARA aspects in the application code.…”

Supporting the Scale-Up of High Performance Application to Pre-Exascale Systems: The ANTAREX Approach

“…In ANTAREX, at runtime, the mARGOt tool [23] configures the available software knobs (application parameters, code transformations and code variants) according to the runtime information coming from application self-monitoring and system monitoring, thus creating an autotuning control loop. Finally, the runtime power manager, PowerCapper, is used to control the resource usage for the underlying computing infrastructure given the changing conditions [24,25].…”

ANTAREX: A DSL-Based Approach to Adaptively Optimizing and Enforcing Extra-Functional Properties in High Performance Computing

“…To do so, we have developed a new power capping run-time based onaset of user space APIs which can be used to define a relative priority for the given task currently in execution on a given core. Thanks to this priority, the run-time is capable of allocating more power to the higher priority process [ 28,29]. In ANTAREX, these APIs can be inserted by LARA aspects in the application code.…”