Abstract

consumption is one of the main limiting factors for designing and deploying ultrascale systems. Therefore, this paper presents challenges and trends associated with energy efficiency for ultrascale systems based on current activities of the working group on Energy Efficiency in the European COST Action Nesus IC1305. The analysis contains major areas that are related to studies of energy efficiency in ultrascale systems: heterogeneous and low power hardware architectures, power monitoring at large scale, modeling and simulation of ultrascale systems, energy-aware scheduling and resource management, and energy-efficient application design.

Highlights

  • Energy consumption is one of the main limiting factors for designing and deploying Ultrascale systems

  • The Intel Running Average Power Limit (RAPL) interface reports per-package estimates of total energy consumed on Intel Sandy Bridge CPUs and later; the Nvidia Management Library (NVML) interface can query instant power draw values from recent Nvidia Tesla Graphics Processing Units (GPUs); some motherboards report power draw value through extensions to the Intelligent Platform Management Interface (IPMI)

  • Green 500 and Green Graph 500 are mainly relevant in the domain of scientific computing, but what about most of data centers? One fact is that GPU accelerators can be considered as valuable accelerators but they still need a larger adoption, especially in industrial and business applications

Read more

Summary

Introduction

Energy consumption is one of the main limiting factors for designing and deploying Ultrascale systems. Moving from pure performance goals to energy consumption and thermal issues For all these areas modeling and simulation techniques are needed to analyse energy efficiency of hardware, applications and whole computing systems at ultrascale level. Future Ultrascale Computing Systems (UCSs) are envisioned as hybrid systems composed of heterogeneous resources and platforms ranging from ”traditional” High Performance Computing (HPC) systems, CC infrastructures and ultra low-power computing systems Another reason for this tendency toward an heterogeneous design is that there is no single approach which is optimal for all computing needs. At an intermediate level (between software and hardware), virtualization is emerging as the prominent approach to mutualize the energy consumed by a single server running multiple Virtual Machines (VMs) instances This approach, commonly designated as Cloud Computing (CC) [17, 87] is increasingly advertised as THE solution to most IT problems.

Power monitoring and profiling of ultrascale context
External devices
Intranode devices
Hardware sensors
Software interfaces
Deploy to ultrascale level
Energy modeling and simulation in ultrascale systems
Simulating multi-criteria scheduling techniques
Modeling complex workloads
Energy efficient cloud computing
Modeling and simulation of thermal processes
Energy-aware modeling and simulation tools
Summary
Resource management and scheduling in extra large scale systems
Workload characteristics
Virtualization
Resource management and scheduling
Towards an Energy Efficient Design of Ultrascale Applications
Green 500 and Hybrid Architectures
Sharing Hardware Accelerators Between Virtual Machines
Low Power SoC
Findings
Discussion
Conclusions
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call