Active Power Management Technology Challenges and Implications for Programming Models
John SHALF, NERSC - LAWRENCE BERKELEY LABS

Click HERE to download presentation

Abstract : Dynamic Voltage and Clock Frequency scaling has dominated the discussion of active power management and power-aware algorithm design. However, there are many finer grained energy savings mechanisms that have yet to be fully exploited in server chip design. This talk will provide a survey of contemporary power management mechanisms incorporated into modern server chip designs as well as the many more aggressive mechanisms employed by mobile and embedded devices. For example, embedded and mobile devices make aggressive use of dark silicon, subthreshold logic design, and even opportunities for using software recovery mechanisms to enable a trade-off of soft error rates to achieve substantial power savings. However, HPC integrators and software designers face daunting challenges of coordinating mechanisms used for local optimal power management into large scale systems. Although these more aggressive techniques could enable enormous energy savings, these methods have a huge impact on the intrinsic performance inhomogeneity of our programming environment. Such changes fundamentally unravel the bulk-synchronous/SPMD programming paradigm that underpins the majority of our current HPC applications. Systemwide coordinated power management control loop cannot operate at the timescale that these local decisions are made. Such dramatic changes drive the study of alternative execution models to overcome the challenges of extreme performance heterogeneity and software-based resilience.

This talk will discuss these emerging technologies for more aggressive local power management and the implications for our programming environment. I will describe recent research into alternative execution models, and describe results from example implementations of these alternative models for computation.