When homogeneous becomes heterogeneous: Wearout aware task scheduling for streaming applications

Recent trends in process technology suggest the need to moni-tor transistor wear-out in future processes. Because of within-die variation and the different computations being run on each core in a multi-core chip, this wear-out causes further imbalance to initial core frequencies as time progresses. Fur-thermore, manufacturing defects mean that cache sizes can vary between cores, adding further imbalance to a system. If we allow different cores to independently control their op-erating frequency we can achieve the best possible perfor-mance for their part of the die. Other parts of the system with slowly degrading performance can include interconnects and Flash-based file caches. In this paper we first explain how conventionally homogeneous multi-core processors can become heterogeneous over time. We discuss possible oper-ating system based solutions to maximize the performance of a system as it wears out and present illustrative theoretical results based on linear programming. We demonstrate that for a class of streaming applications, an intelligent schedul-ing scheme recovers a significant amount of performance lost through wear-out. We advocate the need for multiple accu-rate performance measurements for effective scheduling in a wearout-aware multicore chip. 1