COCKTAIL: Multi-Core Co-Optimization Framework With Proactive Reliability Management

High performance computing (HPC) servers aim to meet an increase in the number and complexity of tasks and, consequently, address energy efficiency challenge. In addition efficiency, it is essential manage lifetime limitations power-hungry components (e.g., cores cache), hence avoiding server failure before its period. Traditional approaches focus on either using hybrid caches reduce leakage power traditional static random-access memory (SRAM) cache, thus or trade-off between multi-core processors. However, these fall short terms flexibility applicability for HPC multi-parametric optimization including quality-of-service (QoS), reliability, efficiency. As a result, this paper we propose COCKTAIL, holistic strategy framework jointly optimize processors context, while guaranteeing reliability. First, analyze best cache technology among SRAM resistive random access (RRAM), within context architectures, improve endurance limits with respect requirements. Second, introduce novel efficient proactive queue policy reorder execution considering their end time possible reliability effects use caches. Third, present dynamic model predictive control (MPC)-based management method maximize task performance, by controlling frequency, temperature, target processor. Our results demonstrate that, consuming similar energy, COCKTAIL provides up 60% QoS improvement when compared latest state-of-the-art techniques context. Moreover, our guarantees design longer than 5 years whole system.