Tackled low utilization and poor energy proportionality in servers using reconfigurable cores which tune performance/power by turning on/off lanes in a superscalar out-of-order core.
Enabled co-scheduling of latency-sensitive & batch apps and handled interference by dynamically reconfiguring cores.
Reconfigurable cores, as against DVFS, offer a solution towards energy proportionality as they reduce static power.
Formulated it as an optimization problem that involves appropriately reconfiguring cores, to maximize throughput for batch apps, and meet the QoS of latency-sensitive apps while operating under server node’s power budget.
Developed online resource manager to rapidly characterize the system and to find near-optimal solution using a combination of collaborative filtering and heuristic search techniques.
Preserved QoS of all latency-sensitive apps and also, achieved up to 3.5x performance improvement on batch apps compared to core-level gating under stringent power budget across different latency-sensitive and batch apps.