Associate Professor Ankur Srivastava
Institute for Systems Research
Department of Electrical and Computer Engineering
Abstract: Increasing levels of silicon integration has resulted in a significant increase in chip power density and thermal level. Temperature and power in computer systems are interdependent quantities partly due to the non linear leakage thermal relationship and also due to increase in cooling energy necessitated by higher power dissipation. This talk presents methods for exploiting the application dependent workload characteristics to adaptively customize the CPU behavior for delivering the required performance at acceptable thermal and power levels. We discuss techniques for developing an on-chip thermal sensing infrastructure that allows online estimation of the thermal state of the chip which in turn is used to develop our control policies. We end our discussion with more complex computer systems such as datacenters.