Publication | Open Access
A Case for an Over-provisioned Multicore System: Energy Efficient Processing of Multithreaded Programs
17
Citations
0
References
2007
Year
Technology scaling has provided system designers with an exploding transistor \nbudget, far more than what was available when the core principles behind many \nexisting commodity microprocessors were envisioned. With this tremendous growth, \nhowever, comes a whole new set of engineering challenges involving power \ndensity, thermal efficiency, programmability and so on. In this paper, we study \nanother important trend in high performance microprocessors: the \nreduction in the Simultaneously Active Fraction (SAF) --- the fraction of the \nentire chip resources that can be active simultaneously, given a target power \nenvelope. As the improvement in the energy efficiency of individual transistor \ndevices is lagging behind the growth in their integration capacity, we find that \nthe SAF is monotonically decreasing for each successive technology generation. \n \n \nGiven this increasing constraint on the SAF, we examine the utility of \ntemporarily suspending computation on a core as a means for reducing the SAF, \nand hence, remain within the confines of cost-effective cooling and power \ndelivery. We investigate a SAF aware over-provisioned multicore system (OPMS), \nwhere only a subset of the available cores are employed to perform active \ncomputation at any given time, by allowing the individual cores to transition \nbetween active and inactive state. Though several possible directions for \nutilizing such an over-provisioned system are possible, this paper focuses on \nenergy efficient dynamic task redistribution. In particular, this paper examines \nthe use of Computation Spreading---a recently proposed technique for runtime \nspecialization of homogeneous multicores---in an OPMS. We show several benefits \nfor such an OPMS design, including reductions in energy, runtime, and superior \nthermal characteristics. Overall, our technique improves the energy-delay \nproduct of the commercial workloads we examine by 5--20%.