Kadjo, David (2015-05). Energy Efficiency and Performance in Multiprocessors Systems on Chip. Doctoral Dissertation. Thesis uri icon

abstract

  • As process technology shrinks, the transistor count on CPUs has increased. The breakdown of Dennard scaling has led to diminishing returns in terms of performance per power. A trend which promises to impact future CPU designs. This breakdown is due in part to the increase in transistor leakage driven static power. We, now, have constrained energy and power budgets. Thus, energy consumption has to be justified by an increased in performance. Simultaneously, architects have shifted to chip multiprocessors(CMPs) designs with large shared last level cache(LLC) to mitigate the cost of long latency off-chip memory access. A primary reason for that shift is the power efficiency of CMPs. Additionally, technology scaling has allowed the integration of platform components on the chip; a design referred to as multiprocessors system on chip (MpSoC). This integration improves the system performance as the communication latency between the components is reduced. Memory subsystems are essential to CPUs performance. Larger caches provide the CPU faster access to a larger data set. Consequently, the size of last level caches have increased to become a significant leakage power dissipation source. We propose a technique to facilitate power gating a partition of the LLC by migrating the high temporal blocks to a live partition; Thus reducing the performance impact. Given the high latency of memory subsystems, prefetching improves CPU performance by speculating future memory accesses and requesting the data ahead of the demand. In the context of CMPs running multiple concurrent processes, prefetching accuracy is critical to prevent cache pollution effects. Furthermore, given the current constraint energy environment, there is a need for lightweight prefetchers with high accuracy. To this end, we present BFetch a lightweight and accurate prefetcher driven by control flow predictions and effective address speculation. MpSoCs have mostly been used in mobile devices. The energy constraint is more pronounced in MpSoCs-based, battery powered mobile devices. The need to address the energy consumption in MpSoCs is further accentuated by the proliferation of mobile devices. This dissertation presents a framework to optimize the energy in MpSoCs. The proposed framework minimizes the energy consumption while meeting performance and power budgets constraints. We first apply this framework to the CPU then extend it to accommodate the GPU.
  • As process technology shrinks, the transistor count on CPUs has increased. The breakdown of Dennard scaling has led to diminishing returns in terms of performance per power. A trend which promises to impact future CPU designs. This breakdown is due in part to the increase in transistor leakage driven static power. We, now, have constrained energy and power budgets. Thus, energy consumption has to be justified by an increased in performance. Simultaneously, architects have shifted to chip multiprocessors(CMPs) designs with large shared last level cache(LLC) to mitigate the cost of long latency off-chip memory access. A primary reason for that shift is the power efficiency of CMPs. Additionally, technology scaling has allowed the integration of platform components on the chip; a design referred to as multiprocessors system on chip (MpSoC). This integration improves the system performance as the communication latency between the components is reduced.

    Memory subsystems are essential to CPUs performance. Larger caches provide the CPU faster access to a larger data set. Consequently, the size of last level caches have increased to become a significant leakage power dissipation source. We propose a technique to facilitate power gating a partition of the LLC by migrating the high temporal blocks to a live partition; Thus reducing the performance impact. Given the high latency of memory subsystems, prefetching improves CPU performance by speculating future memory accesses and requesting the data ahead of the demand. In the context of CMPs running multiple concurrent processes, prefetching accuracy is critical to prevent cache pollution effects. Furthermore, given the current constraint energy environment, there is a need for lightweight prefetchers with high accuracy. To this end, we present BFetch a lightweight and accurate prefetcher driven by control flow predictions and effective address speculation.

    MpSoCs have mostly been used in mobile devices. The energy constraint is more pronounced in MpSoCs-based, battery powered mobile devices. The need to address the energy consumption in MpSoCs is further accentuated by the proliferation of mobile devices. This dissertation presents a framework to optimize the energy in MpSoCs. The proposed framework minimizes the energy consumption while meeting performance and power budgets constraints. We first apply this framework to the CPU then extend it to accommodate the GPU.

publication date

  • May 2015