Won, Jae Yeon (2015-08). Dynamic Voltage and Frequency Scaling Techniques for Chip Multiprocessor Designs. Doctoral Dissertation. Thesis uri icon

abstract

  • Due to chip power density limitations as well as the recent breakdown of Dennard's Scalingover the past decade, performance growth in microprocessor design has largely been driven by core scaling. These trends have led to Chip Multi- Processor(CMP) designs, currently with tens of cores, and expected to grow to the thousands in the pursuit of exascale computing. The more complicated CMP design is more leading power consumption relatively in computer architecture. The increased power consumption generates thermal issues, and so performance degradation. Therefore, it is certain that power efficient algorithm in CMP and main memory are essential. For the power efficiency, we focus on dynamic voltage/frequency scaling (DVFS) techniques for CMP and main memory. In the first work, we focus on the "uncore", consisting of an on-chip communication fabric and shared LLC in CMP. The uncore now occupies as much as 30% of the overall die area, which is not negligible in CMP design, but has rarely researched. We find there are predictable patterns in uncore utility which point towards the potential of a proactive approach to uncore power management. In this work, we utilize artificial intelligence principles to proactively leverage uncore utility pattern prediction via an Artificial Neural Network (ANN). Even though the uncore takes non-negligible portion of CMP power consumption, processor cores still exist as major power consumers. For core DVFS, We explore a novel approach with the potential to achieve synergistic energy-savings and performance gain in chip multiprocessors (CMPs). In current designs, performance must typically be traded-off to achieve energy savings or, conversely, performance gains come with significant energy overhead. Resources shared by processor cores, such as on-chip interconnect and shared memory, play an increasingly critical role in determining the overall CMP performance. Our key observation is that per-core DVFS can be used as a client regulation mechanism for the shared resources. Based on this observation, we propose a new DVFS technique inspired by TCP Vegas, a congestion control protocol from the IP-networking domain. In addition to uncore in CMP, main memory is also critical shared resource in total system. As uncore is critical resource for CMP performance while occupying critical portion of total CMP energy consumed, main memory is also critical for total performance and accounts for large fraction of total energy consumption. Most conventional approaches focused on utilization of cores and memory only for memory power management. We found, however, the uncore plays an important role of total system performance and its utilization must be considered as well for memory power management. From the observation, we propose shared resource utilization aware power management technique for main memory. Our technique chooses low V/F level of memory for some congested case in uncore, and so derives negligible performance degradation while saving more energy by the low V/F level. We also proposed coordination policies to avoid oscillation issues among individual DVFS techniques (i.e. over energy saving or over performance increment). Full system simulations on PARSEC benchmarks show that our coordinated technique reduces total energy dissipation by over 47% across all benchmarks with less than 2.3% performance degradation.
  • Due to chip power density limitations as well as the recent breakdown of Dennard's Scalingover the past decade, performance growth in microprocessor design has largely been driven by core scaling. These trends have led to Chip Multi- Processor(CMP) designs, currently with tens of cores, and expected to grow to the thousands in the pursuit of exascale computing. The more complicated CMP design is more leading power consumption relatively in computer architecture. The increased power consumption generates thermal issues, and so performance degradation. Therefore, it is certain that power efficient algorithm in CMP and main memory are essential. For the power efficiency, we focus on dynamic voltage/frequency scaling (DVFS) techniques for CMP and main memory.

    In the first work, we focus on the "uncore", consisting of an on-chip communication fabric and shared LLC in CMP. The uncore now occupies as much as 30% of the overall die area, which is not negligible in CMP design, but has rarely researched. We find there are predictable patterns in uncore utility which point towards the potential of a proactive approach to uncore power management. In this work, we utilize artificial intelligence principles to proactively leverage uncore utility pattern prediction via an Artificial Neural Network (ANN).

    Even though the uncore takes non-negligible portion of CMP power consumption, processor cores still exist as major power consumers. For core DVFS, We explore a novel approach with the potential to achieve synergistic energy-savings and performance gain in chip multiprocessors (CMPs). In current designs, performance must typically be traded-off to achieve energy savings or, conversely, performance gains come with significant energy overhead. Resources shared by processor cores, such as on-chip interconnect and shared memory, play an increasingly critical role in determining the overall CMP performance. Our key observation is that per-core DVFS can be used as a client regulation mechanism for the shared resources. Based on this observation, we propose a new DVFS technique inspired by TCP Vegas, a congestion control protocol from the IP-networking domain.

    In addition to uncore in CMP, main memory is also critical shared resource in total system. As uncore is critical resource for CMP performance while occupying critical portion of total CMP energy consumed, main memory is also critical for total performance and accounts for large fraction of total energy consumption. Most conventional approaches focused on utilization of cores and memory only for memory power management. We found, however, the uncore plays an important role of total system performance and its utilization must be considered as well for memory power management. From the observation, we propose shared resource utilization aware power management technique for main memory. Our technique chooses low V/F level of memory for some congested case in uncore, and so derives negligible performance degradation while saving more energy by the low V/F level. We also proposed coordination policies to avoid oscillation issues among individual DVFS techniques (i.e. over energy saving or over performance increment).

    Full system simulations on PARSEC benchmarks show that our coordinated technique reduces total energy dissipation by over 47% across all benchmarks with less than 2.3% performance degradation.

publication date

  • August 2015