Optimizing throughput/power trade-offs in hardware transactional memory using DVFS and intelligent scheduling

Abstract

Power has emerged as a first-order design constraint in modern processors and has energized microarchitecture researchers to produce a growing number of power optimization proposals. Almost in tandem with the move toward more energy-efficient designs, architects have been increasing the number of processing elements (PEs) on a single chip and promoting the concept of running multithreaded workloads. Nevertheless, software is still lagging behind and is often unable to exploit these additional resources -giving rise to transactional memory. Transactional memory is a promising programming abstraction that makes it easier for programmers to exploit the resources available in manycore processor systems by removing some of the complexity associated with traditional lock-based programming. This paper proposes new techniques to merge the power and transactional memory domains.An analysis of the per-core and chip-wide power consumption of hardware transactional memory systems (HTMs) pinpoints two areas ripe for power management policies: transactional stalls and aborts. The first proposed policy uses dynamic voltage and frequency scaling (DVFS) during transactional stall periods. By frequency scaling PEs based on their transactional state, DVFS can increase the throughput and energy efficiency of HTMs. The second method uses a transaction's conflict probability to reschedule transactions and clock gate aborted PEs to reduce overall contention and power consumption within the system. The proposed techniques are evaluated using three HTM configurations and are shown to reduce the energy delay squared product (ED2P) of the STAMP and SPLASH-2 benchmarks by an average of 18% when combined. Synthetic workloads are used to explore a wider range of program behaviors and the optimizations are shown to reduce the ED2P by an average of 29%. For a comparison, this work is shown reduce the ED2P by up to 30% relative to previous proposals for energy reduction in HTMs (e.g. transaction serialization).

Application configuration selection for energy-efficient execution on multicore systems

“…In Equation 5, the percentage of C 2,2 and M 2,2 is pre-determined by the workload while in Equation 6, the prediction result of M i,j is determined by both the system architecture and the workload itself. Combine Equation 5 and 6, we obtain Equation 7.…”

“…The speedup of all the benchmarks can be observed from Figure 3 and Figure 4. At the configuration where all eight cores are used, ep.C reaches highest speedup factor, which is 7.60, followed by is.C, which shows 6 Table 6: Percentage of stall in the configuration of (2,2). are used.…”

“…An analytical model is needed to understand the energy efficiency of a workload in a multi-core computing scenario. While allocating a workload to multiple CPUs is an effective way to reduce computation energy consumption, DVFS is usually used to explore slacks during execution to save extra power dissipation [6].…”

Energy-Performance Tradeoffs in Software Transactional Memory

“…Experimental results reveal an average 19% reduction in energy consumption and a 4% speedup. Likewise, Hughes and Li [15] propose two policies to increase energy efficiency in HTM systems: a DVFS technique to save energy during transactional stalls, and a heuristic based on conflict probability to reschedule transactions and clock gate aborted processors. Our work is in tandem with the latter two papers in as much as we also capitalize on reducing the energy spent on aborted transactions and transactional stalls.…”

An integrated approach to system-level CPU and memory energy efficiency on computing systems

“…This is different that the independent policy that is oblivious for the activity occurring in the other resource. (6) Due to the additional information about the other resource, it is expected that Semi-Independent Policy to achieve higher energy savings than Independent Policy, because the decision mechanisms are aware of each other's statistics. However, it cannot estimate the impact of the other resource's frequency on the overall performance since frequency selection of each resource is managed independently.…”