Don't race the memory bus: taming the GC leadfoot

Abstract

Dynamic voltage and frequency scaling (DVFS) is ubiquitous on mobile devices as a mechanism for saving energy. Reducing the clock frequency of a processor allows a corresponding reduction in power consumption, as does turning off idle cores. Garbage collection is a canonical example of the sort of memory-bound workload that best responds to such scaling. Here, we explore the impact of frequency scaling for garbage collection in a real mobile device running Android's Dalvik virtual machine, which uses a concurrent collector. By controlling the frequency of the core on which the concurrent collector thread runs we can reduce power significantly. Running established multi-threaded benchmarks shows that total processor energy can be reduced up to 30 %, with end-to-end performance loss of at most 10 %.

One Process to Reap Them All

“…This happens not only because of its explicit overhead on CPU and memory cycles, but also because of implicit scheduling decisions by the OS and hardware with respect to CPU cores. Therefore, a potential approach to optimize GC cost per single VM is to take advantage of GC idleness and control the frequency of the core on which the concurrent collector thread is running [23,34]. Distributed Controls.…”

“…How-ever, with the introduction of other system components into the cost equation (i.e., scheduling and CPU throttling), the GC scheduling can be tuned by hiding expensive GC operations inside of small, otherwise unused idle portions of application execution, which results in better overall execution [23]. For mobile devices, tuning the GC implementation to meet performance and power goals is exceptionally difficult, because per-app GC cost is defined as a function of several controls [23,34] such as: (i) the power manager reacting to CPU idle-time during memory-bound GC phases; (ii) VM configurations, GC algorithm and the heuristics controlling the heap size; and (iii) the workload and memory stress at run time.…”

“…The reduction of GC energy cost while allowing for better responsiveness and throughput. This can be achieved either (i) by coordinating between the system scheduler and the VM [23], or (ii) by capping the frequency of the core during GC phases [34].…”

“…Several fine-grained profile tools [12,47,49,54] show that apps still have room for significant improvements by reducing idle time and power leaks caused by inefficient software subcomponents [5,18,49,50]. This knowledge fueled the interest in tuning the performance and energy by direct coordination of the application layer and the power managers [23,34,37,40].…”

One Process to Reap Them All

“…VMs substantially increase productivity by abstracting the hardware and offering useful services such as dynamic optimization, class resolution and, garbage collection (GC). However, VM services impose performance and energy costs, ranging from 5% to over 70% [14,18,34,40,53].…”