StreaMorph: A case for synthesizing energy-efficient adaptive programs using high-level abstractions

Abstract

This paper presents the concept of adaptive programs, whose computation and communication structures can morph to adapt to environmental and demand changes to save energy and computing resources. In this approach, programmers write one single program using a language at a higher level of abstraction. The compiler will exploit the properties of the abstractions to generate an adaptive program that is able to adjust computation and communication structures to environmental and demand changes.We develop a technique, called StreaMorph, that exploits the properties of stream programs' Synchronous Dataflow (SDF) programming model to enable runtime stream graph transformation. The StreaMorph technique can be used to optimize memory usage and to adjust core utilization leading to energy reduction by turning off idle cores or reducing operating frequencies. The main challenge for such a runtime transformation is to maintain consistent program states by copying states between different stream graph structures, because a stream program optimized for different numbers of cores often has different sets of filters and inter-filter channels. We propose an analysis that helps simplify program state copying processes by minimizing copying of states based on the properties of the SDF model.Finally, we implement the StreaMorph method in the StreamIt compiler. Our experiments on the Intel Xeon E5450 show that using StreaMorph to minimize the number of cores used from eight cores to one core, e.g. when streaming rates become lower, can reduce energy consumption by 76.33% on average. Using StreaMorph to spread workload from four cores to six or seven cores, e.g. when more cores become available, to reduce operating frequencies, can lead to 10% energy reduction. In addition, StreaMorph can lead to a buffer size reduction of 82.58% in comparison with a straightforward inter-core filter migration technique when switching from using eight cores to one core.

AdaPNet

“…It uses the fusion and fission operators to generate a schedule that maximizes the throughput of the application. StreaMorph [2] is a technique to adapt SDF graphs at run-time by performing a reverse sequence of executions to bring the graph into a known state. In contrast to the previous two approaches, our technique does not assume a static schedule and is therefore applicable to more complex MoCs than SDF graphs.…”

“…Specially, we address this challenge in the context of streaming applications, which are modeled as process networks. In fact, previous works showed that large performance gains in terms of throughput [17,20] or energy consumption [2] can be obtained if the process network is refined at compile-time.…”

“…These steps are not trivial if stateful processes are executed asynchronously and the amount of data being produced or consumed by a process is not known beforehand. Previous works, therefore, limited their approaches to programming models with statically specified data production and consumption rates [2] or to the replication of stateless processes [3].…”

Unified dataflow model for the analysis of data and pipeline parallelism, and buffer sizing

“…It is shown in the context of the StreamIt language [18] that the combination of pipeline parallelism and coarse-grained data parallelism is important [1]. Other methods acknowledge the benefits of the combination of data and pipeline parallelism [19], [20] and use dataflow models to analyze the temporal behavior of data parallel tasks. However, none of these methods determines the amount of data parallelism that is required to meet the throughput constraint of an application.…”