Exploration of Energy and Throughput Tradeoffs for Dataflow Networks

Abstract

The introduction of dynamic power management strategies such as clock gating and power gating in dataflow networks has been shown to provide significant energy savings when applied during idle times. However, these strategies can also degrade throughput due to shutdown and wake-up delays. Such throughput degradations might be particularly detrimental to signal processing systems that require a guaranteed throughput. As a solution, this paper first contributes a linear-program formulation for finding a periodic maximal-throughput schedule of a given so-called self-powering dataflow network where actors, realized in hardware, are allowed to go to sleep whenever not being enabled to fire. Depending on which actors are allowed to power down, tradeoffs between throughput and energy savings can be obtained. As a second contribution, we propose a mixed-integer-linear-program formulation to determine a periodic schedule that satisfies a given throughput while minimizing the overall energy per period by identifying a respective set of actors that is allowed to power down in phases of idleness and which rather not. Finally, as a third contribution, we propose a multi-objective design-space exploration strategy called "Hop and Skip" to efficiently explore the Pareto front of energy and throughput solutions. Experimental evaluations on a set of existing benchmarks and randomly generated graphs witness significant exploration time reductions over a brute-force sweep. Finally, a real-world case study is elaborated, and we report on achievable energy savings and throughputs of the related dataflow network where (a) all actors are always-active, (b) all actors are self-powered, and (c) all optimal energy and throughput tradeoff points as found by the proposed design-space exploration strategy.

Figures (8)

• Figure 1: The shown actor $a_1$ has three input channels (incoming edges) and two output channels (outgoing edges), while the black dots on these edges represent tokens. In (a), $a_1$ is not fireable due to the absence of a token on its upper input channel. In (b), $a_1$ is fireable because at least one token is available on each of its input channels. Finally, (c) depicts the situation after $a_1$ has fired by consuming one token from each input channel and producing one token on each output channel.
• Figure 2: DFG of an AEC application, detailing the flow of data through actors (nodes) interconnected by FIFO channels (directed edges). A single initial token is present on the channel $(a_5, a_2)$, i.e., $\delta(a_5, a_2) = 1$.
• Figure 3: Schedule of the AEC DFG (see \ref{['fig:echo-cancellation']}) with minimal period $P = P_{\mathrm{min}} = 23$ (top). The arrows represent data dependencies. Shown as well is the power profile of actor $a_4$ when operating in always-active mode (bottom). During the execution phase of the actor (blue) a higher power $p^\mathrm{exe}(a_{4})$ is consumed than the power $p^\mathrm{idl}(a_{4})$ in the idle phase (yellow).
• Figure 4: Power consumption and execution phases of actor $a_4$ of the running example when operating in self-powered mode.
• Figure 5: Schedule of a hybrid implementation of the AEC network with $P=P_\mathrm{min}=23$. Actors $a_1$ and $a_6$ are set to self-powered mode, while the others must be set to always-active mode, to guard this minimal period.

Theorems & Definitions (14)

• Definition 1: Dataflow Graph E.Lee87SDF
• Example
• Definition 2: Periodic Schedule
• Definition 3: Maximum Cycle Mean Parhi91OptUnfold
• Example
• Example
• Example
• Example
• Example
• Example