Memory Analysis and Optimized Allocation of Dataflow Applications on Shared-Memory MPSoCs

Abstract

The majority of applications, ranging from the low complexity to very multifaceted entities requiring dedicated hardware accelerators, are very well suited for Multiprocessor Systems-on-Chips (MPSoCs). It is critical to understand the general characteristics of a given embedded application: its behavior and its requirements in terms of MPSoC resources.This paper presents a complete method to study the important aspect of memory characteristic of an application. This method spans the theoretical, architectureindependent memory characterization to the quasi optimal static memory allocation of an application on a real shared-memory MPSoC. The application is modeled as an Synchronous Dataflow (SDF) graph which is used to derive a Memory Exclusion Graph (MEG) essential for the analysis and allocation techniques. Practical considerations, such as cache coherence and memory broadcasting, are extensively treated.Memory footprint optimization is demonstrated using the example of a stereo matching algorithm from the computer vision domain. Experimental results show a reduction of the memory footprint by up to 43% compared to a state-of-the-art minimization technique, a throughput improvement of 33% over dynamic allocation, and the introduction of a tradeoff between multicore scheduling flexibility and memory footprint.

Odyn: Deadlock Prevention and Hybrid Scheduling Algorithm for Real-Time Dataflow Applications

“…Edges in a MEG represent exclusion relations, i.e., the impossibility to share physical memory. In Odyn, a MEG differs from the definition in [19] as it is derived from input I: set of inbound events; H = (T, D): AHSDF dependency graph; T: set of tasks; D: set of dependencies; define R: tasks ready to be fired; W: tasks waiting for input buffers; P t : tasks/events producing input buffers of task t; Start t : start time of task t; End t : end time of task t; Alloc a,b : allocation time of task-a-to-b buffers; Free a,b : deallocation time of task-a-to-b buffers; send_event(node, event, time): inform scheduler of node that event will occur at time; receive_events(): get new event notifications from other schedulers; W ← T; current_time ← now;…”

“…Desnos et al introduced Memory Exclusion Graphs (MEGs) in [19]: a MEG is an undirected weighted graph where vertices represent indivisible memory objects that correspond to communication buffers in an SDF graph, the working memory of SDF actors and feedback FIFOs that store initial tokens in an SDF graph. Edges in a MEG represent exclusion relations, i.e., the impossibility to share physical memory.…”

“…Formally, in Odyn, a MEG is defined as a weighted undirected graph M = < B, R > where the vertices B are the buffers in D p that capture data dependencies between tasks that run on a given PE, T p . Similarly to [19], vertices in B are weighted with the sizes of buffers D p and edges R represent memory exclusion relations. As in [19], in Odyn, we consider that the memory allocated to buffers D p is reserved from the execution start of the producer actor until the completion of the consumer actor in T p .…”

“…Similarly to [19], vertices in B are weighted with the sizes of buffers D p and edges R represent memory exclusion relations. As in [19], in Odyn, we consider that the memory allocated to buffers D p is reserved from the execution start of the producer actor until the completion of the consumer actor in T p . As described in [19], AHSDF graphs can be updated with scheduling constraints.…”

Controlling the Memory Subscription of Distributed Applications with a Task-Based Runtime System

“…A prototype for a more general resource-aware task scheduling was developed in OmpSs [15]. The topic was also studied with keen interest in the domain of embedded computing, where memory resources may drastically be limited [16], [17]. We refer to Section 4.3 of [18] for a survey of theoretical and practical memory-aware algorithms.…”