Radoslaw Szymanek

A constructive algorithm for memory-aware task assignment and scheduling

This paper presents a constructive algorithm for memory-aware task assignment and scheduling, which is a part of the prototype system MATAS. The algorithm is well suited for image and video processing applications which have hard memory constraints as well as constraints on cost, execution time, and resource usage. Our algorithm takes into account code and data memory constraints together with the other constraints. It can create pipelined implementations. The algorithm finds a task assignment, a schedule, and data and code memory placement in memory. Infeasible solutions caused by memory fragmentation are avoided. The experiments show that our memory-aware algorithm reduces memory utilization comparing to greedy scheduling algorithm which has time minimization objective. Moreover, memory-aware algorithm is able to find task assignment and schedule when time minimization algorithm fails. MATAS can create pipelined implementations, therefore the design throughput is increased.

MDD propagators with explanation

Multi-valued decision diagrams (MDDs) are a convenient approach to representing many kinds of constraints including table constraints, regular constraints, complex set and multiset constraints, as well as ad-hoc problem specific constraints. This paper introduces an incremental propagation algorithm for MDDs, and explores several methods for incorporating explanations with MDD-based propagators. We demonstrate that these techniques can provide significantly improved performance when solving a variety of problems.

Partial task assignment of task graphs under heterogeneous resource constraints

This paper presents a novel partial assignment technique (PAT) that decides which tasks should be assigned to the same resource without explicitly defining assignment of these tasks to a particular resource. Our method simplifies the assignment and scheduling steps while imposing a small or no penalty on the final solution quality. This technique is specially suited for problems which have different resources constraints. Our method does not cluster tasks into a new task, as typical clustering techniques do, but specifies which tasks need to be executed on the same processor. Our experiments have shown that PAT, which may produce nonlinear groups of tasks, gives better results than linear clustering when multi-resource constraints are present. Linear clustering was proved to be optimal comparing to all other clusterings for problems with timing constraints only. In this paper, we show that, if used for multi-resource synthesis problem, as it is often used nowadays, linear clustering will produce inferior solutions.

Coordination of software components with BIP: application to OSGi

Coordinating component behaviour and access to resources is among the key difficulties of building large concurrent systems. To address this, developers must be able to manipulate high-level concepts, such as Finite State Machines and separate functional and coordination aspects of the system behaviour. OSGi associates to each bundle a state machine representing the bundles lifecycle. However, once the bundle has been started, it remains in the state Active - the functional states are not represented. Therefore, this mechanism is not sufficient for coordination of active components. In this paper, we present a methodology for functional component coordination in OSGi by using BIP coordination mechanisms. BIP allows us to clearly separate the system-wide coordination policies from the component behaviour and the interface that components expose for interaction. By using BIP, we show how the allowed global states and state transitions of the modular system can be taken into account in a non-invasive manner and without any impact on the technology stack within an OSGi container.

Task assignment and scheduling under memory constraints

Many DSP and image processing embedded systems have hard memory constraints which makes it difficult to find a good task assignment and scheduling which fulfill these constraints. The paper presents a new heuristic developed for task assignment and scheduling for such systems. These systems also have a large number of constraints of different nature, such as cost, execution time, memory capacity and limitations on resource usage. The heterogeneous constraints require new synthesis methods which will take them into account while searching for a valid solution. The heuristic presented in the paper is a part of the CLASS system (Constraint Logic Programming based System Synthesis).

Exogenous Coordination of Concurrent Software Components with JavaBIP

A strong separation of concerns is necessary in order to make the design of domain‐specific functional components independent from cross‐cutting concerns, such as concurrent access to the shared resources of the execution platform. Native coordination mechanisms, such as locks and monitors, allow developers to address these issues. However, such solutions are not modular; they are complex to design, debug, and maintain. We present the JavaBIP framework that allows developers to think on a higher level of abstraction and clearly separate the functional and coordination aspects of the system behavior. It implements the principles of the Behavior, Interaction, and Priority (BIP) component framework rooted in rigorous operational semantics. It allows the coordination of existing concurrent software components in an exogenous manner, relying exclusively on annotations, component APIs, and external specification files. We introduce the annotation and specification syntax of JavaBIP and illustrate its use on realistic examples, present the architecture of our implementation, which is modular and easily extensible, and provide and discuss performance evaluation results. Copyright

An efficient generic network flow constraint

We propose a generic global constraint that can be applied to model a wide range of network flow problems using constraint programming. In our approach, all key aspects of a network flow can be represented by finite domain variables, making the constraint very expressive. At the same time, we utilize a network simplex algorithm to design a highly efficient, and incremental, domain filtering algorithm. We thus integrate two powerful techniques for discrete optimization: constraint programming and the network simplex algorithm. Our generic constraint can be applied to automatically implement effective and efficient domain filterng algorithms for ad-hoc networks, but also for existing global constraints that rely on a network structure, including several soft global constraints many of which are not yet supported by CP systems. Our experimental results demonstrate the efficiency of our constraint, that can achieve speed-ups of several orders of magnitude with negligible overhead, when compared to a decomposition into primitive constraints.

Constraint-Level Advice for Shaving

This work concentrates on improving the robustness of constraint solvers by increasing the propagation strength of constraint models in a declarative and automatic manner. Our objective is to efficiently identify and remove shavable values during search. A value is shavable if as soon as it is assigned to its associated variable an inconsistency can be detected, making it possible to refute it. We extend previous work on shaving by using different techniques to decide if a given value is an interesting candidate for the shaving process. More precisely, we exploit the semantics of (global) constraints to suggest values, and reuse both the successes and failures of shaving later in search to tune shaving further. We illustrate our approach with two important global constraints, namely alldifferent and sum, and present the results of an experimentation obtained for three problem classes. The experimental results are quite encouraging: we are able to significantly reduce the number of search nodes (even by more than two orders of magnitude), and improve the average execution time by one order of magnitude.

A CP-LP approach to network management in OSPF routing

In this paper, we consider a routing problem related to the widely used Open Shortest Path First (OSPF) protocol, which is considered a challenge within the Constraint Programming (CP) community. We address the special version of OSPF which requires unique and symmetrical paths. To solve this problem, we propose a novel hybrid approach which combines CP and Linear Programming (LP). Our approach employs a new global constraint with problem-specific filtering algorithms to efficiently remove inconsistent values from partial solutions. Moreover, this constraint employs two LP relaxations which are used to indicate in-feasible partial solutions due either to network capacity constraints, or to protocol-specific routing constraints. We show the efficiency of our complete approach on backbone networks with hundreds of different demands to route.

Data assignment and access scheduling exploration for multi-layer memory architectures

This work presents an exploration framework which performs data assignment and access scheduling exploration for applications given a multilayer memory architecture. Our framework uses multiobjective criteria during exploration, such as application execution time, energy, bandwidth, and data size. In order to tackle the complexity of the exploration, it is divided into three phases; Pareto diagram composition, data assignment, and access scheduling. The first phase produces multidimensional Pareto points for our application. After this phase, our framework produces distinct data assignments which are represented as Pareto points in a two dimensional space defined by bandwidth requirements and size requirements. Finally, the scheduling phase finds possibly optimal order of the tasks and performs precise scheduling of the tasks. Three feedbacks paths are present which can be used to iteratively improve exploration results. It is possible to trade off the quality of the results and the algorithm runtime. We have evaluated our framework on a medical image processing application. We have shown that our algorithms can perform exploration of the huge design space in an iterative manner and obtains good Pareto diagram coverage.