Interference-Aware Scheduling Using Geometric Constraints

in Euro-Par 2018: Parallel Processing (2018, August)

The large scale parallel and distributed platforms produce a continuously increasing amount of data which have to be stored, exchanged and used by various jobs allocated on different nodes of the platform ... [more ▼]

The large scale parallel and distributed platforms produce a continuously increasing amount of data which have to be stored, exchanged and used by various jobs allocated on different nodes of the platform. The management of this huge communication demand is crucial for the performance of the system. Meanwhile, we have to deal with more interferences as the trend is to use a single all-purpose interconnection network. In this paper, we consider two different types of communications: the flows induced by data exchanges during computations and the flows related to Input/Output operations. We propose a general model for interference-aware scheduling, where explicit communications are replaced by external topological constraints. Specifically, we limit the interferences of both communication types by adding geometric constraints on the allocation of jobs into machines. The proposed constraints reduce implicitly the data movements by restricting the set of possible allocations for each job. We present this methodology on the case study of simple network topologies, namely the line and the ring. We propose theoretical lower and upper bounds under different assumptions with respect to the platform and jobs characteristics. The obtained results illustrate well the difficulty of the problem even on simple topologies. [less ▲]

Scheduling with uncertainties on new computing platforms

in Computational Optimization and Applications (2010), 48(2), 369-398

New distributed computing platforms (grids) are based on interconnections of a large number of processing elements. A most important issue for their effective utilization is the optimal use of resources ... [more ▼]

New distributed computing platforms (grids) are based on interconnections of a large number of processing elements. A most important issue for their effective utilization is the optimal use of resources through proper task scheduling. It consists of allocating the tasks of a parallel program to processors on the platform and to determine at what time the tasks will start their execution. As data may be subject to uncertainties or disturbances, it is practically impossible to precisely predict the input parameters of the task scheduling problem. We briefly survey existing approaches for dealing with data uncertainties and discuss their relevance in the context of grid computing. We describe the stabilization process and analyze a scheduling algorithm that is intrinsically stable (i.e., it mitigates the effects of disturbances in input data at runtime). This algorithm is based on a decomposition of the application graph into convex sets of vertices. Finally, it is compared experimentally to pure on-line and well-known off-line algorithms. [less ▲]

A New Genetic Algorithm for Scheduling for Large Communication Delays

in 15th International Euro-Par Conference (2009, August 28)

In modern parallel and distributed systems, the time for exchanging data is usually larger than that for computing elementary operations. Consequently, these communications slow down the execution of the ... [more ▼]

In modern parallel and distributed systems, the time for exchanging data is usually larger than that for computing elementary operations. Consequently, these communications slow down the execution of the application scheduled on such systems. Accounting for these communications is essential for attaining efficient hardware and software utilization. Therefore, we provide in this paper a new combined approach for scheduling parallel applications with large communication delays on an arbitrary number of processors. In this approach, a genetic algorithm is improved with the introduction of some extra knowledge about the scheduling problem. This knowledge is represented by a class of clustering algorithms introduced recently, namely, convex clusters which are based on structural properties of the parallel applications. The developed algorithm is assessed by simulations run on some families of synthetic task graphs and randomly generated applications. The comparison with related approaches emphasizes its interest. [less ▲]