Scheduling and Load Balancing for Multi-module Applications on Heterogeneous Systems

The main focus of this work is on investigating the efficient scheduling and load balancing methods for parallel heterogeneous execution environments when the computations of a complex multi-module application. In such execution scenarios, there are several levels of heterogeneity that must be explicitly considered in order to attain efficient cross-device collaborative computing. At the system level, it is essential to conduct workload partitioning such that the minimum processing time is achieved and to efficiently exploit the capabilities of all relevant system resources, e.g., the performance of architecturally different devices and the asymmetric bandwidth of communication links. However, the additional level of heterogeneity is also introduced at the application level via data-dependent modules with different characteristics and scheduling opportunities. In order to tackle these problems for a real-world application with the above-mentioned complexity, a specific focus is paid herein on proposing efficient adaptive load balancing algorithms that allow achieving the real-time collaborative video encoding on commodity desktop systems. These algorithms rely on different scheduling and performance modeling techniques to efficiently exploit the parallel execution at several levels, namely: i) at the level of a single module; ii) across several modules, and iii) at the overall application level. In particular, the proposed adaptive load balancing approaches integrate different scheduling techniques based on Functional Performance Modeling, Directed Acyclic Graphs and Linear Programming to explicitly take into account the execution requirements of the individual modules, interand intra-module dependencies, as well as the tight execution constraints of the considered application. Moreover, the algorithm for minimization of the communication volume when two divisible load applications share the access to the single data partitioned buffer is also derived.