A Scheduling Framework Prototype For Heterogeneous Platform High Performance Computing

The massive amount of data available in the field of molecular biology and bioinformatics are required to be processed in a reasonable amount of time. Excessive runtime of single nodes leads to the necessity to assemble and maintain cluster computing systems. As algorithms might have inferior performance on some platforms, cluster computing systems utilizing only a single platform may run into performance slumps when performing some calculations. Thus, a heterogeneous platform type high performance cluster is assembled, consisting of several different types of hardware, including x86-64 CPUs, NVIDIA graphics cards, and the IBM Cell Broadband Engine. To run algorithms efficiently on such a cluster, a scheduling framework is necessary. The scheduling framework prototype to be developed for this cluster requires several smaller modules to work correctly. On the lowest tier, an algorithmic framework is required to allow developers to plug algorithms into the system without any further knowledge but the used APIs. On the middle tier, the resource manager prototype is required to handle and relay information, as well as to execute algorithms. A resource manager will be placed on each compute node. On the highest management tier of the framework, the scheduling prototype will arrange and manage jobs transferred to the system. A sample algorithm will showcase and test the framework prototype. The goal of this thesis is to demonstrate and document the planning, implementation, future developments, and milestones for the scheduling framework. This includes the discussion of several already existing scheduling techniques as well as the concept of anticipation scheduling, a technique to anticipate algorithm runtimes through the help of different types of empirical information.