A Flexible Resource Management Architecture for the Blue Gene/P Supercomputer

Blue Genereg/P are a massively parallel supercomputer intended as the successor to Blue Gene/L. It leverages much of the existing architecture of its predecessor to provide scalability up to a petaflop of peak computing power. The resource management software for such a large parallel system faces several challenges, including system fragmentation due to partitioning, presenting resource usage information using a polling or event driven model, and acting as a barrier between external resource management systems and the Blue Gene/P core. This paper describes how the Blue Gene/P resource management architecture is extremely flexible by providing multiple methodologies for obtaining resource usage information to make scheduling decisions. Three distinctly separate resource management services can be described. First, the Bridge API, a full-featured API suitable for fine tuning scheduling and allocation decisions. Second, a light-weight allocator API for allocating resources without substantial development costs. And lastly, configuring the system into static partitions. Job scheduling strategies utilizing each of the methods can be discussed.