Cooperative hierarchical resource management for efficient composition of parallel software

Abstract

There cannot be a thriving software industry in the upcoming manycore era unless programmers can compose arbitrary parallel codes without sacrificing performance. We believe that the efficient composition of parallel codes is best achieved by exposing unvirtualized hardware resources and sharing these cooperatively across parallel codes within an application. This thesis presents Lithe, a user-level framework that enables efficient composition of parallel software components. Lithe provides the basic primitives, standard interface, and thin runtime to enable parallel codes to efficiently use and share processing resources. Lithe can be inserted underneath the runtimes of legacy parallel software environments to provide bolt-on composability - without changing a single line of the original application code. Lithe can also serve as the foundation for building new parallel abstractions and runtime systems that automatically interoperate with one another. We have built and ported a wide range of interoperable scheduling, synchronization, and domain-specific libraries using Lithe. We show that the modifications needed are small and impose no performance penalty when running each library standalone. We also show that Lithe improves the performance of real world applications composed of multiple parallel libraries by simply relinking them with the new library binaries. Moreover, the Lithe version of an application even outperformed a third-party expert-tuned implementation by being more adaptive to different phases of the computation.