Work

Reconfigurable multiprocessor operating system kernel for high performance computing

Year: 1995
Type: dissertation
Abstract: Operating system kernels typically offer a fixed and limited set of mechanisms and primitives. However, the attainment of high performance for a variety of parallel applications may require the availability of variants of existing primitives or additional low-level mechanisms best suited for specific applications. Furthermore, such primitives must be able to cope with the changing requirements of dynamic applications. One approach to addressing this need is to offer an adaptable and extensible operating system kernel. This dissertation first motivates dynamic lock configuration by experimentation with synthetic workload emulating parallel applications, and by examining locking patterns of sample parallel applications. Then, it presents the implementation of reconfigurable locks that can be reconfigured statically and dynamically. Application level performance improvements gained from dynamic lock reconfiguration are demonstrated using a workload generator. We then present a model and implementation of synchronous adaptive multiprocessor locks which adapt themselves according to application specific adaptation policies to suit any application locking pattern. We evaluate such locks using formal analysis and by experimentation with several well-known multiprocessor applications. Furthermore, we compare the merits and limitations of synchronous adaptation vs. asynchronous adaptation using workload generator and sample multiprocessor applications. A similar model is used to implement reconfigurable thread schedulers which can be statically or dynamically configured. By experimentation with multiprocessor applications, we demonstrate how application-specific adaptation of such schedulers improves the scalability and performance of blocking locks. Finally, we generalize the model to propose a structure for adaptive kernel components and associated mechanisms which may be used for the development of high performance reconfigurable operating system kernels for parallel and distributed systems. These ideas were used to implement a reconfigurable micro-kernel called C-kernel on multiple parallel machines including a 32-node GP1000 BBN Butterfly, an 8-node SGI multiprocessor, a 10-node Sequent Symmetry parallel processor, and a 64-node Kendall Square Supercomputer.
Language: English
Open Access status: green
OpenAlex ID: W150975591
Cites:
Cited by: 2
Related to: 10