Microkernel Technology Support to Parallel Applications

Our topics of research are user-level libraries and microkernels on shared memory multiprocessor architectures. The goal is the promotion of the cooperation between user-level and kernel-level resource schedulers to fit the needs of parallel applications. Mach 3.0 and Cthreads are our departure platform.

More specifically, we look for cooperation between kernel and user levels through processor and physical memory partitioning, affinity and memory conscious scheduling, event notifications to user level: upcalls, scheduler activations, user-level servers (policies) and kernel tunable scheduling mechanisms by processor sets and memory sets.

We are involved in the development of a new parallel execution environment implemented directly on top of the microkernel abstractions composed by servers (loader, pager, parallel i/o server and measurement server) and libraries (user threads and upcalls management). The goal consists in execute parallel applications by use of kernel extensions that enable new functionalities based on physical processors and memory space partitioning techniques, without compromising system reliability.

We have implemented user-level scheduling policies and mechanisms: hints, handoff, user level efficient context switch, priorities, locks with priorities or hints, non blocking and preemptive ready queues management and support to scheduling activations from kernel upcalls.

The validation of our designs is done through porting and evaluating the performance of parallel applications (like those in the SPLASH suite or based on the PARMACS macros) to run on standard Mach 3.0 and our environment.

• Marisa Gil, Toni Cortés, Angel Toribio, and José I. Navarro. Towards User-Level Parallelism with Minimal Kernel Support on Mach . In Workshop On Microkernel Technology for Distributed Systems, OSF , Grenoble (France), June 1993.

• Marisa Gil, Xavier Martorell, and José I. Navarro. The Enhancement of a User-level Thread Package Scheduling on Multiprocessors . In 2nd International Conference on Software for Multiprocessors and Supercomputers, Theory, Practique and Experience , pp. 390-398, Moscow (Russia), September 1994.

• Marisa Gil, Xavier Martorell, and José I. Navarro. The eXc Model: Scheduler-Activations on Mach 3.0. In 7th IAESTED/ISMM International Conference on Parallel and Distributed Computing Systems , pp. 5-10, Washington D.C. (USA), October 1995.