Resource management for task-based parallel programs over a multi-kernel

Trying to attack the problem of resource contention, created by multiple parallel applications running simultaneously, we propose a space-sharing, two-level, adaptive scheduler for the Barrelfish operating system. The first level is system-wide, running close to the OS ’ kernel, and has knowledge of the available resources, while the second level, integrated into the application’s runtime, is aware of its type and amount of parallelism. Feedback on efficiency from the second-level to the first-level, allows the latter to adaptively modify the allotment of cores (domain), intelligently promoting space-sharing of resources while still allowing time-sharing when needed. In order to avoid excess inter-core communication, the system-level scheduler is designed as a distributed service, taking advantage of the message-passing nature of Barrelfish. The processor topology is partitioned so that each instance of the scheduler handles an appropriately sized subset of cores. Malleability is achieved by suspending worker-threads. Two differ-ent methodologies are introduced and explained, each suitable for distinct programming models and applications. Preliminary results are quite promising and show minimal added overhead. In specific multiprogrammed configurations, initial ex-periments proved significant performance improvement by avoid-ing contention. 1