Adaptive Operating System Abstractions: a Case Study of Multiprocessor Locks

Operating system kernels typically ooer a xed and limited set of primitives and underlying mechanisms for use by application programs. However, the attainment of high performance for a variety of parallel applications may require the availability of additional primitives or of variants of existing primitives best suited for speciic applications. Furthermore, operating system mechanisms must also cope with changes over time in the behavior of individual applications. One approach to addressing these needs is to construct conngurable operating system kernels, where conngurability of operating system mechanisms permits their change to suit varying application needs. This paper addresses operating system connguration by considering speciic abstractions. Speciically, we rst motivate dynamic connguration by experimentation with conngurable multiprocessor locks, using synthetic workloads emulating parallel applications. Evaluations of locking patterns of sample parallel codes demonstrate the utility of dynamic lock conngu-ration. Second, an implementation of dynamically conngurable locks results in performance improvements for synthetic multiprocessor programs using them. Third, a generalization of the model and implementation of conngurable locks, termed synchronous, adaptive locks, results in locks that dynamically adapt their behavior to suit the locking patterns of a range of parallel programs. Adaptive locks are implemented and evaluated with application programs on several parallel machines, including a 32-node GP1000 BBN Butterry and a 64-node Kendall Square Supercomputer.