Parallel programs executing on non-dedicated SMPs should be adaptive, that is, they should be able to execute on a dynamically varying environment without loss of efficiency. This paper defines a unified set of services, implemented at the operating system level, which can be used to embed adaptability in any thread-based programming paradigm. These services meet simultaneously three goals: they are highly efficient; they are orthogonal and transparent to the multithreading programming model and its API; and they are non-intrusive, that is, they do not compromise the native operating system's resource management policies. The paper presents an implementation and evaluation of these services in the Linux kernel, using two popular multithreading programming systems, namely OpenMP and Cilk. Our experiments show that using these services in a multiprogrammed SMP yields a throughput improvement of up to 41.2%.
展开▼
