Tera-scale Computing

Runtime Environment for Tera-scale Platforms

RELATED WORK

Our work was inspired by previous projects in the areas of language systems, operating systems design, and high-performance computing. Several operating systems have explored issues related to SMP scalability [3, 4, 17, 23, 26]. Disco [9] and K42 [18] have explored the design of scalable operating systems. Currently, operating systems such as Linux* and MS Windows* treat CMP architectures as an SMP on a chip. While the McRT framework arose out of our experience of these projects, this paper shows why we still need to rethink the system design to fully enable a scalable tera-scale environment.

The threading and synchronization primitives provided by McRT are similar to those seen in traditional user-level threading packages such as pthreads [19], NGPT [25], NPTL [10] and Capriccio [6]. The McRT scheduler is comparable to user-level schedulers that have been explored in the context of kernels such as L4 [20], Exokernel [12], Flux [13], and SPIN [7], but it is more lightweight and configurable. The McRT runtime is also comparable to language runtimes such as CILK [8] and OpenMP [21], but unlike these runtimes it is platform-neutral.

Finally, the specific mechanisms used in the McRT-STM [27] and the McRT memory manager [16] are described elsewhere, while the compiler integration is discussed in [2, 28].

Back to Top