Tera-scale Computing

Runtime Environment for Tera-scale Platforms

CONCLUSION

A tera-scale platform has a number of fundamental differences from a traditional SMP system, which requires that the system software stack be redesigned to provide an effective and scalable runtime environment. In particular, the runtime environment must provide good support for fine-grain parallelism, support new concurrency abstractions that ease parallel programming, and support heterogeneous platforms and applications.

This paper described how McRT addresses the challenges of a many-core environment. To enable efficient fine-grain parallelism, McRT replaces many of the OS-level services with user-level primitives. Our results show that this enables a very scalable runtime stack that scales to more than 64 HW threads.

To ease parallel programming, McRT provides a high-performance TM library that supports a language-level atomic construct. TM provides several software engineering benefits compared to locks such as deadlock freedom, scalable composition, and failure atomicity. Additionally, our results show that transactions achieve the performance of fine-grain locking, yet allow coarse-grain synchronization.

McRT supports heterogeneity by dividing the platform into independent scheduling domains. These domains can be mapped to different hardware resources, and applications can be scheduled on the domain that best fits their requirements. In addition, McRT also supports a number of configurable runtime policies that allow it to adapt to different applications.

Back to Top