HPC applications have many different phases: sometimes computing, sometimes communicating, sometimes doing file I/O. It is desirable to have maximum overlap with minimal interference between these activities. To provide optimal support for applications on HPC clusters, we propose the creation of Right-Weight Kernels (RWK). An RWK is a standard kernel to which we add specific enhancements for the needs of the HPC community: RWKs are a fusion of the ideas of both traditional heavy-weight cluster kernels and Light-Weight Kernels (LWK). The type of enhancements we create will be driven by measures of whole-application performance: measuring all phases of the application, not just one portion
Despite the considerable effort put into LWKs, there has been a lack of general acceptance and use. One reason for this is that while LWKs optimize one type of application activity, they remove many needed capabilities such as file systems, sockets, and security. Full-featured operating systems such as Linux have been quite successful, but there are concerns that the overhead of such operating systems is too high for some applications.
This research will determine if Linux and/or Plan 9 can be used to realize the benefits of light-weight kernels while maintaining the benefits of a full-featured operating system. We will use our expertise in Linux (the most popular and well supported operating system for HPC), Plan 9 (a distributed, light-weight operating system), and real applications to make these two operating systems competitive with or better than a custom LWK.
Publications:
Presentations: