A layered, hierarchical approach to services for a single system image on heterogeneous clusters.

摘要

Workstation clusters are a powerful and often underutilized computing resource. Furthermore, many of these clusters are heterogeneous in nature. Research into heterogeneous computing has shown that heterogeneous collections of computing resources can outperform comparable homogeneous resource sets when the application set places varied demands on the computing nodes and the interconnection networks. Heterogeneous clusters have excellent price to performance ratios and can realize a very high level of aggregate performance, however the lack of single system image (SSI) support for heterogeneous clusters limits their range of application. Without SSI support, the programmer, administrator, and user are all painfully aware of the heterogeneity of the system.; While a distributed operating system (DOS) can realize most of the benefits of an SSI, there are a number of disadvantages in such an approach. The lack of software availability, the need to retrain users, and the generally large overhead involved in maintaining a DOS plague such implementations. Instead of the development of a DOS, this research takes a hierarchical approach to layering key SSI mechanisms within commercial off-the-shelf (COTS) heterogeneous clusters. A number of concepts are developed, analyzed, and presented enabling the presentation of an SSI within a heterogeneous cluster at a fraction of the overhead imposed by a DOS.; This research represents one of the first efforts addressing the SSI needs of heterogeneous clusters. The concept presented is proven by the realization of two of the benefits of an SSI. First, intelligent job scheduling concepts are developed and implemented, demonstrating that using experiential data regarding parallel and sequential resources, overall job performance can be improved over existing scheduling implementations such as NQS and SmartNet. Second, with a hierarchical, layered approach minimizing the software overhead of the cluster management system (CMS), the benefits of a solid cluster management and monitoring concept are extended to heterogeneous clusters. Currently, heterogeneous cluster administrators are generally forced to make changes or implement directives on each machine manually or using ad hoc scripts. The CMS's capability to issue global commands and monitor cluster status greatly reduces the burden of administration.