Abstract

An important argument for the introduction of software managed clusters is that of scale: By constructing the cluster out of commodity compute elements, one can, by simply adding new elements, improve the reliability of the overall system in terms of performance and in availability. The limits to how far such a cluster can be scaled seems to be dependent on the scalability of its management software, which in its core has a collection of distributed algorithms to guarantee the correct operation of the cluster. The complexity of these algorithms makes them a vulnerable component of the system in terms of their impact on the overall scalability of the system.

This paper examines two of the distributed components of the Microsoft Cluster Service [8] that are most likely to have an impact on its scalability: the membership and the global update managers. The first sections of the paper will provide some general background on these distributed services and scalability issues. After that the algorithms used to implement these service are described in detail and an analysis of their impact on scalability is given. The scalability analysis is based on an off-line analysis of the algorithms as well as the results of on-line experiments on a cluster with a, in MSCS terms, large number of nodes.

• View the full text of this paper in HTML form and PDF form.

• If you need the latest Adobe Acrobat Reader, you can download it from Adobe's site.

• To become a USENIX Member, please see our Membership Information.