6.1 I/O subsystems and capacity planning for clusters

Performance and capacity issues that are relevant to standalone non-clustered Windows-based servers are even more so in a clustered environment. Specifically, a poorly designed I/O subsystem in a cluster is likely to be the single major culprit of a poorly performing cluster today. In a cluster, the I/O subsystems on other member nodes can also contribute to unnecessarily lengthy interruptions of service during a failover event. A good I/O subsystem design should take into consideration the clusterwide I/O capacity requirements for every business-critical service that is running on each node in the cluster. That will allow you to determine the I/O subsystem capacity required to support your critical applications when a node(s) goes offline and a failover is initiated, as would be the case when a failover occurs.

This chapter discusses issues surrounding the I/O subsystems used in today's computer systems and how they relate to clustered computers. While we have seen a dramatic increase of the clock rate for CPUs during the past few years, we have not seen the same level of performance gains in the I/O subsystem bandwidth. Multiprocessor-configured cluster nodes in which the CPUs are typically clocked at over a gigahertz are the norm today. Because of this, it has become painfully obvious that the I/O subsystems in today's computers are falling way below the performance curves of the processors available. This chapter presents our outlook for the next generation of I/O subsystems, which are targeted to deliver...