nStor 4100FS, User Manual

Page: 43 / 168

Chapter 2 - Theory of Controller Operation
Minimizing Downtime for Maximum Data Availability
31
High availability is not the only benefit of clustering. In some cases, users may
see advantages in the areas of performance, scalability, and manageability. In
reality, you can expect a 1.6x (80% efficiency) to 1.8x (90% efficiency)
performance increase as you go from one node to two nodes. Going from one
node to a four node cluster generally yields a 2.5x or 3x performance boost.
However, the cluster performance is application dependent. For example, READ
operations may yield a 1.8x performance increase going from one to two nodes,
but in a WRITE intensive application, you may only see a 1.4-1.6x improvement.
Although clusters seem to be relatively simple, they involve complex technology
that can be implemented in a variety of ways. The number of nodes supported
and type of interconnection used, and a number of other features differentiate
cluster implementations. One area of implementation is the manner in which
distributed lock manager is implemented. Some perform this at the user level and
others in the kernel, with the latter enhancing performance.
In addition to the differing features you should consider the following:
Does the cluster:
• have the ability to hot load new nodes without bringing down the
whole cluster?
• provide automatic or manual failover?
• load balance?
• use a journalized file system?
• provide a fast cluster failover?
• allow for the nodes to be geographically located?
4100FS.book Page 31 Tuesday, July 8, 2003 4:35 PM