In response to the July 2 Perspectives column by Jon Oltsik, "Sizzling storage? Just wait":
I read your article with great interest. Of particular note to me was your opinion on the demise of the channel extender and the emergence of Internet Protocol (IP) as the standard for storage over distance.
Because I believe we need a replacement technology for channel extenders, I was glad to see someone point out this important element in this increasingly critical aspect of data center networking. However, IP is not the best solution for distance storage, especially for business continuance and disaster recovery applications like remote mirroring and backup.
All IP networks today are built on the premise that the applications that they support can tolerate a certain amount of loss. Although you can engineer a network to reduce this loss to a small amount, it is always present. This normally isn't a concern, because upper-layer protocols like TCP (Transmission Control Protocol) will look after a retransmission. In addition, the benefits of sharing an IP network with many applications usually outweigh the small performance penalty of a retransmission. When dealing with storage, however, a different story emerges.
Business continuance-focused storage applications are very different; high bandwidth, low latency and a point-to-point topology place unique demands on a network. It is not atypical for this type of application to average several hundred megabits per second of bandwidth over metro distances. Unfortunately, the inherent loss of IP makes this level of throughput unobtainable. For example, it can be easily shown that a typical IP network with a packet delivery ratio of 99.9% and a round-trip latency of 50 milliseconds can only support roughly 5mbps of throughput for a remote disk-mirroring application.
Information technology end users already know about the performance impact that TCP has on end systems--the emergence of the TCP offload engine (TOE). TOEs do fix the central processing unit utilization impact of TCP but do nothing to fix the inherent throughput limitation of TCP itself. The only way to fix this problem is to lease a private line and create a private IP network for the storage extension. This is a very costly proposition and defeats the purpose of using IP in the first place.
Matthew Williams
Frisco, Texas