10.5 The Effects of TCP on Application Performance

In Chapter 9, Ethernet and Fibre Channel were cited as examples of layer-two protocols, and IP was cited as an example of a layer-three protocol. Where end systems are involved, additional networking layers may be used to provide important facilities, such as data integrity. IP uses TCP to guarantee end-to-end data integrity and automatic flow control. The term TCP/IP is used to indicate that both protocols are employed.

Initially, it was thought that TCP/IP could not be used for storage networking applications, as it would place an excessive load on the end systems ”the servers, disk arrays, and tape libraries. TCP requires a large amount of processing power and, until recently, had been handled by the server CPU, limiting the amount of TCP a box can run and still have useful cycles left over.

However, early in 2002, TCP offload adapters became commercially available and changed this workload balance. These specialized adapter cards include a hardware-based TCP offload engine that handles IP storage protocols at full wire speed over a Gigabit Ethernet interface. The loading on the server CPU is no greater than that of a Fibre Channel HBA, so TCP/IP has become a viable storage networking protocol. Vendors now are sampling 10 Gbps TCP offload engines and have promised to deliver IP storage adapter cards that have 10 Gbps Ethernet interfaces during 2003.

Although the limitations of the TCP computational intensity have been overcome , another limitation ”flow control ”still can cause network performance problems. TCP was designed for use in very large networks and, indeed, it is used today by hundreds of millions of users of the world's largest network, the Internet. To ensure fairness and accessibility by millions of simultaneous users, TCP uses a fairly conservative flow-control algorithm.

In effect, each user is granted a modest initial data flow rate that increases as acknowledgments are returned from the destination end system when the IP packets are received without errors. If packets are dropped or contain errors, the sending system must fall back to a much lower flow rate and start once again to ramp its rate. In a well-designed network, TCP has little impact on performance, but in a congested or error-prone network, the overall performance can be cut by more than half.

To improve the operation of TCP/IP, a specification is being developed for a more aggressive version of TCP. It sometimes is called fast-start TCP, as its initial flow rate will be greater, and it will recover more quickly from lost packets. Fast-start TCP is expected to reach volume production by about 2005.