Jumbo frames? No!

Let's get this out of the way: Increasing the size of Ethernet frames can improve a host computer's performance. But how big is the gain? What price must we pay for the improvement? Are there alternatives?

Although proponents claim larger packets improve performance "on the wire," the impact is relatively insignificant. Compare the efficiency of a 9,000-byte large-packet system with a standards-based 1,500-byte system. The standard packet gets 1,500 data bytes out of 1,538 bytes of frame and overhead, or 97.5% efficiency. The nonstandard packet gets 9,000 data bytes out of 9,038 bytes, or 99.6% efficiency. To put it another way, the difference in time required to send a 1M-byte file is only 0.1 msec.

These increases in efficiency and performance are insignificant compared with other delays and negative factors in the large-packet system. Furthermore, substantial costs are associated with diverging from the standards on which your network is based.

The greatest potential benefit of larger packets is improving the host computer's efficiency. In the host, the CPU interacts with the network interface card (NIC) to move packets from the wire to the application. Large packets minimize the CPU's workload by decreasing the number of transfers across the bus, the interrupt stack and the protocol stack. Thus, the challenge is to make the packet appear large in the system to offload the CPU while remaining standards-based on the wire.

To achieve this, today's NICs couple fast bus types with advanced direct memory access schemes to move data efficiently to the host's memory. Techniques such as transmit chaining and interrupt consolidation of received packets serve to group packets together to simulate a larger packet and minimize CPU interrupts. To maximize the protocol stack's efficiency, NICs are taking over many CPU-intensive tasks such as TCP/IP checksumming. Additional schemes will soon be available that will move more protocol stack functions, such as fragmentation and defragmentation of the IP packets, to the NIC. Together, these features provide the benefit of larger packets while still complying to standards and working seamlessly with the installed equipment.

Even if larger packets offer improved performance over other approaches, they exact a cost in terms of interoperability.

To get maximum benefit, large-packet capability would have to be propagated across the whole network. This would require that all existing devices, including NICs, hubs and switches, be replaced.

It is possible for large-packet networks to coexist with standard Ethernet networks by installing both NICs and switches that can handle multiple packet sizes. However, these features increase the cost, complexity and management burden of the devices, and interworking between different packet sizes can negatively affect performance.

Finally, at what point do we say, "Stop, that's large enough?" Is it at 4,000 bytes, as some vendors said when Ethernet ran only at 10M bit/sec; 9,000 bytes, as some assert today; or will the fashionable number be 64,000 bytes next year?

Application and system requirements will never stop changing. The right decision is to resolve the problem where it occurs - in the host, not on the network.