Statistically, most Internet communications pass through an Ethernet network somewhere on the way. This is why Ethernet's 1,500 byte MTU has become the standard MTU, and the main Internet bottleneck. Due to the constant increase in Ethernet spped, and to the fact that Gbit Ethernet has no standard MTU, the discussions of increasing the Ethernet frame size have been rapidly increasing.

Like everything else, Ethernet should evolve with time. High-speed network users have long felt the limitations of Ethernet's small frame sizes, and are among the biggest proponents of Jumbo Frames - extended Ethernet frames that range in size from the standard 1,518 bytes up to 9,000 bytes.

The 1,518-byte frame size was designed to protect against the high bit error rates of yesterday's physical-layer Ethernet components. But computer processing power has increased by an order of magnitude, and the use of switched Ethernet over unshielded twisted pair or fiber media has significantly lowered Ethernet errors.

More importantly, the speed and capacity of today's Ethernets are pushing the processor limits of most installed servers, and more data is being transferred between servers.

For these reasons, extending Ethernet's frame size to reduce server overhead and increase throughput has become an attractive and logical option. Independent tests have verified that the use of Jumbo Frames can deliver a 50% increase in throughput with a simultaneous 50% decrease in CPU utilization.

Frame size has a major impact on server performance. A typical server takes approximately 1200 CPU cycles to process the TCP/IP headers of a single Ethernet frame. Given that it takes 80,000 maximum-sized Ethernet frames (1518 bytes) to fill a Gigabit link, millions of packets every second would interrupt the server's CPU processing tasks. Even the most robust multiprocessor servers can't keep up with all these interruptions.

Users clearly recognize the value of bigger frames, but to get them until now, they had to turn to less-mainstream technologies such as FDDI, High Performance Parallel Interface and Fibre Channel. The problem is that these technologies lack the installed base, cost-effectiveness and seamless interoperability of Ethernet.

So how large should an Ethernet frame be? Ethernet's 32-bit cyclic redundancy check is effective for detecting bit errors at frame sizes under 12,000 bytes, thereby drawing a logical upper limit. Within that, the optimum large frame size can be determined by an application's block size. For example, Network File System (NFS) transfers data in 8,192-byte blocks. So adding room for headers, an attractive maximum Ethernet frame size for NFS applications is 9,000 bytes.

Jumbo frames maintain the same media access control (MAC), frame structure, and frame check sequencing mechanism used for traditional Ethernet frames. Only the payload portion of the frame is extended. Therefore, no modifications would be required for delivering IP over Ethernet, and the new MTU would automatically be used.

Alteon Networks (www.alteonwebsystems.com) which is one of the first and largest promoters of increasing Ethernet frame size, claims the following Jumbo Frame benefits:

• 300% increase in throughput combined with a 50% reduction in host CPU processing.
• Up to 85% reduction in packet processing.
• Jumbo Frames are configurable up to 9000 bytes.
• Can be used in conjunction with IEEE 802.3Q VLAN tags.
• Uses standard Ethernet error correction.
• Increases host protocol stack efficiency.

Alteon is planning to submit Jumbo Frames to the IEEE in hopes of fostering a working group and an eventual standard.

Two basic approaches exist:

• On a port by port basis, where everything "downstream" from a given port is known to support jumbo frames.
• Using 802.1q Virtual LANs, where jumbo frame and non-jumbo frame devices are segregated to different VLANs.

Note that if jumbo frames cross a non-supporting router, the router will break the frames down into 1,500-byte frames.