Phil Dykstra Chief Scientist WareOnEarth Communications, Inc. firstname.lastname@example.org 20 December 1999
Whether or not Gigabit Ethernet (and beyond) should support frame sizes (i.e. packets) larger than 1500 bytes has been a topic of great debate. With the explosive growth of Gigabit ethernet, the impact of this decision is critically important and will affect Internet performance for years to come.
Most of the debate about jumbo frames has focused on local area network performance and the impact that frame size has on host processing requirements, interface cards, memory, etc. But what is less well known, and of critical concern for high performance computing, is the impact that frame size has on wide area network performance. This document discusses why you should care, and about the largely ignored but important impact that frame size has on the wide area performance of TCP.
"Jumbo frames" extends ethernet to 9000 bytes. Why 9000? First because ethernet uses a 32 bit CRC that loses its effectiveness above about 12000 bytes. And secondly, 9000 was large enough to carry an 8 KB application datagram (e.g. NFS) plus packet header overhead. Is 9000 bytes enough? It's a lot better than 1500, but for pure performance reasons there is little reason to stop there. At 64 KB we reach the limit of an IPv4 datagram, while IPv6 allows for packets up to 4 GB in size. For ethernet however, the 32 bit CRC limit is hard to change, so don't expect to see ethernet frame sizes above 9000 bytes anytime soon.
Such local overhead can be reduced by improved system design, offloading work to the NIC interface cards, etc. But however you feel about these often debated local performance issues, it is the WAN that we are most concerned about here.
Throughput <= ~0.9 * MSS / (rtt * sqrt(packet_loss))So maximum TCP throughput is directly proportional to the Maximum Segment Size (MSS, which is MTU minus TCP/IP headers). All other things being equal, you can double your throughput by doubling the packet size! This relationship seems to have escaped most of the arguments surrounding jumbo frames. [Packet_loss may also increase with MSS size, but does so at a sub-linear rate, and in any case has an inverse square effect on throughput, i.e. MSS size still dominates throughput.]
In the local area network or campus environment, rtt and packet loss are both usually small enough that factors other than the above equation set your performance limit (e.g. raw available link bandwidths, packet forwarding speeds, host CPU limitations, etc.). In the WAN however, rtt and packet loss are often rather large and something that the end systems can not control. Thus their only hope for improved performance in the wide area is to use larger packet sizes.
Let's take an example: New York to Los Angeles. Round Trip Time (rtt) is about 40 msec, and let's say packet loss is 0.1% (0.001). With an MTU of 1500 bytes (MSS of 1460), TCP throughput will have an upper bound of about 8.3 Mbps! And no, that is not a window size limitation, but rather one based on TCP's ability to detect and recover from congestion (loss). With 9000 byte frames, TCP throughput could reach about 51 Mbps.
Or let's look at that example in terms of packet loss rates. Same round trip time, but let's say we want to achieve a throughput of 500 Mbps (half a "gigabit"). To do that with 9000 byte frames, we would need a packet loss rate of no more than 1x10^-5. With 1500 byte frames, the required packet loss rate is down to 2.8x10^-7! While the jumbo frame is only 6 times larger, it allows us the same throughput in the face of 36 times more packet loss.
A 9000 byte GigE packet takes the same amount of time to transmit as a 900 byte fast ethernet packet or a 90 byte 10 Mbps ethernet packet. So jumbo frames on gigabit ethernet at worse add less delay variation than 1500 byte frames do on slower ethernets. And no one is suggesting that slower ethernets use 9000 byte frames. As for queueing delay concerns, that could happen whether packets are large or small. If delivery QoS is required, than the routers need to implement some kind of priority or expedited forwarding, regardless of the packet sizes. Tiny frames (including 53 byte ATM cells) may be helpful when multiplexing lower bit rate streams, but they become increasingly ridiculous on gigabit and beyond links.
The economic and bandwidth arguments for GigE NAPs however are compelling. Several NAPs today are based on switched FDDI (100 Mbps, 4 KB MTU) and are running out of steam. An upgrade to OC3 ATM (155 Mbps, 9 KB MTU) is hard to justify since it only provides a 50% increase in bandwidth. And trying to install a switch that could support 50+ ports of OC12 ATM is prohibitively expensive! A 64 port GigE switch however can be had for about $100k and delivers 50% more bandwidth per port at about 1/3 the cost of OC12 ATM. The problem however is 1500 byte frames, but GigE with jumbo frames would permit full FDDI MTU's and only slightly reduce a full Classical IP over ATM MTU (9180 bytes).
A recent example comes from the Pacific Northwest Gigapop in Seattle which is based on a collection of Foundry gigabit ethernet switches. At Supercomputing '99, Microsoft and NCSA demonstrated HDTV over TCP at over 1.2 Gbps from Redmond to Portland. In order to achieve that performance they used 9000 byte packets and thus had to bypass the switches at the NAP! Let's hope that in the future NAPs don't place 1500 byte packet limitations on applications.
If you want high performance, the best network design advice that
I can give for a campus, regardless of the
networking technology being used is this: Every host in the
campus should have a path between it and the wide area network