Solving the bandwidth puzzle
|
|||
 | |||
|
For example, 10G bit/sec is an exorbitant amount of bandwidth for most LANs. Today's fastest servers can't even fill a 1G bit/sec pipe. Yet in a few short years, 10G bit/sec Ethernet links are likely to be readily available.
How much bandwidth is enough? How can users know whether they really need the fastest LAN links on the market or if they can get by with the slowest? With all the hype and media coverage that surrounds any new technology introduction, it's hard to tell. The main thing to look at is utilization.
The traditional measure for shared Ethernet links has been if endstations are consistently using about 30% of the line, it's time for an upgrade. Any higher usage would make collisions and retransmissions likely as all the endstations try to talk at once.
A utilization of 30% is the threshold the Miami Herald looks for in its shared network, says Brian McNab, network administrator there. "Thirty percent gives us concern. At 40%, we definitely take steps, break the network up and switch more," he says.
When upgrading, companies can segment the network, move to switches or bump up the bandwidth. Fully switched networks can get by with utilization closer to 80% because endstations don't have to worry about collisions on their segments. On the network trunk, users should consider upgrading if utilization is greater than 33%, says Tom Nolle, president of CIMI, a consulting firm in Voorhees, N.J. Even if utilization isn't that high, users should upgrade the trunk connections if they notice some applications running slowly during certain periods, he says. A good rule of thumb, Nolle says, is that any switch should have a trunk connection that is 10 times the speed of the input ports. If the network isn't built that way from the start, the network manager can wait for symptoms to pop up and then upgrade the network accordingly.
Because the University of Pennsylvania has more than 100 buildings, the school looks for traffic problems only on the links between a central router and each of the buildings, says Deke Kassabian, technical director there. Links "with sustained utilization of more than 3M bit/sec or 4M bit/sec for long periods or those with peaks in their five-minute utilization averages of 7M bit/sec or more are candidates for a Fast Ethernet connection back to the routing core," he says.
When faster is slower
A common mistake is to upgrade links to the endstations, Nolle says. "Users are more likely to hurt themselves by upgrading ports," he says. "The worst performing networks we've seen have been those with high port speeds."
If users upgrade endstations to Fast Ethernet, it lets the endstations push much more traffic on the network, Nolle points out. This can lead to congestion in other places, which in turn leads to discarded packets and retransmissions, slowing overall performance.
The mixed blessing is that it is inexpensive to upgrade to 100M bit/sec ports. Today, 10/100M bit/sec Ethernet cards cost the same as 10M bit/sec-only cards, and 10/100M bit/sec switches are less than $150 per port. Many users purchase 10/100M bit/sec equipment but continue to run their networks at 10M bit/sec.
"Suddenly, people will find themselves capable of Fast Ethernet," says Dave Passmore, president of NetReference, a consulting firm in Sterling, Va. Next year, Gigabit Ethernet will be relatively inexpensive, too. Copper-based Gigabit Ethernet cards will cost as little as $250, and Gigabit Ethernet ports could be priced under $1,000 (NW, Sept. 14, page 12).
Users that don't want to pay for the next tenfold increase in bandwidth can take smaller steps. Technologies are available from the major switch and router vendors for binding several Fast Ethernet or Gigabit Ethernet lines into one logical link. These technologies are proprietary, requiring that users have the same vendor's equipment at both ends of the lines.
The IEEE has assembled a working group to standardize on one technique for trunking, but that work has just begun.
Watch carefully
Instead of running out and buying more than the network needs, though, the thing to do is watch utilization via Remote Monitoring probes often included in switches or routers. This technique works well for networks that mostly run traditional applications such as file transfers and e-mail, says John McConnell, president of McConnell Associates, a consulting firm in Boulder, Colo. Time-sensitive applications such as voice and video require a different approach.
For these, network managers need to measure latency and - just as important - the variation in latency across a network, McConnell says. A voice call, for instance, may require not just an average 64K bit/sec, but 64K bit/sec for every second or the call won't sound right. McConnell recommends using independent software products that measure response time and latency. Such software is available from NextPoint Networks, Response Networks and VitalSigns Software. The major network management platform vendors, such as Computer Associates, Hewlett-Packard and Tivoli, don't have this capability, McConnell says.
On the other hand, creating enough headroom with wide network links in itself can help ensure that real-time traffic doesn't hit any snags in the network. "Streaming audio and video are on the horizon - with several projects already underway - and I anticipate a phenomenal growth in this area over the next two years," says Mike Myrick, manager of networks at the University of Mississippi. The university is implementing a Gigabit Ethernet backbone to keep ahead of demand. The school is in a rare position and can afford to make the upgrade. Lucent Technologies is giving Ole Miss a trade-in credit on its Lucent ATM backbone switches toward Cajun P550 Gigabit Ethernet switches.
In the majority of cases, the hype about high speeds is irrelevant to day-to-day network operations. A CIMI survey found that fewer than 5% of LANs would benefit from a 100M bit/sec connection to the desktop. More than 70% of desktops are still in shared-media LANs, so companies can get some relief just by migrating to switches, Nolle says.
Numbers like that tend to put 10G bit/sec Ethernet in perspective.
