At 08:41 AM 1/28/02, Peter van Oene wrote: >Related to the subject of ethernet utilization, the following paper, while >dated, provides some interesting insight. > >http://www-2.cs.cmu.edu/~srini/current_class/readings/B+88.pdf
Yes, that's a terrific paper about Ethernet performance. I wonder if it's legal for it to be at the CMU site? It's a DEC (now Compaq!? ;-) paper. Its "official" residence on the Web is here: www.research.digital.com/wrl/publications/abstracts/88.4.html It debunks the theory about a shared Ethernet segment going into meltdown at 40 percent utilization. Allegedly, beyond this limit, the collision rate becomes excessive. But the real answer is "it depends." Consider the case of an Ethernet segment that is shared by only two stations -- a client that sends requests and a server that responds after receiving requests. In this case, it is not a problem if network utilization exceeds 40 percent. There are no collisions because the server and client never try to send at the same time, so the 40 percent rule, which is concerned with collisions, does not apply. The load should approach 100 percent unless the client or server are slow. The 40 (actually 37) percent number was first reported by Metcalfe and Boggs in a 1976 paper that described the original 2.94-Mbps Ethernet. The paper described a simple model that was not meant to resemble real-world networks. The 37 percent limit was also discussed in studies done by the IEEE comparing CSMA/CD to token passing. Token passing makes a node wait for a token before sending. At modest loads, this wait means that token passing results in more delay than CSMA/CD. However, at around 37 percent utilization on a medium shared by 50 stations using 128-byte frames, Ethernet frames experience more delay than token ring frames because of collisions. Note that this is a specific example, not a generic "rule." After many years of hearing the "37 percent rule," Boggs and two other researchers published the 1988 "Measured Capacity of an Ethernet: Myths and Reality" paper that Peter is referencing. And, one last point. None of this is at all relevant on a full-duplex link, which is point-to-point with each side having a dedicated transmit and receive circuit. Collisions should never occur. A station should be able to use 100 percent of the capacity of its transmit circuit. Room should be left for bursty traffic, so an average utilization should be 70 or 80 or somewhere around there. It's not an exact science. ;-) But the 40 percent rule certainly does not apply in this case. Priscilla >At 09:06 AM 1/27/2002 -0500, you wrote: > >Priscilla is absolutely right, its a fuzzy question. I have just two > >things to add. > > > >If the network is "mainly LAN" that suggests that there are some wide > >area links. Because wide area links are usually slower than local area > >media and are used by lots of users, congestion on them is definitely > >worth checking. Also, if there are complaints about response to a > >distant resource, you should also look at delay (which you could check > >with ping). If people are complaining about the time required for a > >complex interaction (one requiring many packets in both directions), its > >possible that a moderate amount of delay can be a problem. > > > >Second, ethernet is different from most media. Because of the way > >ethernet works utilization numbers require some interpretation, at least > >for half duplex operation. This necessarily includes segments used by > >more than two hosts, if you have any. There are no definite definitions > >of what ok and whats too much, but utilization above 40% (and maybe > >less), IMHO, should be considered congested. > > > >HTH, > > > >Jason > > > >Priscilla Oppenheimer wrote: > > > > > At 06:36 PM 1/25/02, Doug Korell wrote: > > > > > >>I have checked individual switches and routers for utilization before but > > >>when asked what the average utilization of an entire network (mainly LAN) > > >>is, what exactly makes up this figure? I am working on getting a packet > > >>sniffer which I know will help take all the variables and give me an >answer > > >>but is there a way to do it without one? How about SNMP queries? If >anyone > > >>can help explain this or knows of a good website, please let me know. > > >> > > > Thanks. > > > > > > That's a rather old-fashioned question. It used to make sense on a shared > > > LAN. You could put a Sniffer or RMON Probe in a shared hub and get a > > > measurement of how much of the overall, shared 10-Mbps capacity was in >use > > > on the LAN. > > > > > > In these days of microsegmented, switched networks, you can't do that > > > easily. You can only monitor the switch ports that you mirror. > > > > > > Each switch port provides full capacity, usually 100-Mbps full duplex. > >(You > > > would have to know if that's true for your network.) Overall capacity is > > > the number of ports times the speed. Overall utilization would be the > > > aggregate of each port utilization divided by the overall capacity, I > > > guess. (But people don't actually tend to make that calculation.) > > > > > > Another capacity issue is the backplane speed of the switches and routers > > > in use. That could actually be more of a bottleneck than overall LAN > > > capacity. > > > > > > Did a pointy-haired boss type ask you to make this measurement? I'm >afraid > > > you might have to explain that it doesn't make sense. Work with them to > > > specify which individual LAN ports need monitoring, rather than trying to > > > find an overall number. The ports that you should monitor are any ones > >that > > > aggregate traffic. Check the utilization on trunk lines and links that go > > > to mission-critical servers. Also, check utilization on an end-user port > > > while doing some typical processes, including logging into the network. >It > > > might also make sense to check other performance metrics such as response > > > time. > > > > > > Hopefully others will respond too in case I have a blind spot with >regards > > > to this, but my initial thought is that this is not the right performance > > > measurement to be considering for a modern LAN. > > > > > > Priscilla > > > ________________________ > > > > > > Priscilla Oppenheimer > > > http://www.priscilla.com ________________________ Priscilla Oppenheimer http://www.priscilla.com Message Posted at: http://www.groupstudy.com/form/read.php?f=7&i=33472&t=33256 -------------------------------------------------- FAQ, list archives, and subscription info: http://www.groupstudy.com/list/cisco.html Report misconduct and Nondisclosure violations to [EMAIL PROTECTED]
