goodloe@xenon.UUCP (Tony Goodloe) (04/27/88)
I have heard (from some forgotten source ) that the access mechanism used on Ethernet will not allow a REAL csma/cd network to achieve more than about 50%-60% network utilisation. Can someone point me in the direction of a paper, article, personal experience, etc. that supports or refutes this claim? Many thanks.
donegan@stanton.TCC.COM (Steven P. Donegan) (04/30/88)
In article <47@xenon.UUCP>, goodloe@xenon.UUCP (Tony Goodloe) writes: > I have heard (from some forgotten source ) that the access mechanism used on Ethernet will not allow a REAL csma/cd network to achieve more than about 50%-60% network utilisation. Can someone point me in the direction of a paper, article, personal experience, etc. that supports or refutes this claim? Many thanks. > I have seen these types of questions so many times that I feel it necessary to respond. A CSMA/CD media protocol is non-deterministic. If only 2 stations are connected via the cable and a transfer is started then the full bandwidth available can be utilized. As the number of stations, distance between stations, cable plant inconsistancies (I've never seen a perfect ethernet segment) and many other factors change it becomes impossible to determine real throughput by prediction. Only real world testing will give you valid results and due to the nature of most networks (constant change) the results of any tests are only valid for a limited time. With the exception of diskless workstation nets I have seldom seen real network utilization exceed 20-30%. Diskless workstation environments can eat up the bandwidth incredibly fast (6 - 10 SUN stations can saturate a segment if you try :-))). Repeaters and bridges can make this question even more impossible to answer. I find continuous performance monitoring the only reliable way to determine the state of your network. The items to watch for are an increasing (slowly, due to traffic; not quickly due to hard problems) level of collisions, crc errors, too short or too long packets. I have watched WD's ethernet grow from 1 building and <100 stations to 16 buildings, 2800+ stations (station=end device, not node - nodes went from 3 to roughly 200 as of today) and have run into many different performance problems, all of which could be solved by segmentation and learning bridge implementations. An ethernet is a synergy of all it's respective components, only extremely careful design and monitoring can keep it running properly. After all the above, it's still far easier than manageing a broadband net :-) -- Steven P. Donegan Sr. Telecommunications Analyst Western Digital Corp. donegan@stanton.TCC.COM
david@ms.uky.edu (David Herron -- One of the vertebrae) (05/02/88)
In article <47@xenon.UUCP> goodloe@xenon.UUCP (Tony Goodloe) writes: >I have heard (from some forgotten source ) that the access mechanism used on >Ethernet will not allow a REAL csma/cd network to achieve more than about >50%-60% network utilisation. Can someone point me in the direction of a paper, >article, personal experience, etc. that supports or refutes this claim? >Many thanks. Well, there's an existence proof that you can get 80-85% utilization, from ONE host at that. Van Jacobsen & Mike Karels has been working on the TCP/IP code in 4.3BSD and fine tuned it to really screem. They've been getting a bit over 8mbit/sec through put from otherwise unmodified Sun 3/50's over ether. The only problem was insufficient buffering in some VMS hosts on their net ... the VMS hosts crashed for some odd reason, probably nobody ever thought anybody could tie up an ethernet so well :-) -- <---- David Herron -- The E-Mail guy <david@ms.uky.edu> <---- or: {rutgers,uunet,cbosgd}!ukma!david, david@UKMA.BITNET <---- Windowing... The Amiga has windowing. The Mac has windowing (echoes of <---- Jonathan Livingston Seagull: "Just flying? A mosquito can do that much!").
mar@athena.mit.edu (Mark A. Rosenstein) (05/03/88)
Yes, there are examples of a pair of hosts using 80% of an ethernet by themselves. But that does not mean that 10 hosts, each doing different things and talking to different peers, could use that much of the net. When only 2 hosts try to use the entire network, the timing is very tightly interleaved so that each sends another packet while the other is listening. But with several pairs of hosts attempting to use the same net, they are likely to have collisions if each pair is attempting to use 80% of the net. In this case several hosts will be waiting for the end of one packet so that they can start sending, and may all start at the same time, causing a collision, then backoff and retry. I don't have any numbers to back this up, but I suspect that the utilization will go down somewhat as more hosts are added, although it would level off pretty quick. I mean that one host could use the entire bandwidth, all it ever needs to do is send. Two cooperating hosts could interleave packets and use nearly the entire network. More pairs will collide a certain amount, but once you've added a certain number of hosts, you will continue to get, say, 75% network usage. If anyone has any simulations, or better yet, real world numbers to back this up, I'd like to hear them. -Mark mar@athena.mit.edu ...!mit-eddie!athena.mit.edu!mar Variables won't and Constants aren't
mogul@decwrl.dec.com (Jeffrey Mogul) (05/03/88)
In article <5038@bloom-beacon.MIT.EDU> mar@athena.mit.edu (Mark A. Rosenstein) writes: > >Yes, there are examples of a pair of hosts using 80% of an ethernet by >themselves. But that does not mean that 10 hosts, each doing >different things and talking to different peers, could use that much >of the net. > [...] >If anyone has any simulations, or better yet, real world numbers to >back this up, I'd like to hear them. > -Mark Dave Boggs has been doing the real-world measurements (in the sense that he's using real computers and networks; it's hard to find real applications that can saturate an Ethernet). We have written a paper, together with Chris Kent, that describes these measurements, summarizes the theory literature, and explains why numbers like "37%" are at best a misunderstanding of the theory. The paper has not yet been published; we will let people know when it is. Some preliminary results: with 1536-byte packets, 1 of our hosts can get 92% of the (10Mbit/sec) bandwidth; 2 hosts together get 99.8%; 22 hosts get 97%. I.e., for large packets the ethernet really is 10 Mbits/sec. Theory says that smaller packets lead to lower utilization, because collisions occur more often. For minimal-length (64-byte) packets, the host we are using could not send them fast enough to get more than 26% utilization with 1 host on the air (that's 3700 packets per second; this has nothing to do with collisions). We get 75% with 3 hosts, 81% with 10 hosts, and it appears to stay at 81% through 22 hosts (11500 packets/sec.) Of course, the delays seen by each host get worse as the number of transmitters go up, but the Ethernet does not suffer congestive collapse even when an awful lot of iron is shouting really loud.
hedrick@athos.rutgers.edu (Charles Hedrick) (05/03/88)
I saw the Ethernet work as Xerox PARC working papers, but Bertsekas and Gallaher reference an article by Metcalfe and Boggs, "Ethernet: Distributed Packet Switching for Local Computer Networks", Commun. ACN [sic - presumably ACM], 395-404. As I recall, the Xerox studies claimed 90% loading. We've heard recently from Van Jacobson that he has gotten over 8MB/sec sustained throughput on an FTP transfer over an Ethernet. Of course this is probably more than 80% Ethernet utilization because of various network overheads. On the other hand, I don't think you'd want to load a network to even 50%. General statistical rules of thumb suggest that when you've got lots of users contending in a random way, your peaks are going to be an order of magnitude above your average, so in order to avoid slowing people down when they need bursts of speed (which is quite typical in the normal Ethernet applications), you don't want steady state loading (i.e. averages over a minute or so) to be above 10%, even during your peak periods. The same consideration would be true for token ring or any other type of medium. That is, the effects due to the very random character of the traffic are likely to dominate the effects due to differences between CSMA, token ring, etc.
bc@halley.UUCP (Bill Crews) (05/04/88)
In article <5038@bloom-beacon.MIT.EDU> mar@athena.mit.edu (Mark A. Rosenstein) writes: > In this case several hosts will be >waiting for the end of one packet so that they can start sending, and >may all start at the same time, causing a collision, then backoff and >retry. Well, sort of. There is no listening for the end of a packet. Just randomizing and trying again and listening for a collision. -bc -- Bill Crews Tandem Computers bc@halley.UUCP Austin, Texas ..!rutgers!im4u!halley!bc (512) 244-8350
holt@inco.UUCP (Mark Holt) (05/04/88)
In article <5038@bloom-beacon.MIT.EDU>, mar@athena.mit.edu (Mark A. Rosenstein) writes: > More pairs will collide a certain amount, but once you've added a > certain number of hosts, you will continue to get, say, 75% network > usage. > > If anyone has any simulations, or better yet, real world numbers to > back this up, I'd like to hear them. > -Mark > > mar@athena.mit.edu ...!mit-eddie!athena.mit.edu!mar Also: (I tried to mail directly ... same old story) News-Path: hadron!sundc!seismo!uunet!ingr!b11!xenon!goodloe > I have heard (from some forgotten source ) that the access mechanism > used on Ethernet will not allow a REAL csma/cd network to achieve more > than about 50%-60% network utilisation. Can someone point me in the > direction of a paper, article, personal experience, etc. that supports > or refutes this claim? Many thanks. > Reference : Stallings, William, Local Networks - An Introduction, Macmillan, 1987, Chapter 9 - Network Performance. The above is a readable, if inconsistant, discussion on some techniques for approximating performance. The book points out that the chief factors affecting the performance of a LAN or a HSLN (Hi Speed Local Net) are: quote Capacity Propagation Delay Number of bits per frame Local network protocols Offered load Number of stations end quote Stallings then goes into greater detail on how some of these parameters affect performance. For example, he shows (by way of derivation of formulas and use of graphs) how the number of bits per frame and the number of stations active can impact the capacity vs actual data rate (ideal : actual data = capacity) curve. Also, he goes on to show how much more efficient (maximum potential utilization) a token ring topology is than CSMA/CD. Now, before I get torched by people crying "real world, real world", Stallings notes that the results he shows "...depend on the nature of the assumptions made and do not reflect accurately the nature of the real-world load." However, I think he has built a pretty good argument... :-) . Hope this helps, my apologies if I have misrepresented Mr. Stallings. mark -- /*--------------------------------------------------------------* * Mark L Holt * * McDonnell Douglas - Inco, Inc. * * ...!seismo!sundc!hadron!inco!holt * * "If you ain't the lead dog, the scenery never changes" * * DISCLAIMER: The opinions expressed are my own and in no way * * reflect the views of McDonnell Douglas or its subsidiaries. * *--------------------------------------------------------------*/
wsmith@umn-cs.cs.umn.edu (Warren Smith [Randy]) (05/11/88)
Sorry for the delay, I'm just catching up on my news here... In article <47@xenon.UUCP> goodloe@xenon.UUCP (Tony Goodloe) writes: >I have heard (from some forgotten source ) that the access mechanism used on >Ethernet will not allow a REAL csma/cd network to achieve more than about >50%-60% network utilisation. Can someone point me in the direction of a paper, >article, personal experience, etc. that supports or refutes this claim? >Many thanks. And then added the following (not all on one line this time): >I feared that my posting didi make it to the net in its entirety, and I can >see now that it must not have. My question was "what maximum utilization >can one expect to see on an Ethernet?" I know that a csma/cd net is >non-deterministic, that's why the question exists, otherwise I could >determine it :) I was looking for more sources like "Performance of an >Ethernet Local Netwotk: A Preliminary Report" by Shoch and Hupp or >"Performance Analysis of CSMA/CDD" by Tobagi and Hunt. I haven't made it >through Tobagi and Hunt, but according to Shoch and Hupp they have seen >a maximum of ~98% utilization down to maximum ~60% utilization depending on the >test conditions. Read the article for more info. I was looking for more >info like this, or real live experience. Our net gets up to about 35% >and we do have individual nodes capable of more. What I am curious about >is what people have seen or caculated (using statistical means (no pun >intended)) on other nets with lots of these beasts on them. Maybe this >posting will make it out in one piece. There are numerous articles and books that review the basic analysis of Ethernet (see Boggs and Metcalfe from CACM sometime in 1976 for the first analysis). Almost every computer networking text includes some portion of a CSMA analysis. Aside from Shoch and Hupp there is very little published on measured Ethernet performance. There was a paper in 1986 by Raj Jain and Shawn Routhier (IEEE Selected Areas of Comm) on some measurements of a 10Mbps token ring at MIT - you may want to track down some of the papers they reference. If you know of other papers in this area please tell me. Here are some of my observations (assuming a network with a large number of stations that aren't synchronizing to avoid collisions): 1) Ethernet can utilize up to about 96% of the 10Mbps bandwidth 2) Higher utilization results in higher delays (due to collisions and deference). 3) Larger (physically) Ethernets waste more time detecting collisions, which are also more likely. 4) Just because the Ethernet itself can reach ~96% utilization doesn't mean your stations or higher level protocols will survive it. 5) At about 50-60% load the delay shoots up dramatically 6) I have monitored an Ethernet with a sustained load of 48% averaged over an hour. This was a network with about 50 nodes, many diskless. Performance was poor but bearable (debatable). This network usually averaged around 25% load during the busy period of the day. 7) Token rings (in general, assuming the same bandwidth) do much better in terms of delay at the higher loads. However delay becomes a significant factor beyond 70-90% load. Of course the token ring gets a slight penalty in terms of delay at low loads. After going back and looking at your question again I see that what I said in (5) resembles what you heard from your forgotten source. The numbers are fuzzy since there are so many other factors involved. For some Ethernets 40% load may be unbearable, on others 70% load no problem. Randy Smith wsmith@umn-cs.cs.umn.edu ...!ihnp4!umn-cs!wsmith -- Randy Smith