Maximum Daisy chain lenght of switches?

[RS1980]

Hi folks,

I am thinking about a concept to connect multiple switches (small 4 Port with 2 devices on each) to one Server. to save cable, i am thinking about daisy chaining all the small switches and only connecting the last one to the Server instead of using a 24 Port switch to whom they all are connected to.
I hope it's clear what I am talking about?

My question is: How many small switches can I put in a chain, is there any limit, or is the only problem the growing lag between the Server and the last switch in the chain? Bandwitch should not be a problem, there is only little data transfer from the devices to the server.

Thanx in advance for any help!
RS1980

 

[pansophic]

I've never heard of any theoretical or physical limit on daisy chaining switches, but the concept is that the contention for access to the port at the server grows with each additional switch. A distant endpoint is N times as likely to have contention and require a retransmit, increasing the traffic on the network.

But there are lots of other reasons not to do what you are thinking about, like the fact that you have N number of class 2 transformers out there to deal with.

You'll have more cables in the network, and we all know that more cables = more problems (the problem is always the cable).

You'll have cascading problems when they occur. In other words, 1 bad cable, bad power supply, etc. means that a group of users get knocked off, but not everyone (necessarily).

Cable management is much more difficult. You may remember that you use port 1 for uplink, 4 for downlink and 2 and 3 for hosts, but will the next guy, or the guy who is responsible while you are on vacation.

Capacity planning is no longer a "no brainer." You have to remember to put all of your common access hosts as high in the chain as possible. Your server and then printers in descending order of usage potentially. Or do your users generate more traffic than the network printers?

And everyone must go through all of these hops to reach the Internet as well.

It may not be all that bad in the end, but it is a bad precedence to set. I've worked on too many networks that are just hobbled together, and you are considering making another one even when it appears that you have a choice.


pansophic

 

[technome]

Ethernet has rules for repeaters and seagment limits

http://www.webopedia.com/TERM/5/5_4_3_rule.html

........................................
Chernobyl disaster..a must see pictorial

http://www.kiddofspeed.com/default.htm

 

[RS1980]

@technome:
I think this only matters when using hubs and repeaters, not switches/bridges.

@panasonic:
Thanx for your detailed answer. I should have explained what I want to use it for more precisely, my fault...
It's about to be used in a switchgear to connect single monitoring devices to a Master PC who is controlling them all. So all the problems regarding "users and administration" are out of scope (luckily
Each field has 2 devices that are connected with electrically to the local field switch (which has 2 electircal and two optical connections)
The cables between the switches would be optical (MTRJ, MT or MC plug).
The configuraion will not change for a long time, so there should not be a cable Problem, once it is working. If so, the problem would also exist when connecting all field switches to a master switch.

Reliability is a problem, but that could be handled with a managed switch on top and making a ring with all devices (weher either the connection between firstfield-mainswitch or lastflied-mainswitch is active, caused by spanning tree).

so what do you think under those conditions?

RS1980

 

[jimbopalmer]

RS1980 assuming you do not need any more than the speed of one cable, that should work fine. Daiseychained, all devices will share 100 meg (or 10 meg if these are 10 meg switches) while if you used a traditional star pattern each switch would have its own 100 meg to share on that switch. If your server is only 100 meg as well, then the server is the bottleneck either way, if all data goes to the server.

5 100 meg switches coming back to a gig master switch attached to a gig server is only superior if you are generating more than 100 meg of data.



I tried to remain child-like, all I acheived was childish.

 

[pansophic]

It sounds like a FDDI ring...only with Ethernet.

I, personally, would probably still prefer a star topology, mainly because I feel that it makes life easier, but you are using the optical links to move from switch to switch, and using the electrical to attach to your switchgear (I'm assuming that we are talking high-power here). With that in mind, having the long runs as optical is excellent, because you avoid all of the possible noise that the switchgear can cause on your network.

I was also assuming a Windows network, which are noisy machines from a networking perspective. These are embedded controllers, so they will not be noisy. They will probably literally only transmit when they actually need to.

It really sounds a lot better once you've provided the details.


pansophic

 

[PalmTest]

To go back to the original question, I'd daisychain no more than max 8 switches.
Every switch has to process the network packets. So every switch introduces some latency in the complete chain. To avoid timing problems, don't chain too much.

 

[bitb0y]

My thoughts.
First, there are no limits on switched hubs only the media you choose (fiber,copper).
Next, Spanning tree will work if you need.
Next, You need to hard set the full/half duplex so that the autoneg doesn't hang in the event of a power outage. Also, apps don't talk full duplex well if at all so you won't lose speed/efficiency by going 10/100 half duplex. Nail the port to 100/half and leave it. Save the configs for future diags.
Latency may be an issue depending on the apps so keep that in mind when you add nodes.
The good thing is that if a switch goes down you don't have a single point of failure. However, BPDU traffic may be too high to be efficient so watch that too. Collisions are normal so don't panic if you see a lot of red initially. Also during convergence you'll have no traffic flow really.
Just my thoughts.
R.

 

[RS1980]

Thanx for the answers and suggestions!

The overall bandwitch should really be no problem. The normal data flow from each device to the server is some kb every some seconds. Form time to time some 100 kb can be transferred, but that's all about traffic.

The delay over the chain is another problem, because the devices will by synchronized with the server through NTP Protocoll. But I think, even when there is a "big" load of data transferred to the server, the switch backplane should not be to busy to store the NTP packets for too long. So I am guessing that the delay between Master and - lets say the 8th device - will be < 5ms. Do you agree?

@bibt0y
Thanx for the duplex thing, haven't thought about it. I did't know that the duplex autonegotiantion can cause even more problems when power fails than the power failure itself? Or what's the matter?
About the BPDU traffic: If only the Master switch on top supports spanning tree, is the traffic still an issue?

RS1980

 

[vipergg]

I'll blatently disagree that anything should be hardcoded to half duplex unless that is all the nic will support (very uncommon in todays world) . You are basically cutting your bandwidth in half when you do this as this permits one way traffic at a a time whereas full duplex you can transmit and receve at the same time. I certainly would not do this if you go with the daisychaining model . This turns on a collision domain and anything over 30% utilization of that 100 half duplex link and you will start to see major degradation of the network uplinks.

 

[jimbopalmer]

First, I agree that I never hard code duplex unless Auto-negotiation is failing or the link includes fiber.

Second, even if some links are half duplex, it is still a switch, so the old hub rules about 30% usage do NOT apply.

I tried to remain child-like, all I acheived was childish.

 

[vipergg]

That is correct in that user ports at half duplex are isolated but if he uses the uplinks (daisychain) at half duplex it still applies.