160 person LAN -- is this enough? or will it lag?

[Quad]

ok. i'm not actually doing this but i want to know the outcome.
for a 160 person lan, will the following scheme work (link below). assume that this lan is for gaming and file transfers. my ultimate goal is to have a low ping, 0 packet loss environment.

please note: when i say 40port hub, i mean some sort of combination of uplinked hubs to allow 40 connections (ie. 5 8-port hubs uplinked together)
http://members.nbci.com/quad6969/150_network.jpg


since i want a silky smooth network, should the hubs be 100mb? or will 10mb be fine since they connect to the switch?

thx

 

[Russ]

<< 5 8-port hubs uplinked together >>



You cannot daisy-chain more then two hubs, so you'll need to use all switches. It also makes no sense to even consider 10Mb.



<< my ultimate goal is to have a low ping, 0 packet loss environment. >>



There is no such thing as ZERO packet loss. Collisions are part of networking. All you can do is minimize it with good equipment. A hub is NOT defined as good equipment in this day and age.

If you are networking 160 systems, you also might want to consider adding a router or two to the mix, and splitting it in to multiple physical segments.

Russ, NCNE

 

[Quad]

<< You cannot daisy-chain more then two hubs, so you'll need to use all switches. It >>



why can't u daisy-chain more than two? i believe that i've seen it done before. what are the implications of doing so?

thx

 

[Russ]

<< i believe that i've seen it done before >>



If so, then the installation was done by someone who didn't know what they were doing.

The spec calls for a maximum of two because beyond that the potential for collisions grows dramatically. A hub does not manage traffic the way a switch does.

Russ, NCNE

 

[Quad]

thx russ for your quick responses.

so if i wanted to create a 160 person lan, is this more feasible:

get 7 24-port 100mb hubs
connect the hubs to an 8port 100mb switch

or is there something more economical/cheaper/better ?

 

[Russ]

<< get 7 24-port 100mb hubs >>



No. You want SWITCHES. You don't even need the extra eight porter since seven 24 porters gives you 161 stations.

Russ, NCNE

 

[Quad]

oh ok. so you're saying get 7 24port switches. but wouldn't that cost a hell of a lot of money?

 

[Russ]

<< but wouldn't that cost a hell of a lot of money? >>



Did you think setting up a 160 station LAN was going to be cheap? There is not that much price difference now between a switch and a hub, and you cannot daisy-chain that many hubs anyway.

Russ, NCNE

 

[Quad]

ok russ. would u reccomend any places to buy switches of this sort?

 

[Russ]

Just pick a good quality brand such as SMC, Netgear or Linksys and buy it where you can get the best price. Per port, 16 port switches are going to be quite a bit less then 24 port, so they'd be a better choice.

You would need 11 switches and, for example, the 16 port SMC EZnet is $145 at buy.com.

Russ, NCNE

 

[Xanathar]

It also depends what you want to do, will you need the ability to remotly turn off and on ports? Monitor mac addresses? if so you will need manageable switches. Do you want to be manageing 1 switch, or 10?. Look at the whole picture also, like cost to maintain and cost if something is down.

 

[PELarson]

<< ok russ. would u reccomend any places to buy switches of this sort? >>



Intel Express 510T Switch - 24-port - Rackmount for $299 from Coumptergeeks.

 

[ScottMac]

Assuming a single server, IMHO, the performance is gonna suck. This is a scenario where you need to have a big pipe to the server, and for games, probably a whompin' server as well.

Since you have 160 machines pretty much aimed at one place, through one pipe, the final aggregating segment should be as large as possible. If all of the stations are at 10, and the server-link segment is 100, you're 16 times oversubscribed on the link to the server. If each stations are 10 meg, and the server-link is a gigabit, you are only 1.6 times oversubscribed, and much more likely of getting packets through smoothly.

The next issue, in a similar vein, relates to the peripheral switches. Same scenario, 23/24 (or 47/48) ports aimed at the same exit. A gig uplink would do wunnerful things for the rapid egrees of your data. And though a switch with gig uplinks is more expensive, you could easily go to 48 port switches and not oversubscribe the uplink (assuming 10 meg stations).

Put single gig switch as a concentrator feed to the server, and you got a pretty decent setup.

Many-to-one is bad for a switch, unless the gozoutta is much bigger then the collective gozintas.

And...a collision is NOT a dropped packet. The collision occurs as the packet is transmitted, the conflicting transmitters fall back for a short (standards-defined random) time, then they give it another shot. That's why the spec syas you can only have so many hops, and limits the segment length, so the collision can be sensed in time to kill the transmission and fall back. Too many hops, or excessive segment length would allow the packet to be completely transmitted before the collision is sensed (a &quot;Late Collision&quot; usually shown on an analyzer as a series of A's or 5's in the captured packet).

Dropped packets are purely a phenominon of the switched or routed environment. The packet is sent, buffered, and due to (usually) high latency (like slow/busy switches, slow/busy routers, satellite transit times, etc), the packet times-out or gets dumped for priority traffic. Then the higher-level protocol (which usually takes longer than collision-fallback-retransmit) has to decide things are taking too long and crank out another packet (which is subject to the same path latencies or policies)...and hope this one gets through.

If you put a switch in the wrong place, or implement it improperly, it will amplify everything bad about your network.

FWIW / .02

Scott

 

[spidey07]

Yeah, 160 person lan is not going to be cheap if you want it to perform well at all. Russ is right on with the two hub limitation with 100 megabit hubs. year is 2001 ya know.

Ideal setup would be some kind of eight port gigabit switch star network with 48 port switches as the leafs. Server(s) are gig attached to the core eight (or 16 if you have that many servers) attached. From that core gig switch connect your (4) 48 port switches giving you a very compact high performance network.

Better yet, at 160 nodes i'd use some soft of chassis based switch. Maybe 30K or so.

To add to SCOTTMAC'S point, any kind of packet loss on a LAN given today's technology is totally unacceptable. If you're dropping frames on a LAN you have VERY serious problems. 160 nodes on a single broadcast domain is fine so you don't need some kind of router, depending on the protocols you use you can go as high as 300 or 400 nodes per bcast domain before wan't to do LAN routing.

hope this helps!

by the way, silky smooth network and hubs don't belong in the same sentence, 2001 ya know.

 

[trikster2]

Hey

Your multiple hubs connected to a central switch is a valid fast ethernet configuration; it would be one way to keep costs down at the sacrifice of some performance.

 

[Russ]

<< Your multiple hubs connected to a central switch is a valid fast ethernet configuration >>



trikster2,

No, it is not a valid &quot;fast&quot; ethernet configuration. First, his diagram uses 10Mb hubs, and second, he's daisy-chaining too many.



<< Assuming a single server, IMHO, the performance is gonna suck. >>



Scott,

I think it would suck no matter the size of the pipe, unless the employees are just sitting around playing solitaire all day (or the company has a HELL of a lot of money to spend on the server).

Personally, if a client came to me with this project, I would recommend multiple servers, probably a total of four. With 160 stations, I'd assume that there was some important work going on, and some redundancy was in order.

Of course, I always look at these things from a business standpoint, and I may not be up to speed on the needs of a gaming LAN.

Looking at the diagram and the question, though, I figured he was more interested in simplicity then anything else.

Russ, NCNE

 

[CoolTech]

this is a massive undertaking, first tell us your budget, then we can tell if what you want to do is feasible in an efficient manner.

 

[Quad]

wow. ok, i'm really misinformed. what i'm getting at here is: how is it done at those huge lan parties? what kind of network setup do they have? for example: The gathering 2001. i believe it was something like 5000 people. what would be the network layout for something as massive as that?

 

[ScottMac]

I dunno. limiting the clients to 10Meg (more'n enough for gaming) and keeping the fat pipes in the horizontal and the server would probably do OK.

If it's a commercial system, then yeah, cluster a couple servers or put in a load sharing switch. If it was for file/print/web/ftp...drop in a Novell box. IT'll easily handle four + times the load of the MS servers. Or a Linux/UNIX box, Solaris..whatever.

Extreme let's you trunk gig ports switch-to-switch if you need *really* fat pipes, oe they have a DWDM blade that'll do (pretty sure) eight lambda (one gig per lambda)...and it's only US$60,000.00 each (needs two, of course).

ATM would work really well for him too. The multiplexing effect of per-VC queuing would give everybody a fair shot at the resources, even if it's a bit oversubscribed. They would also get the benefit of REAL QOS and policing, traffic shapeing, traffic management, the whole enchalada (instead of waiting another year or so for end-to-end MPLS to be inplemented (and that's for the business switches, consumer switches with MPLS are probably WAAAAAAAAY out there).

There's lots of solutions, but he started out with daisy-chained hubs....if they've got the extra cash for a switched-star config with fat pipes where they'll do the most good, I figure they're ahead of the game. Nuts, even if they star'd a bunch of hubs to a central killer switch &amp; a fat pipe to the server(s) they'd be way ahead of what they were planning on.

Other than a few suggestions to the local Mech Warrior parlor to improve their network, I don't have much experience with gaming stuff either. I just figured it can't much worse than a hard-core database network with a bunch of inexperienced programmers and managers. You'd probably either laugh or cry if I told you what (at least) one of the major credit card companies use for their core network.

What can you do? Folks is Folks, they're gonna do what they're gonna do. &quot;My mind is made up, don't confuse me with the facts.&quot; They'll do what they can do with the cash at-hand. Advice was offered, they get to choose what and how much to accept or ignore.

That's the business. It all pretty much boils down to credibility, tempered with the availability of cash. There's almost always a compromise.

At Networld + Interop 2000/2001 (12,000 nodes, 850 booths) they used 10/100 drops to the end-users, concentrated the drops to Gig links to a &quot;super-concentrator&quot; node that aggregated the gig links and fed the services resources, which included 2 DS3's to the Internet, and an OC3 to Internet2. There was a branch like this in each hall. Some services (DNS, etc) were located on servers (quad-Xeon 500's, 1 gig RAM, RH Linux (6.2?))in each hall.

I'm babbling again, sorry.

FWIW / .02

Scott

 

[L3Guy]

I agree with ScottMac, but would like to add another perspective.

ALL Lans are a compromise of cost, performance and complexity.

I think that the lan Quad has proposed would be disappointing as a game lan, but is very low cost and is better than the lan I had at work until recently.

If I was going to have 5000 people in a lan party, I would split it up into smaller lans with only one game per lan with NO connectivity between. One server, one game, one lan. Just my personal opinion.

I agree with Scott, that for a game lan, I might limit my clients to 10 Mb to give all players more equal access to the server. We have seen in the lab that (slightly) faster machines can have far greater access to servers than slower machines.

My suggestions would range as follows:

Slowest:40 5 port 10 meg hubs, using 4 12 port 10/100 switches, cascading with 100 mb into a 12 port 10/100 switch. hang 4 servers with 2 nics capable of link aggregation (fast etherchannel)

Next, 7 24 port 10 MB switches with dual 100 mb uplinks cascading into a 24 port 100 mb switch,
using 2 port link aggregation.

Fastest, find a 160 port 10/100 L2 switch such as a CAT 5500/6500, Extreme Alpine, 3Com CB 9000,
Foundry Big Iron, Nortel 8000 or equivalent.

My impression is that 250 Mbps or so is the practical maximum of a Gig nic running on a 32 Bit PCI bus. Also, multiple 100 Mb NIC's will not exceed this without multiple pci busses.
I would size my servers accordingly.

Just my 2 cents.

 

[trikster2]

Russ:

I was refering to Quads alternate proposal, which IS a valid fast ethernet configuration:

>so if i wanted to create a 160 person lan, is this more feasible:

>get 7 24-port 100mb hubs
connect the hubs to an 8port 100mb switch

And this is similar to what is showed:
http://members.nbci.com/quad6969/150_network.jpg

Both are valid Fast Ethernet configurations given that the center of the star is a switch.

For sharing internet access (where they probably have something like a 1.54Mb/s T1), email, and light file sharing the above configurations are economical and adequate.

jeez: I've seen the same setup with 110 users hanging off each 10-base-T spoke; performance was horrible but no one complained.

 

[Russ]

trikster2,

Sorry, I thought you were referring to his original diagram. Yes, that second approach is valid. But, I still think that on a project of this size, it makes more sense to get switches.

How much price difference is there between a 100Mb 24 port hub and the same sized switch? Maybe $75 to $100. On a LAN this large, that shouldn't be a consideration.

A switch gives you a lot more configuration flexibility then a hub, and I think it makes sense to be looking at, and planning for, the future.

Russ, NCNE

 

[spidey07]

48 port 10/100 switches with 2 gig ports are only like 4500 a piece. Then a gig ethernet switch at the core is another 4000.

pretty cheap to support 160 people is you ask me. Still 160 people, i'd use a chassis based switch and take all the guesswork out of it. Plus it would be way easier to troubleshoot and SCALE.

what kind of budget we talkin here? Oh one more thing quad, engineering 160 node networks differ significantly from a 5000 node one. 5000 node you HAVE to route and plan. You can't just star everything into one switch because you can overload the switch. Spanning-tree design comes into big play as well. There are many ways to support a 5000 node network and all the different designs focus on where and how you want to route the frames. Also distance limitations of copper, I doubt you can squeeze 5000 computers into a 100 meter radius.

<---------/me need white board darnit!

 

[Quad]

ok awesome. at least now i think i have a better idea of what needs to be done.



<< Also distance limitations of copper, I doubt you can squeeze 5000 computers into a 100 meter radius. >>


Check out this 5000 person lan

but this network is not for a big business, with database work and dual t3's. it's strictly for gaming (and file sharing). Russ and Scott, thx for your explanations on packet loss. i understand that u can't have zero packet loss, but i would like to keep it to a minimum, because dropped packets are a real pain in games.

say we went with the 7 24port 100mb hub idea, all connected to an 8port 100mb switch. what are some obvious problems with this configuration, given what the network will be used for (stated above). what modifications would you make to this simple setup?

 

[ScottMac]

The problem as I see it remains that you'll have somewhere between 1.6 GIG and 16 GIG of traffic all trying to squeeze through a single 100 meg pipe to get to the server (or are there going to be more than one server?).

You might as well spec in Apple LocalTalk, you're not gonna get much more than 256K through anyhow.....

JM.02 / FWIW

Scott