1Gbs or 100Mbs to the desktop ?

[Face-Man]

It's been recommended that I cut the speed to the desktop down to 100Mbs. Our network bandwith usage doesn't seem anywhere near capacity so is there any advantage in cutting the speed ?

[kmount]

You've been recommended to _cut_ the speed?? - Why?

If there's no problems with it, I'd leave it like that.

[dhicks]

It's been recommended that I cut the speed to the desktop down to 100Mbs. Our network bandwith usage doesn't seem anywhere near capacity so is there any advantage in cutting the speed ?

I can't think, offhand, of any advantages. Did whomever it was who recommended you cut the speed down say why they recommended that?

--
David Hicks

[Theblacksheep]

It's been recommended that I cut the speed to the desktop down to 100Mbs. Our network bandwith usage doesn't seem anywhere near capacity so is there any advantage in cutting the speed ?

Only reason for putting them on 100MB instead of 1G would be to free the 1G ports up for something else that needs it (servers/switches).

[srochford]

Have you currently got gigabit switches etc in place and you're being asked to reduce the speed to 100Mbit or have you got future plans for gigabit and you're being asked not to implement it?

If you're already running gigabit I can't see any reason to stop!

[maniac]

I think gigabit to desktop should be standard in any new network, as pretty soon we're going to have 10Gbs links to servers as standard, and the whole thing will shift up a gear like it did with the wide spread introduction of Gb links years ago. With more wide spread useage of video and multimedia content and storage being rediculously cheap meaning everything can be stored on the network storage with no problems, gigabit to desktop is going to be a necessity pretty soon.

Why or earth would anyone suggest changing it to a slower speed? Unless it's a plan on the drawing board, and they're trying to save cash? Even then, the cost difference is a fraction of what is used to be, gigabit switches are very affordable now, and most workstations come with gigabit NICs as standard these days anyway.

Mike.

[Galway]

If you have allot of clients with gigabit, and a server connected at gigabit to a gigabit backbone, there could be problems that the sever is swamped with requests. It some situations, can can be beneficial to keep the backbone at gigabit and hold the clients at 100mb to allow the server at logon times to work as efficiently as possible. Adding an additional network card and teaming them to double the tranfer rate at the server allows the disk subsystem to try to keep up but ultimately depends on the number of network clients and the services the server provides.

[torledo]

If you have allot of clients with gigabit, and a server connected at gigabit to a gigabit backbone, there could be problems that the sever is swamped with requests. It some situations, can can be beneficial to keep the backbone at gigabit and hold the clients at 100mb to allow the server at logon times to work as efficiently as possible. Adding an additional network card and teaming them to double the tranfer rate at the server allows the disk subsystem to try to keep up but ultimately depends on the number of network clients and the services the server provides.

...or take the more progressive step of implementing 10gbps to servers that really need it. It really depends on the type of traffic between clients and servers, gigabit ethernet adapters are incredibly smart able to free up CPU cycles for multicore servers - ofcourse the issue of saturization or overwhelming a server by pushing CPU usage to 100% could be an issue depending on the number of gigabit clients and types of applications communicating with the server, but this is exactly why baselining and monitoring load, latency, utlization is so important. You can't go round throwing 10gbps card in willy nilly, nor should throttling clients be a good idea becuase of a 'perceived' performance impact. As you've said adapter teaming is the cheap method of balancing load but i agree with maniac that eventually 10gbps will be the dominant server connect. And while there is still a premium on 10gbps adapters the situations where they are needed in real world scenarions are few and far between for schools. Am happy to be proved wrong by those who have the metrics to prove otherwise.

I think the obvious candidates for 10gbps at the moment are virtualized servers because of the mix and match of different server applications being hosted.

[dhicks]

It's been recommended that I cut the speed to the desktop down to 100Mbs.

Had a think about this as I ambled home. Did think of a (vaguely) possible reason: as suggested above, it could be that the switches to your desktop machines have a limited bandwidth available. What's the backplane capacity of your edge switches? If it's less than the theoretical total available with all the switch's network ports running at maximum speed, then maybe explicitly telling some ports to run slower will stop your switch being overloaded. Similarly, maybe telling some ports to run slower gives higher priority to faster-running ports (i.e. your gigabit uplink ports).

My advice is to check with whoever told you about the idea in the first place. Also read your switch's manual (PDF document is probably available somewhere in Google land if you don't have the actual printed manual to hand).

Bearing in mind you can get a perfectly decent (and fanless) 24-port gigabit HP switch with a backplane capable of over 48Mbit/s for £300, I'd say if this is the case then just replace your crusty old switches with shiny new ones.

--
David Hicks

[Galway]

...or take the more progressive step of implementing 10gbps to servers that really need it.

And if you have a gigabit server and 300 computers on gigabit this helps the situation or makes it worse?

The whole point is if 3 classes of 30 students logon, the server gets an avalanche of data to transfer and it becomes very inefficient in getting it to the clients. If the clients are kept at 100mb, the server completes the transfers quicker and therefore the students log on quicker because the server is working more efficient.

Its like copying 3 x 1 gig files. If you do them 1 at a time the transfer is faster then doing all 3 in parallel since the system is more efficient is accessing and transferring the data. I have actually witnessed this effect and can personally say in some situations has a significant impact on the network performance. All you need is one client moving media clips/cd images data about to effectively swamp the server to capacity.

[torledo]

And if you have a gigabit server and 300 computers on gigabit this helps the situation or makes it worse?

The whole point is if 3 classes of 30 students logon, the server gets an avalanche of data to transfer and it becomes very inefficient in getting it to the clients. If the clients are kept at 100mb, the server completes the transfers quicker and therefore the students log on quicker because the server is working more efficient.

Its like copying 3 x 1 gig files. If you do them 1 at a time the transfer is faster then doing all 3 in parallel since the system is more efficient is accessing and transferring the data. I have actually witnessed this effect and can personally say in some situations has a significant impact on the network performance. All you need is one client moving media clips/cd images data about to effectively swamp the server to capacity.

Again, i think that's tackling the problem from the wrong angle. If you've got 300 computers on gigabit and just a single server on gigabit it's pretty obvious the main problem is that you have only one server servicing up to 300 clients NOT that gigabit to the desktop is a bad idea. The whole point of centralized computing - whether that be client-server or centralized storage using SANs - is that management is easiest at the centre. Adding more RAM to a server, adding another blade to a blade center, adding another CPU or another server to a cluster, adding another domain controller. That's the way to tackle the problem of server load not handicapping network speeds of client computers. Ofcourse if the budget isn't there to make those server upgrades (and definition of server spec will vary depending on architecture and type of server application being hosted) then ofcourse a sysadmin must do what they need to do to not impact overall performance of the network.

Putting a 10gbps in a single core, single socket server which is plainly not up to the task servicing hundreds of client computers is not a result of the problem with the idea of having 1gbps to the desktop and 10gbps to the server, it's more a problem of implementation. 10gbps isn't a magic bullet, you can't just plug it into a compatible server and expect all performance issues to be eradicated. But, a modern 10gbps adapter in a suitably sized server(s) is preferable to multiport gigabit adapters or mutiple gigabit NICs in a server from a performance, price and footprint point of view.

[Face-Man]

wow thanks for the replies.

The reason it has been sugested is that another school has implemented it and seen inprovement especially for their over worked internet contection. We have been having no internal problems but lots of internet ones (especially using https which avoid caching) and yes our RBC has been getting it in the ear. However our bandwidth appear to be under used.

To sum it up the arguement is that cutting the speed to 100 reduces the contention ratio on our switchs and backbone. What I'm not sure about is if ethernet is at all bothered by this at low capacity.

[localzuk]

If you've not been getting any problems with bandwidth internally, changing the settings on your desktop machines shouldn't change anything.

Do you have a proxy server? Is it running well? Or is it running full on? What edge router do you have? Is this being overloaded?

[plexer]

The fact that your desktops are connected to gigabit switch ports does not mean that your client are runnign at gigabit speeds, what is the actual peak utilisation on those cards? nowhere near 100% I bet.

This is unlikely to affect your internet connection.

Ben

[CyberNerd]

The fact that your desktops are connected to gigabit switch ports does not mean that your client are runnign at gigabit speeds, what is the actual peak utilisation on those cards? nowhere near 100% I bet.

+1

You really need to get some baseline information first to analyse where the bottlenecks are before spending money. Without any hard figures it is difficult to work on assumptions.
I reckon the biggest bottleneck will be on your clients,somewhere between disk access and the network card.(32bit pci bus?)