What network performance is required for thin clients?

Question

I am building a network with approx. 1000 thin clients. I intend to run them over VNC.

What type of LAN bandwidth and latency are required? What type of hardware will provide that?

If, instead of VNC, I were to use RDP,ICA,NX or another graphics-based protocol, how would that change things?

UPDATE

1. The clients will be used for web browsing (yes, we could just use local browsers, but, there are reasons outside of the scope of this question why we're not)
2. I imagine, at peak times, 100 people will be browsing the web at the same time
3. I'm looking for a rough order of magnitude, not a final answer. Someone to say "Well, we have 200 thin clients in our environment, used for running MS Office, and our LAN at 10Mbs couldn't handle it, we upgraded to 100Mbs and its fine." Yes, the final solution will need to iterate and experiment. But, as initial boundaries, gaining the experience of the community - especially from people who operate similar networks - is an invaluable first step.

Also: I'm not asking about the CPU, the server hardware, or anything else. Only the network aspect: latency and bandwidth, other network considerations.

Answer 1

Your question seems to be a lot more complicated than you think. You may want to do some further research.

You haven't stated what you will be running on these thin clients, different software has different requirements for both bandwidth and latency. Some require low latency, others require high bandwidth, other still require both low latency on high bandwidth. Some software uses very little bandwidth and is usable with substantial latencies, while other software will become completely unusable with the slightest latency, or create massive bandwidth demands bogging down high-end generic networks with as few as 10 clients.

They also aren't as you would expect, for instance a stock Firefox install running server-side can eat through bandwidth like no tomarrow, but can be modified to dramatically reduce it's bandwidth requirment with only minor noticable difference. Meanwhile, a stock LibreOffice install doing graphic layout work will actually not use too much bandwidth at all. However, Maya and Blender are even worse than Firefox, demanding massive requirements from their servers and requiring incredibly high bandwidth and low latency unless you are using a specialize solution for such programs.

Depending on the exact demand, different server hierarchy and network layouts are ideal.

You really can't simply state the number of clients and expect anyone to give you a solid estimate on what sort requirements you need. If anyone not familiar with your software needs gave you an estimate, I would either anticipate that they think you are running what they see as a "typical setup" or they are just giving you the needs for a setup that will support any setup of that size, which will likely be grossly overpowered for your needs.

All I can tell you for sure is that for what you are proposing, which is a very large network of thin clients relying on a centralized host server setup (either cluster, NUMA or extremely high-powered monolithic) require a very specialized setup, and getting the exact hardware and optimized setup you need you is going to save you a lot of money on hardware you don't need to make up for inadequacies.

You really need someone willing to work with you one on one to go through each available solution, and you need to keep an open mind on getting a custom solution. Due to the scale of the project, you don't want a cookie cutter solution. A cookie cutter solution will give you massive headaches and inflate hardware demands. The extra consulting costs of a custom solution will pay off many times in saved expenses, especially down the road if expansion and upgrades are necessary.

I wish I could tell you more, but I can't without more knowledge of your situation.

Answer 2

The speed of the individual links to the Thin Clients will be mostly meaningless (though I would not run them on anything less than 100Mb, and 1GbE is the widely accepted default for TC running a graphical protocol like VNC, RDP, ICA, etc in new installations).

The switches have to have sufficient fabric for the traffic, almost all modern switches have full fabric internally. Then you have to worry about aggregating links, or stacking links - this will depend entirely on what model switches you get. For 1000 TCs on a budget I would snag large modular switches, like a ProCurve 5412zl, 10x 24 Port 1GbE access ports and 2x 2 Port 10GbE uplink ports. It would take 5 such modular switches to do 1000 TCs.

Then you'd have to worry about the core network. That last paragraph would be called the Access Network (the switches that users get "access" to the network from). The "core" network is the set of switches where the servers connect. It would take quite a few servers to run 1000 TC, the exact number would vary wildly with usage - I could imaging anywhere from 10 to 100 realistically. Less servers would need redundant 10GbE connections to the core network. More could probably get away with 4-8x bonded 1GbE connections; but might benefit from simple redundant 10GbE. With a low quantity of servers, say 12 (10 necessary + 2 redundant), you'd need 2x12 ports of 10GbE for the servers and 4x5 Ports of 10GbE for the access network backhaul; something like a pair of ProCurve 5820 switches could provide the ports and redundancy. Also, at 12+ servers you should be considering Blade servers, which are likely to have 10GbE Switch Modules available (though their external connections might not be enough for this application, it would really depend on your setup; something like a HP C7000 with 4x Flex10 switches would be sufficient, if you only have 16 servers; more servers and you could drop to 2x switches in each chassis).

Another note: Firefox and most web browsing is quite graphics heavy. The software graphics of most session hosts is unimpressive, however some can use passthrough graphics to boost performance. This would require an acceptable graphics card in the servers (pedistal or rack servers can usually use "normal" PCIe workstation graphics cards; blade servers have Graphics Mezzanines available) and software that supports as much (some, like recent versions of Windows support this natively, others require licensing, others don't support it at all).

Hopefully the Walls of Text Robert and I have provided give you some idea of the depth of knowledge required to setup something like this properly. It ain't cheap either...

Answer 3

I managed a system of around 200 thin clients for a college on 100mbs copper to switch, 10 gig fiber to wan. Mainly 60 in a library, Office, web browsing, etc. The rest were faculty systems. The network handled it with ease. I would guess with your recommended load that 100mbs copper would be just fine.

I might be a little worried if you ever got into the 500+ systems operating at once.