During my studies about QFX-5100-48T virtual chassis I noted that there is a kind of lack in the documentation about the switch over-subscription's topic.
The 2 sets of built-in interfaces on QFX-51000-48T, 48 10GbE (connecetd to end hosts) and 6 40GB (uplink) can offer a switch-oversubscritpion of 2:1, (480Gb : 240Gb). So far so good. Jumping now on virtual-chassis topic (please note I am not talking about virtual chassis fabric)I further read that the traffic going from one switch to another in a ring topology is subject to nondeterministic over-subscription, depending on how many devices are between the source and the destination. My question now is: how can I compute the switch over-subscription in case of a virtual-chassis with 10 Node's members? If the books mentions non-deterministic over-subscription, is it really impossible to analise or plan the performance, in terms of over-subscription?
If you really have the need for a lot of east-west (probably server) traffic, then VCF (or some type of Fabric solution) is probably a better solution to consider. If the traffic is more north-south, then the oversubscription for a VC is more related to uplink capacity, then VC capacity. Now you can make a VC type fabric with upto 7 x QFX5100s and ensure the total oversubscription is 2:1 but creating something like my [very crude] drawing, but for traffic between any 2 switches the oversubscription could be worse case 12:1, if all 48 input 10GE wanted to talk to interfaces on only one other switch. Therefore we you physically locate your servers within a fabric could well matter.
Hope this helps. I hope the drawing comes through, is readable and makes sense. For the 7 node VC "fabric" all 6 x 40GE ports would need to enabled as VCP interfaces.
The usual answer is to go leaf - spine rather than full mesh. In order to guarantee 2:1 oversubscription in a 7 switch fabric, you'd have to get 2 spine switches (QFX5100-24Q), and run 3 uplinks out of each leaf into each spine. The question is if you need 2:1. You may only need 4:1 or 6:1.
I don't think you can guarantee a 2:1 oversubscription in a ring given that the bandwidth of your 6 uplinks is a shared resource across multiple switches, and though there may be a way to figure out the oversubscription ratio between two given specific hosts, I think it would be very complicated, and would not be consistent oversubscription across all hosts. It would be affected by how much traffic is going east vs west, and the number of switch hops between any two given hosts.