Load balancing for custom client server app in the cloud

Question

I'm designing a custom client server tcp/ip app. The networking requirements for the app are:

1. Be able to speak a custom application layer protocol through a secure TCP/IP channel (opened at a designated port)
2. The client-server connection/channel needs to remain persistent.
3. If multiple instances of the server side app is running, be able to dispatch the client connection to a specific instance of the server side app (based on a server side unique ID).

One of the design goals is to make the app scale so load balancing is particularly important. I've been researching the load-balancing capabilities of EC2 and Windows Azure. I believe requirement 1 is supported by most offerings today. However I'm not so sure about requirement 2 and 3. In particular:

1. Do any of these services (EC2, Azure) allow the app to influence the load-balancing policy, by specifying additional application-layer requirements? Azure, for example, uses round-robin job allocation for cloud services, but requirement 3 above clearly needs to be factored in as part of the load balancing decision, i.e. forward based on unique ID, but uses round-robin allocation if the unique ID is not found at any of the server side instances.

2. Do the load-balancer work with persistent connection, per requirement 2? My understanding from Azure is that you can specify a public and private port-pair as an end-point, so the load-balancer monitors the public port and forward the connection request to the private port of some running instance, so basically you can do whatever you want with that connection thereafter. Is this the correct understanding?

Any help would be appreciated.

Answer 1

Windows Azure has input endpoints on a hosted service, which are public-facing ports. If you have one or more instances of a VM (Web or Worker role), the traffic will be distributed amongst the instances; you cannot choose which instance to route to (e.g. you must support a stateless app model).

If you wanted to enforce a sticky-session model, you'd need to run your own front-end load-balancer (in a Web / Worker role). For instance: You could use IIS + ARR (application request routing), or maybe nginx or other servers supporting this.

What I said above also applies to Windows Azure IaaS (Virtual Machines). In this case, you create load-balanced endpoints. But you also have the option of non-load-balanced endpoints: Maybe 3 servers, each with a unique port number. This bypasses any type of load balancing, but gives direct access to each Virtual Machine. You could also just run a single Virtual Machine running a server (again, nginx, IIS+ARR, etc.) which then routes traffic to one of several app-server Virtual Machines (accessed via direct communication between load-balancer Virtual Machine and app server Virtual Machine).

Note: The public-to-private-port mapping does not let you do any load-balancing. This is more of a convenience to you: Sometimes you'll run software that absolutely has to listen on a specific port, regardless of the port you want your clients to visit.