When I run a worker role instance on Azure, is it a complete VM running in a shared host (like EC2)? Or is it running in a shared system (like Heroku)?
For example, what happens if my application starts requesting 100 GB of memory? Will it get killed off automatically for violation of limits (á la Google App Engine), or will it just exhaust the VM, so that the Azure fabric restarts it?
Do two roles ever run in the same system?
It's a whole VM, and the resources allocated are based directly on the size of VM you choose, from 1.75GB (Small) to 14GB (XL), with 1-8 cores. There's also an Extra Small instance with 768MB RAM and shared core. Full VM size details are here.
With Windows Azure, your VM is allocated on a physical server, and it's the fabric's responsibility of finding such servers to properly allocate all of your web or worker role instances. If you have multiple instances, this means allocating these VMs across fault domains.
With your VM, you don't have to worry about being killed off if you try allocating too much in the resource dep't: it's just like having a machine, and you can't go beyond what's there.
As far as two roles running on the same system: Each role has instances, and with multiple instances, as I mentioned above, your instances are divided into fault domains. If, say, you had 4 instances and 2 fault domains, it's possible that you may have two instances on the same rack (or maybe same server).
I ran a quick test to check this. I'm using a "small" instance that has something like 1,75 gigabytes of memory. My code uses an ArrayList to store references to large arrays so that those arrays are not garbage collected. Each array is one billion bytes and once it is allocated I run a loop that sets each element to zero and then another loop to check that each element is zero to ensure that memory is indeed allocated from the operating system (not sure if it matters in C#, but it indeed mattered in C++). Once the array is created, written to and read from, it is added to the ArrayList.
So my code successfully allocated five such arrays and the attempt to allocate the sixth one resulted in System.OutOfMemoryException. Since 5 billion bytes plus overhead is definitely more that 1,75 gigabytes of physical memory allocated to the machine I believe this proves that page file is enabled on the VM and the behavior is the same as on usual Windows Server 2008 with the limitations induced by the machine where it is running.
来源：https://stackoverflow.com/questions/10925218/is-a-windows-azure-worker-role-instance-a-whole-vm