Shrink yourself to the size of an ant and think of a classic server as a mini data centre. Banks upon banks of memory, multiple disks, many CPU cores and even network interfaces. The OS manages all of the many processes and threads, sharing the resources amongst them and when they demand more resource it makes it available to them – on demand and apparently instantly. Importantly all the resources in my mini data centre genuinely appear as one enormous pool available to all my applications at once. My application runs in a process and doesn’t care whether it consumes one core or 8. So what is a cloud OS?
Cloud OSes do appear to be quite different because somehow I need to play a part traditionally played by the OS. When my application demands more resources I get involved. Typically I need to allocate more virtual machines or storage and quite likely I need to tell my application to extend itself into this new resource (i.e. install my application components and bring it into the pool/farm/cluster/you-name-it). Then when that demand passes I need to consider releasing those resources.
Consider for a moment too the utility model, or to use a common analogy: “cloud computing is like the power grid – on demand, just in time, scale up and down”. When I need to consume more power, the statistically multiplexed grid allows me to draw as much as I want. There are even special burst providers like pumped storage schemes to handle the gigantic surge of kettles going on after a football game. If I need a milliamp for a second or 10 amps for an hour it makes no odds to me(!).
Again the cloud seems to break this analogy. My application will consume up to the limit of pre-allocated resources and stop. The data centre might have a potential for a billion petaflops but the cloud OS can’t just grant them to my app. It is a bit like knowing that I am about to plug in my electric car and ringing the grid to provision me more current.
Once upon a time the server OS used to be quite crude but they evolved and we take their sophistication for granted now. Most people struggle to even describe what an OS does. At the moment I need to resort to special software like Appistry to help me virtualise my application so that when I auto scale up/down my Amazon machine images the application can do so too. How long before these capabilities are part of the cloud OS allowing applications to genuinely scale up and down very discretely? Will all the CPUs in the cloud data centre eventually present themselves as a single virtual resource with the cloud OS presenting abstractions, much like a regular process, as execution containers or in the cloud OS will the virtual machine always get in the way?
It feels as if the virtual machine is a convenient way of *currently* spanning large amounts of hardware, but is it a stop gap? Will a genuine cloud OS come along that behaves like a regular server OS atop a giant data centre? It will be happy to present 24TB of apparently contiguous RAM and 1000 cores to a single “process”. It will be like NUMA and move application data across the server farm to the CPU that has been scheduled to work on it. Do we need it? Are we content for the next few decades to worry about how to span virtual machines and make them collaborate and auto-scale? Is the VM the new Process?
P.S. I appreciate that in many clouds, storage capacity seems to have been the first to become a truly virtualised and discrete resource, although storage performance hasn’t (i.e. most solutions are too blunt to configure for different performance characteristics etc.)