You can call it a “Cloud operating system”, an “adaptive infrastructure framework” or simply “IT management middleware” (my vote) as you prefer. It’s the software that underpins the automation engine of your Cloud. You can’t have a Cloud without an automation engine, unless you live in a country where IT admins run really fast, never push the wrong button, never plug a cable in the wrong port, can interpret blinking lights at a rate of 9,600 bauds and are very cheap. The automation engine is what technically makes a Cloud. That engine is an application whose business is to know what needs to be done to maintain the IT environment you use in a state that is acceptable to you at any point in time (where you definition of “acceptable” can evolve). Like any application, you want to keep its business logic neatly isolated from the mundane tasks that it relies on. These mundane tasks include things like:
- collecting events and delivered them to the right place
- collecting metrics of the different IT elements
- discovering available resources and accessing them (with or without agents)
- performing coordinated actions on IT elements
- maintaining an audit of management actions
- securing the management interactions
- managing long-running tasks and processes
That’s what management middleware does. It doesn’t automate anything by itself, but it provides an environment in which it is feasible to implement automation. This middleware is useful even if you don’t automate anything, but it often doesn’t get called out in that scenario. On the other hand, automation means capturing more business logic in software which makes it imperative to clearly layer concerns, at which point the IT management middleware can be more clearly identified within the overall IT management infrastructure.
This is happening in many different ways. I can count seven roads to IT management middleware, seven ways in which it is emerging as an identifiable actor in data centers. Each road represents a different history and comes with different assumptions and mindsets. And yet, they go after the same base problem of enabling IT management automation. Here is a quick overview of these seven roads.
Road #1: “these scripts have to grow up”
This road starts from all the scripts common in IT operations and matures them. It’s based on the realization that they are crucial business assets, just like the applications that they support. And that they implement reusable patterns. Alex Honor described it well here. Puppet and Powershell are in this category.
Road #2: “it’s just another integration job”
We’ve been doing computer integration almost since the second piece of software was written. There are plenty of mechanisms available to do so. IT management is just another integration problem, so let’s present it in a way that allows us to use our favorite integration tools on it. That’s the driver behind the use of Web Services for management integration (e.g. WSDM/WS-Management): create interfaces to manageable resources so that existing middleware (mostly J2EE application servers, along with their WS stack) can be used to solve the “enterprise IT management” integration problems in a robust and reliable way. The same logic is behind the current wave of REST-based IT management efforts (see this presentation). REST is a good integration approach, so let’s turn IT resources into RESTful resources so we can apply this generic integration mechanism to enterprise IT management. Different tool, but same logical approach. Which is why they can be easily compared.
Road #3: “top-down”
This is the “high road”, the one most intellectually satisfying and most promising in the long term. But also the one with this highest hurdle off the gate. In this approach, you create a top-down model of your system and you try to mediate management actions through this model. But for this to be practical, you need to hit the sweetspot in many dimensions. You need composable sub-models at a level of granularity that makes them maintainable. You need to force enough uniformity but not so much as to loose all optimizations. You need to decide which of the myriads of configuration variables you include in your model. Because you can’t take the traditional approach of “I’ll model it and display it to my user who can decide what to do with it”. Because the user now is a piece of software and it can’t make a judgment of whether it is ok if parameter foo differs from the desires state or not. This has been worked on for a long time (remember HP’s UDC?) with steady but slow progress. Elastra has some of the most interesting technology there, and a healthy dose of realism and opportunism to make it work.
Think of it as SCA component/composites but not just for software artifacts. Rather, it’s SCA for all IT elements, with wires and policies that are just rich enough to allow meaningful optimization but not too rich to be unmanageable. If you can pull off such model-driven IT management middleware, then the automation code almost writes itself on top of it.
Road #4: “management integration is another feature of our management console”
That was the road followed by the Big Four. Buy enough of their products (CMDBs, network management console, operations console, service desk, etc…) and you’ll get APIs that allow you to leverage their discovery, collection, eventing and process management features. So you can write your automation on top. At least on paper. In reality, these APIs are too inconsistent and import/export-oriented to really support SOA-style (or REST-style) integration, even though they usually have a SOAP and/or plain HTTP option available. It’s a challenge just to get point to point integration between these products, even more a true set of management services that can be orchestrated. These vendors know it and rather than turning their product suites into a real SOI (Service Oriented Infrastructure) they have decided to build/buy automation engines on the side that can be hard-wired with the existing IT management products. That’s your IT management middleware but it comes bundled with the automation engine rather than as an independent layer.
Road #5: “management integration is another feature of our hypervisor”
If the virtual machine (in the x86 virtualization sense of the term, a.k.a. a fake machine) is the basic building block of your IT infrastructure then hypervisor interfaces to manipulate these VMs are pretty much all you need in terms of middleware to build data center automation on top, right? Are we done then? Not really, since there is a lot more to an application than the VMs on which it runs. Still, hypervisors bring the potential of automation to what used to be a hardware domain and as such play a big part in the composition of the IT management middleware of modern data centers.
Road #6: “make it all the same”
From what I understand about how Yahoo, Google (see section 2.1 “System Health Infrastructure”), Microsoft (see the “device manager” and “collection service” parts of Autopilot) and others run their Web applications, they have put a lot of work in that management middleware and have made simplicity a key design goal for it. To that end, they are willing to accept drastic limitations at both ends of the IT infrastructure chain: at the bottom, they actively limit the heterogeneity of resources in the data center. At the top, they limit the capabilities exposed to the business applications. In an extreme scenario, all servers are the same and all the business applications are written to a few execution/persistence/communication environments (think GAE SDK as an example). Even if you only approximate this ideal, it’s a dramatic simplification that makes your IT management middleware much simpler and thinner.
Road #7: “it’s the Grid”
The Grid computing and HPC (High Performance Computing) communities have long been active in this area. There is a lot of relevant expertise in all the Grid work, but we also need to understand the difference between IT management middleware and the Grid infrastructure as defined by OGSI. OGSI defines a virtualization layer on which to build applications. It doesn’t define how to deploy, manage and configure the physical datacenter infrastructure that allows OGSI interfaces to be exposed to consumers. With regards to HPC, we also need to keep in mind that the profile of the applications is very different from your typical enterprise application (especially the user-driven apps as opposed to batch jobs). In HPC environments, CPUs can run at full capacity for days and new requests just go in a queue. The Web applications of your typical enterprise don’t have this luxury and usually need spare capacity.
All these approaches can complement each other and I am not trying to pin each product/vendor to just one approach. In this post (motivated by this podcast), Stu Charlton discusses the overlap and differences between some of these approaches. Rather than a taxonomy of products, this list of seven roads to IT management middleware is simply a cultural history, a reading guide to understand the background, vocabulary and assumptions of the different solutions. This list cuts across the declarative versus procedural debate (#1 is clearly procedural, #3 is clearly declarative, the others could go either way).
[UPDATED 2009/6/23: Stu has a somewhat related (similary structured but much more entertainingly writen) list of Cloud Computing approaches. I feel good that I have one more item in my list than him.]