William Vambenepe's blog

Grid computing is moulting and, to no surprise, the new skin has “cloud” written all over it.

That’s one way to interpret the announcement today that HP, Intel and Yahoo are going to launch a compute cloud. Seeing Intel and HP work together on this is no surprise. Back at HP I had some involvement with the collaboration between HP Labs and Intel on PlanetLab.

I have only read the Gigaom article and Steve’s, so this post is not an analysis of the announcement. Just a few questions that come to mind. They can be most concisely expressed by trying to understand the difference with Amazon’s EC2. The quotes below all come from the Gigaom article.

“six physical locations” -> Amazon has availability zones, including the choice of three geographies.

“between 1,000 and 4,000 mostly Intel cores” -> According to this well-publicized story, Amazon can deliver 5,000 servers (each linked to at least one physical core) to one customer without breaking a sweat.

“We want, unlike other partnerships including Google and IBM’s where the lower-level stacks are not provided in a open manner to the world, open access to all levels of the hardware” -> The quote seems to conveniently avoid comparison with EC2 which provides a much lower abstraction level: virtual machines with mountable raw block storage devices. How much lower can you go without handing out access cards to physically walk into the datacenter? Access to the BMC on the motherboard? Access to some internal bus? Remote-controlled little robots that will slide cards in and out of a chassis?

“researchers will be able to access the cloud through a proposal process later this year” -> Ec2 offers pay-as-you go, which tends to be a good driver for people to use the infrastructure efficiently. And of course someone can always give researchers a grant in the form of EC2 rent money.

Just to be clear, I am not belittling the announcement because for one thing I haven’t read much about it and for another I probably know many of the HP Labs people involved and they are part of the “mucho sapiens” branch of “homo sapiens”. I know they wouldn’t bother putting this out if it was nothing more than giving researchers some free EC2 time.

But these are the questions I’ll be trying to answer for myself as I read more about this project.

I have nothing against Animoto. From what I know about them (mostly from John’s podcast with Brad Jefferson) they built their system, using EC2, in a very smart way.

But I do have something against their story being used to set the benchmark for infrastructure flexibility. For those who haven’t heard it five times already, the summary of “their story” is ramping up from 50 to 5000 machines in a week (according to the podcast). Or from 50 to 3500 (according to the this AWS blog entry). Whatever. If I auto-generate my load (which is mostly what they did when they decided to auto-create a custom video for each new user) I too can create the need for a thousands of machines.

This was probably a good business decision for Animoto. They got plenty of visibility at a low cost. Plus the extra publicity from being an EC2 success story (I for one would never have heard of them through their other channels). Good for them. Good for Amazon who made it possible. And who got a poster child out of it. Good for the facebookers who got to waste another 30 seconds of their time straining their eyes. Everyone is happy, no animal got hurt in the process, hurray.

That’s all good but it doesn’t mean that from now on any utility computing solution needs to support ramping up by a factor of 100 in a week. What if Animoto had been STD’ed (slashdoted, technoratied and dugg) at the same time as the Facebook burst, resulting in the need for 50,000 servers? Would 1,000 X be the new benchmark? What if a few of the sites that target the “lonely guy” demographic decided to use Animoto for… ok let’s not got there.

There are three types of user requirements. The Animoto use case is clearly not in the first category but I am not convinced it’s in the third one either.

1. The “pulled out of thin air” requirements that someone makes up on the fly to justify a feature that they’ve already decided needs to be there. Most frequently encountered in standards working groups.
2. The “it happened” requirements that assumes that because something happened sometimes somewhere it needs to be supported all the time everywhere.
3. The “it makes business sense” requirements that include a cost-value analysis. The kind that comes not from asking “would you like this” to a customer but rather “how much more would you pay for this” or “what other feature would you trade for this”.

When cloud computing succeeds (i.e. when you stop hearing about it all the time and, hopefully, we go back to calling it “utility computing”), it will be because the third category of requirements will have been identified and met. Best exemplified by the attitude of Tarus (from OpenNMS) in the latest Redmonk podcast (paraphrased): sure we’ll customize OpenNMS for cloud environments; as soon as someone pays us to do it.

As far as I know, the W3C is still reviewing the proposal that was made to them to create a new working group to standardize WS-Transfer, WS-ResourceTransfer, WS-Enumeration and WS-MetadataExchange. The suggested name, “Web Services Resource Access Working Group” or WS-RAWG is likely, if it sticks, to end up being shortened to WS-RAW. Which is a bit more cruel than needed. I’d say it’s simply half-baked.

There are many aspects to the specifications and features covered by the proposal. Some goodness, some badness and some ugliness. This post analyzes the good, points at the bad and hints at the ugly. Like your average family-oriented summer movie.

The good

The specifications proposed for W3C standardization describe a way to provide some generally useful features for SOAP messages. Some SOAP messages can get very long. In some cases, I know ahead of time what portion of the long messages promised by the contract (e.g. WSDL) I want. Wouldn’t it be nice, as an optimization, to let the message sender know about this so they can, if they are able to, filter down the message to just the part I want? Alternatively, maybe I do want the full response but I can’t consume it as one big message so I would like to get it in chunks.

You’ll notice that the paragraph above says nothing about “resources”. We are just talking about messaging features for SOAP messages. There are precedents for this. WS-Security can be used to encrypt a message. Any message. WS-ReliableMessaging can be used to ensure delivery of a message. Any message. These “quality of service” specifications are mostly orthogonal to the message content.

WS-RT and WS-Enumeration provide a solution to the “message filtering” and “message chunking”, respectively. But they only address them in the context of a GET-like operation. They can’t be layered on top of any SOAP message. How useful would WS-Security and WS-ReliableMessaging be if they had such a restriction?

If W3C takes on part of the work listed in the proposal, I hope they’ll do so in a way that expends the utility of these features to all SOAP messages.

And just like WS-Security and WS-ReliableMessaging, these features should be provided in a way that leverages the SOAP processing model. Such that I can judiciously use the soap:mustUnderstand header to not break existing services. If I’d like the message to be paired down but I can handle the complete message if need be, I’ll set this attribute to false. If I can’t handle the full message, I’ll set the attribute to true and I’ll get an error if the other party doesn’t understand this extension. At which point I can pick an alternative way to get the task accomplished. Sounds pretty basic but it’s amazing how often this important feature of SOAP (which heralds from and extends XML’s must-ignore semantics) is neglected and obstructed by designers of SOAP messages.

And then there is WS-MetadataExchange. While I am not a huge fan of this specification, I agree with the need for a simple, reliable way to retrieve different types of metadata for an endpoint.

So that’s the (potential) good. A flexible and generally useful way to pair-down long SOAP messages, to chunk them and to retrieve metadata for SOAP endpoints.

The bad

The bad is the whole “resource access” spin. It is not actually intrinsically bad. There are scenarios where such a pattern actually fits. But the way that pattern is being addressed by WS-RT and friends is overly generalized and overly XML-centric. By the latter I mean that it takes XML from an agreed-upon on-the-wire interchange format to an implicit metamodel (e.g. it assumes not just that you agree to exchange XML-formated data but that your model and your business logic are organized and implemented around an XML representation of the domain, which is a much more constraining requirement). I could go on and on about this, especially the use of XPath in the PUT operation. In fact I did go on and on with it, but I spun that off as a separate entry.

In the context of the W3C proposal at hand, this is bad because it burdens the generally useful features (see the “good” section above) with an unneeded and limiting formalism. Not to mention the fact that W3C kind of already has its resource access mechanism, but I’ll leave that aspect of the question to Mark and various bloggers (see a short list of relevant posts at the end of this entry).

The resource access part might be worth doing (one more time), but probably not in the same group as things like metadata discovery, message filtering and message chunking, which are not specific to “resource access” situations. And if someone is going to do this again, rather than repeating the not too useful approaches of the past, it may be good to consider alternatives.

The ugly

That’s the politics around this whole deal. There is, as you would expect, a lot more to it than meets the eye. The underlying drivers for all this have little to do with REST/WS or other architecture considerations. They have a lot to do with control. But that’s a topic for another post (maybe) when more of it can be publicly discussed.

A lot of what I describe in this post was already explained in the WS-ManagementHammer post from a couple of months ago. But that was before the W3C proposal and before WS-MetadataExchange was dragged into the deal. So I thought it might be useful to put the analysis in the context of that proposal. And BTW, this is a personal opinion, not an Oracle position (which is true in general for everything on this blog but is worth repeating specifically for this post).

I just found out that I completly missed some interesting information about Oslo-related efforts at Microsoft. Back in February, Mary-Jo Foley reported on a new modeling language (code-name “D”, apparently) that is part of this initiative. And more recently she reported that David Chappell gave a presentation about Oslo (and more generally Microsoft’s SOA plans) at TechEd. He reportedly said that we should expect a new “schema language” (which Mary-Jo thinks is “D”). What I want to know is what its relationship is with SML/SDM and SCA.

Mary-Jo might not know about SCA and SML but I know that David does. He wrote this white paper about SCA and an article arguing that “Microsoft Should Not Support SCA” (based on an a questionable assessment that SCA is only about portability). He and I also had a little back-and-forth about SCA, SML and Microsoft in the comments section of his post. Unfortunately, David hasn’t blogged about Microsoft’s SOA strategy for a while for us non-TechEd people.

In addition to Mary-Jo’s report, the only information I was about to quickly dig out about David’s presentation is this blog post on Microsoft’s Israel site. Looks like David gave the same presentation at TechEd Israel 2008. Anyone who understands Hebrew cares to translate the blog? Fortunately there is a two-minutes video (also available here) in which we can hear David talk (in English). During the second of the two minutes you’ll hear and see something that could come straight out of a SCA presentation…

For some reason, David’s TechEd Israel presentation doesn’t seem to be listed here and TechEd online tells me that “Featured videos are unavailable at this time”. That’s both for IT Professionals and Developers. But of course they forced me to install Silverlight before telling me that.

[UPDATED 2008/8/11: Here is a 14 minutes video interview of David Chappell providing an update on Oslo.]

There were a few announcements relevant to the evolution of IT management over the last week. The most interesting is VMware’s release of the open-source (BSD license) VI SDK, a Java API to manage a host system and the virtual machines that run on it. Interesting that they went the way of a language-specific API. The alternatives, to complement/improve their existing web services SDK, would have been: define CIM classes and implement a WBEM provider (using CIM-HTTP and/or WS-Management), use WS-Management but without the CIM part (define the model as native XML, not XML-from-CIM), use a RESTful HTTP-driven interface to that same native XML model or, on the more sci-fi side, go the MDA way with a controller from which you retrieve the observed state and to which you specify the desired state. The Java API approach is the easiest one for developers to use, as long as they can access the Java ecosystem and they are mainly concerned with controlling the VMWare entities. If the management application also deals with many other resources (like the OS that runs in the guest machines or the hardware under the host, both of which are likely to have CIM models), a more model-centric approach could be more handy. The Java API of course has an underlying model (described here), but the interface itself is not model-centric. So what with all the DMTF-love that VMWare has been displaying lately (OVF submission, board membership, hiring of the DMTF president…). Should we expect a more model-friendly version of this API in the future? How does this relate to the DMTF SVPC working group that recently released some preliminary profiles? The choice to focus on beefing-up the Java-centric management story (which includes Jython, as VMWare was quick to point out) rather than the platform-agnostic, on-the-wire-interop side might be seen by the more twisted minds as a way to not facilitate Microsoft’s “manage VMWare today to replace it tomorrow” plan any more than necessary.

Speaking of Microsoft, in unrelated news we also got a heartbeat from them on the Oslo project: a tech preview of some of the components is scheduled for October. When Oslo was announced, there was a mix of “next gen BizTalk” aspects and “developer-driven DSI” aspects. From this report, the BizTalk part seems to be dominating. No word on use of SML.

And finally, SOA Software (who was previously called Digital Evolution and who acquired Blue Titan, Flamenco and LogicLibrary, in case you’re trying to keep track) has released a “SOA Development Governance Product”. Nothing too exciting from what I can see on InfoQ about it, but that’s a pretty superficial evaluation so don’t let me stop you. Am I the only one who twitches whenever “federation” is used to mean at worst “import” or at best “synchronization”? Did CMDBf start that trend? BTW, is it just an impression or did SOA Software give InfoQ a list of the questions they wanted to be asked?

“If you have a stove, a saucepan and a bottle of cold water, how can you make boiling water?”

If you ask this question to a mathematician, they’ll think about it a while, and finally tell you to pour the water in the saucepan, light up the stove and put the saucepan on it until the water boils. Makes sense. Then ask them a slightly different question: “if you have a stove and a saucepan filled with cold water, how can you make boiling water?”. They’ll look at you and ask “can I also have a bottle”? If you agree to that request they’ll triumphantly announce: “pour the water from the saucepan into the bottle and we are back to the previous problem, which is already solved.”

In addition to making fun of mathematicians, this is a good illustration of the “fake machine” approach to utility computing embodied by Amazon’s EC2. There is plenty of practical value in emulating physical machines (either in your data center, using VMWare/Xen/OVM or at a utility provider’s site, e.g. EC2). They are all rooted in the fact that there is a huge amount of code written with the assumption that it is running on an identified physical machine (or set of machines), and you want to keep using that code. This will remain true for many many years to come, but is it the future of utility computing?

Google’s App Engine is a clear break from this set of assumptions. From this perspective, the App Engine is more interesting for what it doesn’t provide than for what it provides. As the description of the Sandbox explains:

“An App Engine application runs on many web servers simultaneously. Any web request can go to any web server, and multiple requests from the same user may be handled by different web servers. Distribution across multiple web servers is how App Engine ensures your application stays available while serving many simultaneous users [not to mention that this is also how they keep their costs low -- William]. To allow App Engine to distribute your application in this way, the application runs in a restricted ’sandbox’ environment.”

The page then goes on to succinctly list the limitations of the sandbox (no filesystem, limited networking, no threads, no long-lived requests, no low-level OS functions). The limitations are better described and commented upon here but even that article misses one major limitation, mentioned here: the lack of scheduler/cron.

Rather than a feature-by-feature comparison between the App Engine and EC2 (which Amazon would won handily at this point), what is interesting is to compare the underlying philosophies. Even with Amazon EC2, you don’t get every single feature your local hardware can deliver. For example, in its initial release EC2 didn’t offer a filesystem, only a storage-as-a-service interface (S3 and then SimpleDB). But Amazon worked hard to fix this as quickly as possible in order to be appear as similar to a physical infrastructure as possible. In this entry, announcing persistent storage for EC2, Amazon’s CTO takes pain to highlight this achievement:

“Persistent storage for Amazon EC2 will be offered in the form of storage volumes which you can mount into your EC2 instance as a raw block storage device. It basically looks like an unformatted hard disk. Once you have the volume mounted for the first time you can format it with any file system you want or if you have advanced applications such as high-end database engines, you could use it directly.”

and

“And the great thing is it that it is all done with using standard technologies such that you can use this with any kind of application, middleware or any infrastructure software, whether it is legacy or brand new.”

Amazon works hard to hide (from the application code) the fact that the infrastructure is a huge, shared, distributed system. The beauty (and business value) of their offering is that while the legacy code thinks it is running in a good old data center, the paying customer derives benefits from the fact that this is not the case (e.g. fast/easy/cheap provisioning and reduced management responsibilities).

Google, on the other hand, embraces the change in underlying infrastructure and requires your code to use new abstractions that are optimized for that infrastructure.

To use an automotive analogy, Amazon is offering car drivers to switch to a gas/electric hybrid that refuels in today’s gas stations while Google is pushing for a direct jump to hydrogen fuel cells.

History is rarely kind to promoters of radical departures. The software industry is especially fond of layering the new on top of the old (a practice that has been enabled by the constant increase in underlying computing capacity). If you are wondering why your command prompt, shell terminal or text editor opens with a default width of 80 characters, take a trip back to 1928, when IBM defined its 80-columns punch card format. Will Google beat the odds or be forced to be more accommodating of existing code?

It’s not the idea of moving to a more abstracted development framework that worries me about Google’s offering (JEE, Spring and Ruby on Rails show that developers want this move anyway, for productivity reasons, even if there is no change in the underlying infrastructure to further motivate it). It’s the fact that by defining their offering at the level of this framework (as opposed to one level below, like Amazon), Google puts itself in the position of having to select the right framework. Sure, they can support more than one. But the speed of evolution in that area of the software industry shows that it’s not mature enough (yet?) for any party to guess where application frameworks are going. Community experimentation has been driving application frameworks, and Google App Engine can’t support this. It can only select and freeze a few framework.

Time will tell which approach works best, whether they should exist side by side or whether they slowly merge into a “best of both worlds” offering (Amazon already offers many features, like snapshots, that aim for this “best of both worlds”). Unmanaged code (e.g. C/C++ compiled programs) and managed code (JVM or CLR) have been coexisting for a while now. Traditional applications and utility-enabled applications may do so in the future. For all I know, Google may decide that it makes business sense for them too to offer a Xen-based solution like EC2 and Amazon may decide to offer a more abstracted utility computing environment along the lines of the App Engine. But at this point, I am glad that the leaders in utility computing have taken different paths as this will allow the whole industry to experiment and progress more quickly.

The comparison is somewhat blurred by the fact that the Google offering has not reached the same maturity level as Amazon’s. It has restrictions that are not directly related to the requirements of the underlying infrastructure. For example, I don’t see how the distributed infrastructure prevents the existence of a scheduling service for background jobs. I expect this to be fixed soon. Also, Amazon has a full commercial offering, with a price list and an ecosystem of tools, why Google only offers a very limited beta environment for which you can’t buy extra capacity (but this too is changing).

I have been involved in various capacity with five different specifications that define a GET (or GET-like) operation that takes as input an XPath expression used to pinpoint the subset of the XML document that should be retrieved (here is a quick history as of a couple of years ago, more has happened since). And I must shamefully admit that all but one are simply impossible to implement in an interoperable way.

That’s because they instruct implementers to return an XPath nodeset in the response SOAP message but say nothing about how to serialize the nodeset. While an XPath nodeset contains the kind of things that make up an XML document, it is not an XML document by itself. There is an infinite number of possible ways to serialized an XPath nodeset into XML. To have any hope of interoperability on this, a serialization algorithm has to be clearly described by the specification. Which hasn’t happened.

Let’s start with WS-ResourceProperties (WS-RP). It has a QueryResourceProperties operation that takes an XPath expression as input. The specification says that “the response MUST contain an XML serialization of the results of evaluating the QueryExpression against the resource properties document“. Great, thanks. The example provided happens to return a nodeset with only one node (a boolean), which is implicitly serialized into the text representation of that boolean. What if there is more than one node in the nodeset? What about other types of nodes?

Moving on to WS-Management, which defines a SOAP header that uses XPath to qualify a WS-Transfer GET request such that it only retrieves a subset of the target XML document. While it does a better job than WS-RP at describing the input (e.g. it specifies the context node and what namespace declarations are in scope for the XPath evaluation) it is even more cavalier than WS-RP in describing the output: “the output (lines 53-55) is like that supplied by a typical XPath processor and might or might not contain XML namespace information or attributes“. By “a typical XPath processor” we should understand MSXML I suppose. But as far as I know a “typical XML processor” doesn’t return XML, it returns language-specific data structures (e.g. a C# or Java object, like a nu.xom.Nodes instance). And here too, the examples only use single-node nodesets.

WS-ResourceTransfer (WS-RT) was supposed to be the convergence of these two efforts, so presumably it would have learned from their mistakes. While it is better written in general than its predecessors, it fails just as badly with regards to specifying the nodeset serialization. And once again, the example provided uses a nodeset with just one node.

And then came the CMDBf query operation which, for some unclear reason, was deemed in need of a built-in XPath transformation of records. As I pointed out in my review of CMDBf 1.0 at the time, this feature was added without taking the pain to define the XML serialization of the resulting nodeset. And there isn’t even an example of the XPath serialization.

It is sad in a way, but the only specification that acknowledges the problem and addresses it came before any of the four above even got started. It is the WSMF (Web Services Management Framework) work that we did at HP, and more specifically the “note on dynamic attributes and meta information” (not available at HP anymore but available from archive.org) . This specification was the first one to define a GET operation that is qualified by an XPath expression. Unlike its successors it also explicitly narrowed down the types of nodes that could be selected (”The manager MUST NOT send as input an XPath statement that returns a nodeset containing nodes other than element, attribute and namespace nodes“). And for those valid types it described how to serialized them in XML (”When a node in the result nodeset is an attribute node, for the sake of the response it is serialized as an element node which has the same name as the name of the original attribute (see example 4 for an illustration). The element is in the same namespace as the namespace the attribute it represents is in. This applies to namespace nodes as well, they are serialized like an attributes in the xmlns namespace“). Turning an attribute into an element of the same QName might not be the smartest thing in retrospect (after all there may be an element by that QName already) but at least we recognized and addressed the problem.

But all is good now, I am told, because XPath 2.0 is here, along with a clean data model and a well-described serialization.

Not so. Anyone wanting to use XPath for a SOAP-based query language still would have to specify a serialization.

The first problem with the W3C serialization is that the XML output method doesn’t work for all nodesets. Try to use it on a nodeset that contains a top-level attribute node and you get error err:SENR0001. And even for the nodesets it accepts, it sometimes returns less-than-useful results. For example, if your XPath is of the form /employee/name/text() and you have four employees, the result will look something like this:

“Joe SmithKathy O’ConnorHelen MartinBrian Jones”

Concatenated text values without separators. I guess W3C is like a department store, they don’t offer complimentary wrapping anymore…

That’s why the nux.xom.xquery.ResultSequenceSerializer class had to define its own wrapping mechanims to produce a useful XML serialization. The API gives you the choice between the W3C_ALGORITHM and the WRAP_ALGORITHM.

Bottom line, and however much some would like to think of it that way, XPath (1 or 2) is not an XML subsetting/transformation mechanism. It could be used to create one (as XSLT does), but you have to do your own plumbing.

In addition to the technical aspects of this discussion, what else can be learned from this sad state of things? The fact that all these specifications define an XPath-driven query mechanism that is simply broken (beyond the simplest use cases) withouth anyone even noticing tells me that there isn’t a real need for full XPath query over SOAP (and I am talking about XPath 1.0, the introduction of XPath 2.0 in CMDBf is even more out there). A way to retrieve individual elements (and maybe text values) is all that is needed for 99% of the use cases addressed by these specifications. Users would be better served (especially in a version 1.0) by specifications that cover the simple case correctly than by overly generic, complex and poorly documented features. There is always time to add features later if the initial specification is successful enough that users encounter its limitations.

Spring’s Rod Johnson writes today about the future he sees for Java Bloatware (his unkind term for Java EE middleware). Of course, as Mr. Spring, he is far from neutral. Of course he is focusing on a certain class of applications (web-centered, mostly greenfield, which is a huge - and sexy - segment, but not in any way the only kind of applications). Of course he underestimates how established technology that works remains used long after it may have ceased to be the optimal solution for new developments. But even taking all that into account, he makes some good points about the proliferation of rarely-used capabilities in Java EE and the associated cost. Most of those points are well understood and are driving the more modular approach taken by Java EE 6. As well as the adoption of OSGi (see here and here for BEA’s example). In addition, as Rod mentions, the JCP now has to share the playground with other framework standardization efforts like SCA.

The most interesting part of Rod’s post from my perspective, is this prediction:

“The market will need to address the gap between Tomcat and WebLogic/WebSphere. Currently an important part of the market is neglected. The majority of Java web applications are most at home on Tomcat. A minority actually want some of the more esoteric functionality of a full-blown application server, such as JCA, or specialized capabilities such as distributed transaction management. But a larger minority need some of the operational and management features of those products, but are not interested in the esoteric APIs and the bloat they bring along with them. As more and more end user companies look to phase out legacy application servers in favor of better suited technologies, there will inevitably be a response to market demand, with products that hit the sweet spot and bridge this gap.”

Right on. This is the second time in a week that we see an acknowledgment of the importance of application manageability coming from SpringSource. Whether this mid-point demand will be met from the top down by a more modular Java EE stack or from the bottom up by building on top of Tomcat (or some non-Java HTTP server) remains to be seen. The two aren’t exclusive either.

I expect that the hosted application frameworks like the recently announced Google App Engine will also aim at that “more than Tomcat, less than J2EE” sweetspot. But the cost/benefit formula of a more full-featured (or “bloated” if you prefer Rod’s terminology) environment might turn out to be different in a “hosted framework” situation.

I see the tide rising for semantic technologies. On the other hand, I wonder if they don’t need to fail in order to succeed.

Let’s use the Grid effort as an example. By “Grid effort” I mean the work that took place in and around OGF (or GGF as it was known before its merger w/ EGA). That community, mostly made of researchers and academics, was defining “utility computing” and creating related technology (e.g. OGSA, OGSI, GridFTP, JSDL, SAGA as specs, Globus and Platform as implementations) when Amazon was still a bookstore. There was an expectation that, as large-scale, flexible, distributed computing became a more pressing need for the industry at large, the Grid vision and technology would find their way into the broader market. That’s probably why IBM (and to a lesser extent HP) invested in the effort. Instead, what we are seeing is a new approach to utility computing (marketed as “cloud computing”), delivered by Amazon and others. It addresses utility computing with a different technology than Grid. With X86 virtualization as a catalyst, “cloud computing” delivers flexible, large-scale computing capabilities in a way that, to the users, looks a lot like their current environment. They still have servers with operating systems and applications on them. It’s not as elegant and optimized as service factories, service references (GSR), service handle (GSH), etc but it maps a lot better to administrators’ skills and tools (and to running the current code unchanged). Incremental changes with quick ROI beat paradigm shifts 9 times out of 10.

Is this indicative of what is going to happen with semantic technologies? Let’s break it down chronologically:

1. Trailblazers (often faced with larger/harder problems than the rest of us) come up with a vision and a different way to think about what computers can do (e.g. the “computers -> compute grid” transition).
2. They develop innovative technology, with a strong theoretical underpinning (OGSA-BES and those listed above).
3. There are some successful deployments, but the adoption is mostly limited to a few niches. It is seen as too complex and too different from current practices for broad adoption.
4. Outsiders use incremental technology to deliver 80% of the vision with 20% of the complexity. Hype and adoption ensue.

If we are lucky, the end result will look more like the nicely abstracted utility computing vision than the “did you patch your EC2 Xen images today” cloud computing landscape. But that’s a necessary step that Grid computing failed to leapfrog.

Semantic web technologies can easily be mapped to the first three bullets. Replace “computers -> computer grid” with “documents/data -> information” in the first one. Fill in RDF, RDFS, OWL (with all its flavors), SPARQL etc as counterparts to OGSA-BES and friends in the second. For the third, consider life sciences and defense as niche markets in which semantic technologies are seeing practical adoption. What form will bullet #4 take for semantic technology (e.g. who is going to be the EC2 of semantic technology)? Or is this where it diverges from Grid and instead gets adopted in its “original” form?

I heard tonight for the first time of a company called Elastra. It sounds like they are trying to address a variation of the data center automation use cases covered by Opsware (now HP) and Bladelogic (now BMC). Elastra seems to be in an awareness-building phase and as far as I can tell it’s working (since I heard about them). They got to me through John’s blog. They are also using the more conventional PR channel (and in that context they follow all the cheesy conventions: you get to “unlock the value”, with “the leading provider” who gives you “a new product that revolutionizes…” etc, all before the end of the first paragraph). And while I am making fun of the PR-talk I can’t help zeroing on this quote from the CEO, who “wanted to pick up where utility computing left off – to go beyond the VM and toward virtualizing complex applications that span many machines and networks”. Does he feels the need to narrowly redefine “utility computing” (who knew that all that time “utility computing” was just referring to a single hypervisor?) as a way to justify the need for the new “cloud” buzzword (you’ll notice that I haven’t quite given up yet, this post is in the “utility computing” category and I still do not have a “cloud” category)?

The implied difference with Opsware and Bladelogic seems to be that while these incumbent (hey Bladelogic, how does it feel to be an “incumbent”?) automate data center management tasks in old boring data centers, Elastra does it in clouds. More specifically “public and private compute clouds”. I think I know roughly what a public cloud is supposed to be (e.g. EC2), but a private cloud? How is that different from a data center? Is a private cloud a data center that has the Elastra management software deployed? In that case, how is automating private clouds with Elastra different from automating data centers with Elastra? Basically it sounds like they don’t want to be seen as competing with Opsware and Bladelogic so they try to redefine the category. Which makes it easier to claim (see above) to be “the leading provider of software for designing, deploying, and managing applications in public and private compute clouds” without having the discovery or change management capabilities of Opsware (or anywhere near the same number of customers).

John seems impressed by their “public cloud” capabilities (I don’t think he has actually tested them yet though) and I trust him on that. Knowing the complexities of internal data centers, I am a lot more doubtful of the “private cloud” claims (at least if I interpret them correctly).

Anyway, I am getting carried away with some easy nitpicking on the PR-talk, but in truth it uses a pretty standard level of obfuscation/hype for this type of press release. Sad, I know.

The interesting thing (and the reason I started this blog entry in the first place) is that they seem to have created structures to capture system design (ECML) and deployment (EDML) rules. From John’s blog:

“At the core of Elastra’s architecture are the system design specifications called ECML and EDML. ECML is an XML markup language to specify a cloud design (i.e., multiple system design of firewalls, load balancers, app servers, db servers, etc…). The EDML markup provides the provisioning instructions.”

John generously adds “Elastra seems to be the first to have designed their autonomics into a standards language” which seems to assume that anything in XML is a standard. Leaving the “standard” debate aside, an XML format does tend improve interoperability and that’s a good thing.

So where are the specifications for these ECML and EDML formats? I would be very interested in reading them, but they don’t appear to be available anywhere. Maybe that would be a good first step towards making them industry standards.

I would be especially interested in comparing this to what the SML/CML effort is going after. Here are some propositions that need to be validated or disproved. Comparing SML/CML to ECML/EDML could help shade light on them:

• SML/CML encompasses important and useful datacenter automation use cases.
• Some level of standardization of cross-domain system design/deployment/management is needed.
• SML/CML will be too late.
• SML/CML will try to do too many things at once.

You can perform the same exercise with OVF. Why isn’t OVF based on SML? If you look at the benefits that could be theoretically be derived by that approach (hardware, VM, network and application configuration all in the same metamodel) it tells you about all that is attractive about SML. At the same time, if you look at the fact that OVF is happening while CML doesn’t seem to go anywhere, it tells you that the “from the very top all the way down to the very bottom” approach that SML is going after is very difficult to pull off. Especially with so many cooks in the kitchen.

And BTW, what is the relationship between ECML/EDML and OVF? I’d like to find out where the Elastra specifications land in all this. In the worst case, they are just an XML rendering of the internals of the Elastra application, mixing all domains of the IT stack. The OOXML of data center automation if you want. In the best case, it is a supple connective tissue that links stiffer domain-specific formats.

[UPDATED 2008/3/26: Elastra's "introduction to elastic programing" white paper has a few words about the relationship between OVF and EDML: "EDML builds on the foundation laid by Open Virtual Machine Format (OVF) and extends that language's capabilities to specify ways in which applications are deployed onto a Virtual Machine system". Encouraging, if still vague.]

[UPDATED 2008/3/31: A week ago I hadn't heard of Elastra and now I learn that I had been tracking the blog of its lead-architect-to-be all along! Maybe Stu will one day explain what a "private cloud" is. His description of his new company seems to confirm my impression that they are really focused (for now at least) on "public clouds" and not the Opsware-like "private clouds" automation capabilities. Maybe the "private clouds" are just in the business plan (and marketing literature) to be able to show a huge potential markets to VCs so they pony up the funds. Or maybe they really plan to go after this too. Being able to seamlessly integrate both (for mixed deployments) is the holly grail, I am just dubious that focusing on this rather than doing one or the other "right" is the best starting point for a new company. My guess is that despite the "private cloud" talk, they are really focusing on "public clouds" for now. That's what I would do anyway.]

[UPDATED on 2008/6/25: Stephen O'Grady has an interesting post about the role of standards in Cloud computing. But he only looks at it from the perspective of possible standardization of the interfaces used by today's Cloud providers. A full analysis also needs to include the role, in Cloud Computing, of standards (app runtime standards, IT management standards, system modeling standards, etc...) that started before Cloud computing was big. Not everything in Cloud computing is new. And even less is new about how it will be used. Especially if, as I expect, utility computing and on-premise computing are going to become more and more intertwined, resulting in the need to manage them as a whole. If my app is deployed at Amazon, why doesn't it (and its hosts) show up in my CMDB and in my monitoring panel? As Coté recently wrote, "as the use of cloud computing for an extension of data centers evolves, you could see a stronger linking between Hyperic’s main product, HQ and something like Cloud Status".]