Azul Systems is coming to Europe. The company, which sells what it calls network attached processing -- think NAS - was set up in the dog days of 2002 by CEO Stephen DeWitt, and succeeded in acquiring funding despite the economic climate of the time.
Among the promises for its hardware, which has just started shipping, are at least 50 percent consolidation of data centre servers. It works on the premise that a large and growing percentage of enterprise applications run inside Java VMs or using .Net frameworks - in other words, inside a virtual machine that neither knows nor cares about the underlying hardware or OS.
The problem it's addressing is that those boxes are expensive to buy and maintain, and that they're not designed to do what is now being asked of them. "There's now a significant disconnect between systems' architectures and the way people build applications", said DeWitt. "Java-based applications are highly multi-threaded but today's CPUs are all about minimising single thread latency. Yet Intel will maybe have an eight-core chip by the end of the decade."
DeWitt said: "We believe its critical to transform how CPU power is delivered. Look at the IBM/360 mainframe: ever since then we either cluster small servers or build big iron. We cut through those two axes of computing. How was it done in networking and storage? Everyone shares a common fabric. NFS allowed us to take storage out of the server and treat it as a unified storage fabric that is easily scalable.
"The same thing can apply to processing. Existing servers, unmodified, can mount shared processor resources - no-one has done that before."
Customers acquired so far include Wall Street banks, telcos, big web properties, reservations specialist Pegasus, and outsourcer and data centre management company EDS.
What Azul is selling is a box, purpose-built for accelerating JVMs -- and sometime later, .Net machines too. It's built using simple, proprietary 64-bit processors that don't need to run x86 or other large-vendor code, so they're optimised for virtual machine work, such as multi-threading and garbage collection. All the chips are multi-core and can be strung together in boxes that offer up to 384 cores in a single 11U rack. And because they can achieve much higher throughput than an x86 or SPARC box, according to DeWitt, it means that power consumption is reduced. There are also fewer pieces of hardware to maintain - all of which lowers costs, ROI, Capex and Opex, said DeWitt.
"It costs $89k for a 96-way server to $500k for 384-way. On TCO - management cost, power, space, moves adds and changes - all represent extraordinary savings."
The way it works is to replace a server's JVM with Azul's own. "It's 100 per cent certified as Java-compliant", said DeWitt. This means, he said, that there are no religious issues, that no-one has to re-write their code -- the mistake made by Transmeta, the only other vendor that in recent times tried -- but failed -- to attract mainstream buyers to a new processor architecture.
The big plus, according to DeWitt, is that, like NAS, it's scalable. "When you need more computing power for your enterprise applications, you plug in another box, and the Azul JVM discovers it automatically and uses it as required", said DeWitt.
"We have created a segmented VM that loads our VM onto a standard server. It discovers the compute pool, and moves the entire VM and application onto our machine in a one-off operation. This then runs on our hardware, and it runs fast because it's 64-bit. We can run at one-millisecond granularity on our multi-core processors which are designed to be shared by thousands of applications sharing a common pool.
"We took the ASIC approach. This means that, if more power is needed, we can use the silicon real estate to throw more CPU cores at the problem, which was only possible with 130nm geometry. It means lower power consumption - our 96-way is in a 5U form factor and uses only 700W. Even the 384-way uses only 2600W", DeWitt said. "Compare that to a standard rack of servers."
Pluses and minuses
Disadvantages of the multiple processor approach to adding computing power include the overhead created in trying to manage large numbers of threads, the question of added latency in moving processes around the data centre, and the thorny issue of software licensing.
On the issue of overhead, DeWitt said: "We have built a full command and control system. We are mainframe-like, which means we can set application schedules, maxs & mins for processing power as well as memory. It's all under control.
"We also have a big bucket of sharable 64-bit memory and our full management environment gives holistic view of how much power you have, how its shared out, and the ability to bill back to departments. We've put as much effort into management as any other aspect."
As for latency, DeWitt said, "We do introduce latency because we're a hop on the wire - but you only notice that when the system is not loaded, and people care more about what happens when the system is loaded. We eliminate the latency that occurs on every server - when it comes to processing, we don't have one revolving door, we have 384 of them with no queuing for garbage collection."
On licensing, DeWitt was less bullish. "This problem is getting worse by the day", he said. "But Microsoft's licensing along with all the others is changing. For instance, if you're running BEA on a Lintel box, we're cost-neutral. But every ISV's model is different, so the per-CPU licensing model will evolve rapidly. We have licensing deals with major vendors including IBM, Oracle, and SAP."
So, can Azul sail on where others have foundered? It's a big gamble, which DeWitt admits. But as the founder of Web server vendor Cobalt Systems, he got richer to the tune of $2 billion by selling Cobalt to Sun - it was one of the last big sales of the bubble. He argues that, if anyone knows how to sell appliances to the data centre, he must be numbered among them.
On the plus side, the system is about as compatible as it needs to be. When the platform from the enterprise perspective is the JVM, the underlying hardware becomes much less of an issue. Real world usage will be the ultimate test.
Can he beat IBM, Sun, Dell and HP? Tough one; it depends on their response - including the possibility that a deep-pocketed competitor might decide to take out the young pretender. DeWitt said he was alive to that eventuality. An alternative would be for IBM, Sun or HP to develop something similar, but it might be hard, given both their internal dynamics and existing customer bases, to market it convincingly.
Maybe the biggest question mark is timing. Is the enterprise ready for so radical an approach? Data centre managers didn't get where they are by taking risks. On the other hand, if Azul's small but growing list of enterprise customers are prepared to go on the record, then others will surely follow suit.
We might look back in two years and agree that Azul's truly is a disruptive technology.