Azul’s Network-Attached Processing to Shake Up Server Market
October 4, 2004 Timothy Prickett Morgan
The advent of n-tier computing, where the database, application, and presentation layers of an application are spread out over multiple, incompatible servers, has allowed much more diverse and flexible applications than what would have been possible or affordable using central hosts or even simple client/server architectures. But n-tier architectures have their limitations, and Azul Systems, a startup secretly founded in April 2002, thinks it has an answer to the nightmare of managing those application servers.
Given that Azul, based in Mountain View, California, was founded by Stephen DeWitt, whose Cobalt Networks server company practically invented the server appliance business, you would expect that Azul’s solution to the n-tier quandary would have an appliance angle to it. And the network-attached processor scheme Azul will pioneer when it ships its specialized application servers, sometime in the first half of 2005, do just that.
DeWitt has had some time to think about what to do with his time since he sold his Linux server appliance company to Sun Microsystems at the peak of the dot-com boom in September 2000 for $2 billion. And he has gathered a team of technologists to attack the application tier and simplify it, one of them being Scott Sellers, who is chief technology officer of Azul and is a chip designer from 3dfx Interactive, a graphics card maker that helped to commercialize 3D graphics in the late 1990s, then rose to $400 million in sales, and went bust in 2000 as the retail IT market headed south.
DeWitt and Sellers say that they identified one big problem and one big trend, which, together, presented the technologists and backers of Azul with a substantial business opportunity. The big problem is that application servers, despite the advent of provisioning, virtualization, and autonomic systems management programs, are expensive and difficult to manage. That is just the nature of heterogeneous computing, and if you think of computing as a server operating system doing computing, it seems difficult to imagine how it could get much better. The trend that Azul was watching, however, was that more and more applications are being created to run in virtualized runtime environments, specifically Sun’s Java/JVM/J2EE combo and Microsoft‘s C#/CLR/.NET combo. The VMs ride above the hardware and, depending on how you want to look at it, above, inside, or side-by-side with the operating system level. (It’s my guess that the Azul box has a stripped down Linux kernel and a homegrown JVM. I am also guessing that .NET support will come through the open source Mono clone of .NET.) According to a report from Gartner that Azul cited as it officially launched its company last week, 80 percent of all new e-business applications will be written for Java or .NET by 2008. And that means an intrepid vendor that can make a server appliance that just provides the Java or .NET VM layer can shake up the server business and quite possibly take a big bite out of the future general-purpose server sales that might have ended up supporting Java or .NET application servers had the Azul machine (or the inevitable others that will be announced that will behave like Azul machines) not existed.
“We want to bring to computing what EMC and NetApp have brought to storage,” says Sellers. Having identified a big problem (doing capacity planning on application servers is a pain, and so is maintaining them), the Azul team took a hard look at why server appliances have not become the norm and why network-attached storage has become a vibrant part of the storage business. Sellers said that Azul figured out that one of the main benefits of NAS arrays is that they don’t care what attaches to them. So any Azul application appliance would have to provide the transparency that later NAS devices offered and that earlier NAS gear, which was technologically superior in many ways, did not. In addition to the transparency and heterogeneity, an application appliance would have to offer a very large amount of processing capacity that could be partitioned dynamically as servers requested processing of Java or C# code. The answer, Azul reckoned, was not to use many blade servers, but to have very large SMP servers that presented a single image to servers and that could have their resources configured on the fly to meet changing needs. The idea, says Sellers, is to have an Azul box more or less eliminate the problem of doing capacity planning at the application server tier. It’s so big and so inexpensive (relatively speaking) that you just stop caring.
Sellers says that when Azul was just getting started, the team sat down and looked at using off the shelf components, such as X86 processors from Transmeta or AMD. With the former, you could not build a cache-coherent SMP machine, since the Efficeon chips do not support SMP, and, moreover, the Efficeon is not a 64-bit chip. With the current Xeon or Opteron chips, you can have 64-bit processing (which is necessary because VMs need lots of main memory), and you can even build a reasonably powerful SMP. But there are problems. These processors have a lot of electronics to support legacy code and do things that a virtual machine environment like a JVM or CLR runtime do not need. Moreover, these architectures do not scale very linearly, and the best anyone could do is to make a big 32- or 64-way SMP box like the current high-end server vendors already do. When Azul was talking about big, cache-coherent SMP scalability, it had a much larger box in mind. And that’s why Azul decided to take the hard road and build a custom processor to do only those functions it would need, like garbage collection or running object-oriented programming routines, to support VMs and their application workloads.
So how big will the Azul box be when it starts shipping, sometime in the first half of 2005? Try 384 processors with 256 GB of main memory, all solely dedicated to running Java and its VM environment. (Support for the .NET environment is not expected until a later date.)
Sellers and his chip team have created their own processor with its own instruction set. This processor has 24 cores on a single die, which suggests that Azul is taking an ultra-minimalist approach. Azul is not saying what the clock speed or cache structure will be on this processor, nor will it say what chip process it is using to create it. But the odds favor a 130 nanometer or 90 nanometer process, and each core should be extremely simple, compared with general purpose processor cores, like the Xeon, Sparc, Power, Opteron, or Itanium, which dominate IT today. All that Sellers would say about the chip at the heart of the Azul box is that it will not be using the bleeding-edge technology embodied in the 65 nanometer processes that have not been perfected as yet. Azul has chosen Taiwan Semiconductor Manufacturing Company as its foundry. A single Azul appliance will have 16 of these 24-core chips and a maximum of 256 GB of main memory, all in a single SMP image, packed into an 11U form factor. And, unlike NAS arrays, which are stand-alone boxes that cannot be seamlessly networked and virtualized to look like a single NAS to the network, the Azul compute arrays will be able to be scaled, thanks to virtualization software, and that just by plunking in another Azul box and plugging it in, customers can double their computing power, Sellers says.
Having bought one of these Azul boxes, what customers will do is point their WebSphere, Weblogic, Java Enterprise System, JBoss, or whatever J2EE application server layer, at the Azul box, which will look like any other JVM running on the main production machines or in the network. (When .NET support is delivered “further down the road,” the Azul machine will be able to host Java and .NET workloads side by side.) The Azul server will link into networks through four Gigabit Ethernet ports, but, Sellers says, the company can tweak the system to support InfiniBand or 10G Ethernet–whatever customers want. “We don’t have religion,” says Sellers. “We couldn’t care less who wins the battle of 10G Ethernet versus InfiniBand. It’s all just I/O to us.”
DeWitt says that he believes this network-attached processor idea will evolve into a hot area, and that Azul will have a “material lead” over competitors, which will jump in once they see the merit of the concept. He also believes that his team knows how to stay ahead of the pack. “The rate of evolution at Azul will be swift, and our lead is not measured in weeks or months, but in many, many quarters,” he says. Azul is being cagey about pricing and performance for these VM appliances, but DeWitt says that the company will sell the Azul boxes itself with a “high-touch, direct” method, which Cobalt also used. The company is also getting together an integrator and reseller channel plan, and it is debating whether it can sign distribution deals with the big server vendors. As for pricing, all that DeWitt would say is that the cost of computing on an Azul box would be “monumentally less expensive, in terms of capex and opex,” compared with general-purpose servers, referring to the initial capital expense of buying a server, as well as the operating expenses associated with its use.