Azul Systems Revamps Compute Appliances with 48-Core Vega2 Chip
December 11, 2006 Timothy Prickett Morgan
Server appliance maker Azul Systems is not letting its legal battle with Sun Microsystems stop innovation with its Compute Appliance servers, which offload Java workloads from general purpose servers. Last week, Azul announced four new systems that are based on its 48-core Vega2 processor, which doubles the scalability of the Compute Appliances.
According to Scott Sellers, one of the co-founders of Azul Systems and the company’s chief operating officer, the second generation of the Compute Appliances includes some other tweaks, such as using common system boards across the line, to make them upgradeable in the field. He says Azul Systems, which announced its Java co-processors in the fall of 2004 and delivered them in early 2005, has over 100 customers using the appliances in field trials to boost the performance of their Java workloads. “We also have some customers who are building their infrastructure starting with the Compute Appliances as the foundation,” he says.
One of those customer wins is at Openreach, the B2B gateway that British telecom giant BT had to set up as the United Kingdom deregulated telecommunications in that country. The Openreach portal will allow third-party telecom providers to access services on the BT network, which was opened up as part of that deregulation.
According to BT’s benchmark tests, which pitted a mix of general purpose Unix servers with Compute Appliances and their Java Virtual Machines against regular Unix servers supporting Java, the Compute Appliances were able to provide four times the throughput on the Openreach portal applications, which have to be accessible or BT gets in trouble with the law. Clive Selley, the chief information officer for BT Wholesale, which operates the Openreach portal, says that even highly tuned Unix servers with Java application servers were “unacceptably slow” compared to the Compute Appliances, which had very little tuning. Neither Azul Systems nor BT would say how big the processing complex that Openreach runs on will be, but did say that as far as they know, it will be the largest B2B gateway in the world.
The Compute Appliances are designed to take advantaged of the multithreaded nature of Java applications. The original Compute Appliances were based on the 24-core Vega1 chip, a homegrown 64-bit processor designed specifically to support Java Virtual Machines using a fairly simple core designed specifically for running virtual machines of many kinds. These original Vega1 machines scaled up to 384 cores in a shared cache, SMP cluster.
With the Vega2 chips, Azul is doubling the core count to 48 per chip, and can deliver 768 cores in a single system image to support many JVMs. Sellers won’t say what the clock speed is on the Vega1 or Vega2 chips–except that it is below 1 GHz–but given that there are a large number of cores on the chip, it is fair to assume the clock speed is very low. Which is fine for multithreaded workloads, since memory speeds are kept more in line with processor core speeds. The CPU is not tapping its foot waiting for memory to catch up and deliver some information. It also looks like Azul Systems goosed the clock speed of the Vega2 chips by about 50 percent, since these boxes are deliver three times the throughput as the Vega1-based machines, which have half the cores.
Sellers says that in moving to the Vega2 chip in the second generation of the Compute Appliances, Azul Systems actually changed the instruction set of the processor cores, but because these cores run virtualized software anyway, this change had no effect on compatibility. This kind of system design–virtualizing at what amounts to the microcode level–is smart, and it is a wonder that more machines are not designed this way. (This approach is akin to how IBM has been able to change processor technology on the System/38, AS/400, iSeries, and System i5 midrange server line for almost 30 years but still maintain application compatibility.)
There are three models in the Vega2 Compute Appliance line. The Model 3210 comes with two Vega2 chips, or 96 cores, on two processor boards. Each board has 24 DDR2 DIMM slots. (The Vega1 machines supported DDR1 main memory.) The Model 3210 supports up to 48 GB of main memory and comes in a 5U rack-mounted form factor. It has two Gigabit Ethernet links, which hook back into the general purpose servers that call on the Compute Appliances to field Java application processing requests. This box, which costs $49,995, burns about 580 watts in a typical configuration, according to Sellers.
The Model 3220 has four Vega2 chips, or 192 cores, and supports from 96 GB to 192 GB of DDR2 main memory for the JVMs. It has four Gigabit Ethernet ports and fits in the same 5U form factor. It actually uses two Vega2 boards instead of one, and burns around 1,000 watts. Next year, Azul Systems will ship boxes that will hold eight and 16 of the Vega2 chips, which means 384 cores and 384 GB of memory and 768 cores and 768 GB of memory, respectively. Presumably, this will be called the Model 3240 and Model 3280 machines, but Azul is not saying for some reason. Boards used in all of these machines are compatible, so customers can start small and build out. They can also mix and match Vega1 and Vega2 systems in the same Compute Appliance pool.
Azul is not announcing pricing on the remaining machines, but if prices scale roughly linearly with performance, that means a top-end box with 768 cores should cost around $400,000.
So what does this kind of money buy? Sellers says that Azul Systems will report that the top-end box will be able to process 872,972 operations per second on the Java-based SPECjbb2005 test from SPEC, one of the industry standard benchmark providers in the server racket. This did not outgun a 128-way PrimePower 2500 server from Fujitsu-Siemens, which hit 1,251,024 operations per second. This machine uses 2.08 GHz Sparc64 V processors from Fujitsu. This Vega2 performance is also a little shy of the 1,164,995 operations per second that a Sun Fire E25000 with 72 dual-core, 1.5 GHz UltraSparc-IV processors handled on the SPECjbb2005 test. But, these two big, bad Unix servers cost many millions of dollars.
The presentation put together by Azul Systems for the Vega2 launch took some swipes at Sun to make its point about the performance/watt and compactness that it could bring to Java workloads. Azul Systems says that a 192-core Vega2 Compute Appliance with 192 GB of memory can deliver about the same performance as eight of Sun’s “Niagara” T2000 servers. Azul Systems does it inside a 5U form factor and within a 1,000-watt thermal envelope. Those Niagara boxes will take up 16U of rack space and consume around 2,200 watts. If you want to use Sun’s Opteron servers, Azul Systems says it will take six two-way, dual-core Opteron servers, in 12U of space and consuming 1,800 watts of power, to do the same job. (Yes, if you choose the right Opteron processor, it can deliver better power and thermals than Niagara.) And finally, it would take a full-sized, 42U Sun Fire E20000 server with 36 dual-core processors, eating up around 20,000 watts, to do the same job as the Azul Systems appliance.
Of course, to be more broadly accepted, Azul Systems needs to have more than BEA Systems certifying its Java application servers to run with the Compute Appliances. Sun has not certified the box, and considering the ongoing litigation earlier this year–Azul sued Sun over its Java licensing practices, and then Sun sued Azul for infringing its patents–it seems unlikely that the Java middleware stack from Sun will be certified to use the Compute Appliances. And even though IBM Global Services is offering support for the boxes that Azul Systems sells, it has thus far not certified its WebSphere middleware to work on the Compute Appliances, either. Sellers says that Azul Systems is “engaged with” IBM, and it could turn out that they work together to upend server sales at competitors. But IBM is obviously worried about self-impact on its own system sales.
The Compute Appliances can interoperate with Linux, AIX, HP-UX, and Solaris operating systems and with WebLogic, JBoss, Oracle Application Server, WebSphere, Apache Tomcat, and Caucho Resin Java application servers. Any kind of network link–Ethernet, Gigabit Ethernet, InfiniBand, whatever–can be used to link Compute Appliances to the general purpose servers that are running the application layer.