A Dose of Opteron

 
 
By Chris Preimesberger  |  Posted 2006-07-17
 
 
 

Sun Microsystems next steps into the world of x86 servers simultaneously include ventures into the most competitive area of the market and another space essentially abandoned by rivals.

The Santa Clara, Calif., systems maker on July 11 rolled out the latest of its Opteron-based Galaxy servers, including a blade server and another system that can scale from four to 16 processors.

The third is a hybrid data server, a combination of both traditional server and storage technology.

The Sun Blade 8000 represents Suns first foray into the blade market since it pulled its SPARC-based Sun Fire B1600 last year. The blade space is among the fastest-

growing server markets, with research company IDC predicting it to grow to $15 billion by 2009. In the first quarter of 2006, IBM topped the market with 40.1 percent share, followed by Hewlett-Packard at 35.6 percent and Dell at 11.1 percent.

Meanwhile, the Sun Fire X4600 can scale from four to 16 processors, representing Suns entry into a part of the industry—the x86 space for systems with eight or more processors—largely abandoned by its rivals. Dell servers scale to four sockets, while HPs x86 ProLiants also scale up to four sockets, with anything larger coming from its Itanium 2-based Integrity line.

David Lawler, director of product definition and strategy at Suns Systems Group, said previous eight-way systems were primarily 32-bit servers that lacked the memory footprint for such scale-up environments. However, with the dual-core capabilities—moving to quad-core in 2007—of Advanced Micro Devices Opteron, and with 64-bit computing features, those concerns are gone, Lawler said.

It was the X4600 that the Tokyo Institute of Technology used to build TSUBAME, the supercomputer that ranks as the seventh most powerful computer in the world. The supercomputer—a cluster of systems connected via InfiniBand—uses 10,480 Opteron chips and has a sustained peak of 38.18 teraflops (trillion floating-point operations per second).

"[Sun] competitors did not have x86-based fat nodes, i.e., those with a large number of CPUs and large memory," said Satoshi Matsuoka, the professor responsible for computing infrastructures at the schools Global Scientific Information and Computing Center. "This is necessary because good, general-purpose supercomputers are typically built out of such fat nodes, giving various benefits both from user and system administrative perspectives, such as various algorithmic advantages by having large shared memory, support of both shared memory and message-passing programming models, [and] lower node count for reliability and manageability."

The last of the new systems is the dual-core Opteron-based Sun Fire X4500 archival storage server, which can hold up to 24TB of data within its 48 hot-swappable disk drives. Code-named Thumper, it has recorded high throughput numbers in testing: 1GB per second from disk to network and 2GB per second from disk to memory, said Andy Bechtolsheim, systems designer and the architect of the new servers.

The systems will add to the one- to four-socket servers Sun already has in its Opteron-based suite, a key part of Suns makeover as it attempts to return to profitability after several years of operating losses. Sun officials pointed to continued gains in sales and revenues in its x86 business, and said the new servers represent a marked improvement over similar servers from competitors.

"These are all completely unique in their architecture and in their performance," Bechtolsheim said during a press preview June 29. "They represent an entirely new chapter in the history of Sun Microsystems."

The servers, internally code-named Andromeda, are different from previous Sun Fire versions in that they have easily upgradable CPUs, much higher I/O capability ("six to 10 times the throughput of most current blades," said John Fowler, executive vice president of Suns Systems Group) and will be easier to service. They will also cost less in the long run, Fowler said.

"We have separated the I/O from the blade CPU," Fowler said. "Thats one of the big differences in our design. Once you put it in, you can replace anything inside it without taking the chassis out. These are built for clustering."

All three systems—available immediately—were built with virtualization in mind, and to help with the heating and cooling issues that are becoming key concerns in modern data centers. The servers use an unusual new flow-through system, in which air is forced directly through the disk drives and processors via narrow channels.

Rocket Fuel