Galaxy remakes Sun's server strategy

(continued from previous page)

Bechtolsheim's skills, it's hard to stand out: "They're going to struggle to provide the proof of why this is so different and so unique in the marketplace to get customers comfortable buying from HP, Dell or IBM to consider it."

Dave Turek, vice president of deep computing at IBM, predicted a tough business case for Sun with the low-profit margin x86 server business. "There's a short-term promise of some growth, but it's unclear it will be matched with growth in profitability," he said, and he predicted Sun's x86 work will dilute its ongoing Sparc-based server work.

And though Opteron has some advantages, Intel enjoys pointing out its own product line strengths: the company's move this year to a chip manufacturing process with 65-nanometer features, while most of the rest of the industry is still at the earlier 90 nanometers; integrated chip features for virtualization, management, security and communications; and partnerships and products to help software developers.

A rocky x86 start
Sun's first x86 servers were largely a dud--the LX50, introduced in 2002, was obsolete almost before it shipped, and successors such as the V60 and V65 didn't fare much better.

Sun failed to realize that customers demanded more than generic systems, Bechtolsheim said: "Success in this market isn't free. We can't just rebadge or relabel third-party systems and deliver value to customers."

But things changed in 2003 with Opteron, whose built-in memory controller and 64-bit memory addressing features gave it a leg up over Intel's dominant Xeon. Sun dropped its Intel designs and pared its x86 line down to the dual-processor V20z and four-processor V40z.

Those systems were designed outside Sun and lacked many high-end features the company's customers liked, such as remote management abilities and components that could be swapped without shutting the server down. But they helped Sun boost its x86 market share considerably and take advantage of AMD's switch earlier this year to dual-core chips--models that combine two processing engines onto a single slice of silicon. Intel's mainstream dual-core Xeon chips are scheduled to arrive later this year for four-processor systems and early next year for dual-processor systems.

Now, with the Galaxy overhaul, Sun sells three in-house Opteron servers, all rack-mounted models: the Sun Fire X2100, a low-end 1.75-inch thick system with a single processor; the Sun Fire X4100, a 1.75-inch thick system with dual Opterons; and the X4200, a 3.5-inch thick model with dual Opterons and more storage capacity. All the systems include Solaris.

The X4100 has a starting price of $2,195, but a midrange configuration with dual Opteron 254 processors, redundant power supplies and 2GB of memory costs about $5,095. The X4200 starts at $2,595; the price increases to $7,795 for two dual-core Opteron 275 processors, 4GB of memory, two 73GB drives and redundant power supplies.

Sun also is selling the X2100 for $40 per month if purchased with a three-year service plan.

The X4100 and X4200 use 2.5-inch Serial Attached SCSI hard drives--a new, smaller standard that eases air flow and therefore cooling. The machines can be completely managed remotely using a service processor and the built-in networked keyboard-video-mouse controls. They use hard drives, fans and power supplies that can be replaced while the server is running. And software certified to work on one 4000-series machine will run on the others.

Reliability for x86 may sound wierd, but there is need for it

Mark Hudson from HP said: "They're going to struggle to provide the proof of why this is so different and so unique in the marketplace to get customers comfortable buying from HP, Dell or IBM to consider it."

Comment: If Sun (or whoever else) will provide reasonably reliable 4CPU x86 servers, it will be a superhit. Current 4-way x86 servers from HP and IBM are just a joke - we have hardware MTBF of less than 2 months for x365 and 3 months for DL585! And it's not just a singular failure on a single server, it's a normal thing - we've got at least 3 different motherboard failures and 1 RAID controller failure. What's wierd, failures happen mostly on heavily loaded boxes (has it something to do with heating?). What's even more wierd, it's that MTBF for 4CPU boxes is about an order of magnitude worse than MTBF for cheap commodity 2CPU boxes (even with 2CPU ones running @ 50-100% avg CPU for weeks).

So having reliability in place (preferably with real MTBF measurements) would be great for x86 market - even if this reliability comes with a price (there are lots of mission-critical and pretty reliable Windows server applications out there, and having more hardware failures than application/database/OS failures combined is definitely not a good thing).