The Unix Guardian--Sun's Opteron-Based Galaxy Servers Launch September 12

Newsletters

Subscriptions

Forums

Store

Career

Media Kit

About Us

Contact

Home

Volume 2, Number 32 -- August 25, 2005

Sun's Opteron-Based Galaxy Servers Launch September 12

by Timothy Prickett Morgan

The great mystery surrounding the much-anticipated "Galaxy" line Opteron-based servers is about to be resolved as Sun Microsystems is expected to make the Galaxy machines the star attraction at its Network Computing 2005 Q3 announcements in New York on September 12. Sun is not likely to launch its "Panther" UltraSparc-IV+ processors, which are an improved version of the dual-core Sparc chips that Sun started shipping last year, but could change its mind.

While high-end Sun Fire server customers and those who are enamored with the Sparc architecture (quite possibly because they have code that has not yet been ported to Solaris/Opteron) might be eager to see the Panther chips come to market at this Sun announcement, they may not (a few months ago, Sun was saying they would come out in the fourth quarter), since Sun has only said that it would get the Panther chips to market at then end of 2005 and roll them out across the Sun Fire server line into early 2006. That could mean Panther chips and their related "Amazon+" server frames come out in an announcement, perhaps in November or December, which is about when Sun is expected to do its NC05Q4 product announcements.

The invitation from the Sun event, which I received this week, says merely that Sun is announcing "record-breaking servers for the data center," which doesn't say much. And intentionally so. People at Sun are not saying beans about what will be announced, except that there will be "network-driven desktops." I think that means when a Windows virus takes over your desktop, but perhaps not. (If you want to find out more about the Panther chips, see the links below in the Related Stories section.)

In any event, the Galaxy machines are widely expected to be the cornerstone of the show, and very little is known about them except that they are based on commodity technologies that Andy Bechtolsheim, one of the five founders of Sun, is weaving into new kinds of systems. Bechtolsheim was Sun's original chief technology officer, and he came back to the company when Sun acquired super-secret Opteron server startup Kealia back in February 2004, which was just after Sun announced it was going to adopt the Opteron processors from AMD for its entry and midrange server lines.

The rumor mill is suggesting that the Galaxy machines will be comprised of 1U and 2U form factor rack servers that have two or four processor sockets per box. Sun will obviously support single- and dual-core Opterons in such machines. But for Galaxy machines to be different from the current Opteron-based Sun Fire V20z and V40z servers--and therefore be more appealing to server customers who currently do not buy X86/X64 machines from Sun--they will have to have some pretty compelling architectural advantages over regular rack servers based on Intel chipsets and Intel processors.

Back in June 2004, John Fowler, general manager of Sun's Network Systems Group, which is responsible for X64 servers at Sun, said that future Opteron machines would be based on an idea called "diagonal scaling," which seeks to strike a balance between the vertical scaling of big SMP systems, like the Sparc-based Sun Fire machines, and the horizontal scaling of Unix and Linux clusters, some of the latter of which are based on Sun's own Opteron servers these days. In many cases, what workloads really need are big four-way or even eight-way nodes that are clustered together reasonably tightly, using very fast interconnection, but because they are clusters, they are not as costly to build as big SMP boxes. Still other customers might do well having such tight coupling across a number of compact, two-socket servers that sit next to each other in a rack.

Sun knows how to cluster such machines--it created its own "WildCat" clustering for Sparc-based supercomputers and its knows lots about InfiniBand and, now, HyperTransport--and I think there is a fair possibility that Galaxy machines will not be standalone machines so much as machines that can be linked together using NUMA technologies in the event that customers need to create an more scalable server on the fly. If Cray can marry AMD's HyperTransport architecture to a proprietary high-bandwidth, low-latency interconnection scheme for both its XD1 (OctigaBay) and XT3 (Red Storm) supercomputers, I would be willing to bet that Sun can do something similar to make such machines by augmenting HyperTransport. For all anyone knows, Sun has adopted the "Horus" chipset being developed by the Newisys unit of contract server manufacturer Sanmina-SCI; the Horus chipset scales from four to 32 processors in a single system image. Sun's V20z and V40z servers already come from Newisys, but that is no guarantee that Sun will use the Horus chipset in the Galaxy. Newisys was expected to get machines based on Horus out in mid-2005, which is about now. We'll have to see.

I also think that the Galaxy boxes would have a sophisticated blade server design with integrated storage arrays, integrated switching and server interconnects for scaling, and lots of server and storage virtualization features. And, because Sun is worried about power consumption and is interested in making "green machines" (just like I am), I think that at least some of the Galaxy boxes (if not all) will use the low-powered variants of the Opteron processors and work on 110/120 volt power rather than the 220/240 volt power that most blade server chassis require. Most rack servers use 110/120 volt power, but they do not offer the density and shared power of a blade chassis. It would also be interesting if the Opteron-based servers and the future "Niagara" Sparc-based servers, due in early 2006, could be lashed together--one to do infrastructure workloads at very low power, the other to do heavier workloads at more power, but with a certain amount of pep.

So I expect Galaxy machines to be a hybrid between blade servers and rack-based SMP servers that employ NUMA connections for scalability. The NUMA links for diagonal scaling is not a new idea, by the way. IBM's current "Squadron" server line, using Power5 processors and embodied in the pSeries and iSeries servers, uses such interconnections build out servers in blocks of four cores. You use fast light pipes and special memory and I/O bus extensions to extend a four-core server into one with eight, 12 or 16 processors in the p5 570 and i5 570 lines of servers. Moreover, IBM is on its third generation of "Summit" xSeries chipsets, which do exactly the same thing, and with the latest chipset, the "Hurricane" X3, IBM can scale up to 32 Intel Xeon processors. That Sun would provide similar extensibility--and one that might be reconfigurable--for the Galaxy machines would make perfect sense.

Finally, given that Sun has invented the "uniboard" concept for its Sparc-based Sun Fire servers, which has been a boon to easy upgrading for customers, I find it hard to believe that Sun would not create the Galaxy machines using a uniboard architecture. A uniboard is a cell board that has a chipset, processor sockets, and main memory all on the same card. Uniboards from early UltraSparc-III processors dating from late 2001 are compatible with more recent Sun Fire machines, which allow a certain amount of investment protection. These uniboards are not strictly speaking blade servers, but they are hot-pluggable and they can be added and removed on the fly. A Galaxy uniboard might have two or four sockets, and I am betting on two sockets, based on the fact that the AMD 8000 series chipset is based on gluing two-socket boards into servers with two, four, or eight sockets using NUMA links. But the AMD 8000 chipset and the Horus chipset both support four-socket cell boards, too, which means Sun could also do some four-socket uniboards.

But remember: Most of this is speculation for the sake of amusement.

One last thing: There is another rumor going around that Sun might call these Galaxy boxes the Sun Fire X Series. This cannot be true--and if it is, it should not be considering that rival IBM is already using the xSeries moniker. What on Earth is wrong with calling them the Galaxy servers? I have a hard time believing anyone could trademark the word "galaxy." But then again, with lawyers and judges and government bureaucrats working together, anything nonsensical is not just possible, but probable.