SMP? Why Bother?
All your eggs, one basket: Does this make sense?
By Steven J. Vaughan-Nichols
April 22, 1999
I've been thinking a lot about symmetric multiprocessing (SMP). A lot of people seem to think that it's the be-all and end-all of high-end server computing. In other words, if you want super-powered servers, you must be using SMP.
SMP is just another scheme for getting the maximum amount of scalability and performance from one box. It's got some things going for it. As a shared memory, common bus design, it's relatively simple to implement. It's also almost the only show in town for NT to get multiple CPUs pulling together.
It's not just an NT show, though. NetWare does SMP. Linux does SMP. And the commercial Reduced Instruction Set Computer (RISC) Unixes, especially IBM's AIX, does it best of all. Indeed, in cases where someone is bound and determined that SMP is the be-all and end-all of computing, a reseller would do well to set them up with an IBM RS/6000 and AIX solution.
But, say that customers really want more than a hot name; they want the whole deal. You know, performance, reliability and a low price. Maybe, just maybe, instead of trying to convince them that they really need to fork over the big bucks for a quad-Xeon III box with hot-swappable, Level 5 Redundant Array of Inexpensive Disks (RAID), a built in UPS and a kazoo, maybe resellers should point them in the direction of an older, safer technology: clustering.
Clustering isn't as sexy as it used to be. I remember when the department server of choice in many businesses was a DEC VAX/VMS cluster. Even so, clustering, by allowing you to get several servers to work together as if they were one computer, still has advantages over SMP.
My personal favorite is that it doesn't put all your eggs in one basket. There's a school of computing thought that it's OK to put all your computing power in one box if you make it so strong--so that like Jason in the Friday the 13th movies, it keeps coming back no matter what you do to it. That works in the movies. Not real life.
Oh, no question, a properly provisioned SMP box can take a beating and keep on ticking. But, eventually, something will take it down. Properly maintained, however, a cluster of RISC/Unix or AS/400 boxes will keep delivering customer services until the entire facility burns down around them. Clusters are also, in my experience, easier to maintain.
It's also not like clustering is a dead-end. It's still useful technology. Microsoft is continuing to advance its clustering technology. IBM is improving MS Cluster Server, which has been around for years and enables you to cluster a pair of NT servers. IBM's newest project, codenamed "Cornhusker," will empower you to cluster up to eight NT servers at a time.
Other companies are also putting a new shine on old clustering approaches. Sun, for instance, is releasing the first pieces of its Full Moon Clustering. With this technology, and the rest of its datacenter.com plans, Sun wants to make high-end, SMP SPARC clusters an alternative for mainframe customers.
Whether resellers combine SMP and clustering, or just make clustering an alternative to SMP, they owe it to their customers to look beyond SMP for powerful server solutions. SMP's good, no question about it, but clustering has a place as well.