Slashdot Mirror
Sun Considers Opteron
Sanjay writes "Official from Sun spokesman. Sun is considering using AMD's Opteron chip in a server it expects to deliver to the market shortly.
Intead of fighting Win of Wintel (like Redhat is doing), Sun can choose to fight both with Linux AMD's servers and also fight with HP/IBM as Itanium is anyway a non starter.
Sun can rise again!
"
Which sounds like the perfect definition for a Slashback story. We don't need another full-blown story on this just because Sun confirmed it. All we need is a paragraph in Slashback saying, "By the way, remember this story about Sun and the Opteron? Sun's confirmed it."
It never ceases to amaze me how ill-informed people are. Sun has been battered heavily in the stock market, but they aren't "drying" up. There is no way PC architectures will replace "real computers"...they just aren't made to do that kind of job properly. If you're going clock for clock on silly single-user apps, maybe but when are you going to find a partitioned 512 processor PC (in one box) that heals itself and scales as well as Starfire. x86 is junk pure and simple. Opterons are being considered by Sun in the same way as Intel procs made their way into Sun's product list in the form of Cobalt. Sun isn't replacing Sparc with Opteron... that would be utterly retarded and Sun isn' retarded like HP is (read scrapping Tandem, PA-RISC, and Alpha for an unproven architecture that they don't control). Check out the specs on the USIV. Opteron is all about migrating customers off of 32 bit machines onto the Solaris platform. It is a step up to Sparc. Sparc based systems have been 64 bit since Solaris 7 and the Ultra 1 140. That's roughly 1995 or so.
I'm pretty sure Compaq and others already have hot-swap for cards, and support terabytes of storage (which is really just a matter of having enough FC bandwidth for whatever you're doing and plugging into the same standard storage arrays the Sun can).
Hot-swapping CPUs and RAM is trickier, but Sun only offers that on high end models which have no direct counterpart in the PC marketplace. Even then, it's a dicey situation at best.
With the E10K generation, you can hotswap CPU boards (there's 16 of them, each holding up to 4 processors, 4G ram, and two I/O busses (4x Sbus cards or 2x PCI cards). Thus you very much have to plan ahead to make sure you can "swap out" a given board without losing anything (oops, the failed memory is on the board with the only controller for this scsi disk over here, or the only one with this gigabit network connection). Assuming you built the machine right so that no single board is a single point of failure, you hit the next problem: If a CPU or memory module were to actually fail during runtime, it is still just as likely to cause an OS crash. The advantage is that in most cases the offending peice of hardware (1 CPU, 1 bank of RAM, etc) will be blacklisted and not used at all when the machine reboots from the panic (now you have a 15 CPU machine instead of 16). Then after that reboot, you can go about hot-swapping in a replacement with the OS online. You run some commands which basically tell the scheduler to stop scheduling on those CPUs, and tell the VM to not allocate any more physical ram in a certain region - then it goes about paging all the allocated RAM off to other ram or swap until it has emptied the board - then you can swap in the new stuff and re-add the CPU/mem into the OS.
On the newer SunFire architecture (3800s, 6800's, 15K, etc), they finally split the I/O boards from the CPU/Mem boards to make this a bit less painful, thank god. Still, in either case, you dont get a 4-way that scales to 64. You could buy a 64-capable machine (or higher now with SunFire architecture), and only populate it with 4 CPUs because you expect growth - but an E10K with just 4 CPUs in is a huge waste of cash - we're talking at least several hundred thousand dollars, for the hardware equivalent of what other companies sell for just a few thousand dollars. I think at one point a few years ago my company bought one 1/4 configured (16 CPU 16 GB ram) and left the other 3/4 open for expansion, and the cost was on the order of around $1,300,000. Do you really want to pay 50x+ over the same hardware capacity of a top end x86 just to be able to expand and have better support?
And in any case - these solutions, ultimately, may have slightly better sigma numbers on uptime, but they are still riddled with single points of failure, and ultimately no Sun solution is truly reliable with resorting to redundant clustering of oen sort or another. Once you resort to a redundant cluster, you're saying "I don't care if the hardware fails occasionally, my cluster will handle it while we do the maintenance". At that point, are you going to spend that much more money to make the difference between 99.9% and 99.999%?
Lets make a rough real example - a 24/7 Oracle database. In the Sun world, to get 24/7 uptime, you'd build out two machines of appropriate power (let's say 2x 6800s), and drop Oracle's OPS or RAC (or whatever they call the next generation) on it for a fully fault-tolerant cluster. You'd attach it to an FC SAN of appropriately configured redundant storage.
On the x86 side, you'd rack up the equivalent in I/O and CPU horsepower worth of 1U boxes (let's say 32x 1U dual processor large-ram crap-reliability boxes from Penguin Computing or something).
Either one is going to be very reliable because of Oracle's nonstop clustering stuff. You'll experience more failures/year on the x86 solution, but losing one of 32 machines is no biggie for a few hours while you drop in a spare.
Two fully loaded 6800's is gonna run you about $2.0 million. 32 high end-ish (lets say 10K a pop) 1U machines is gonna run you $0.32 million. You do the math.
11*43+456^2