Slashdot Mirror
Mainframe Programming to Make a Comeback?
ajw1976 writes to tell us that IBM has released a series of announcements today "introducing many new software tools, academic programs, and support for outside developers." The new releases are designed to help entice programmers and businesses back to the mainframe. From the article: "The announcements, according to analysts briefed on them in advance, signal a shift from defense to offense in the company's mainframe strategy. Last month, I.B.M. introduced a machine priced at $100,000, about half the previous starting price for its mainframes, which can run up to several million dollars. The announcement of the low-end mainframe was made in China, which I.B.M. regards as a promising market for the machines."
""Imagine what a beowulf cluster of these things could do?""
Probably not a whole lot.
High performance computing is not a Mainframe's purpose. A typical personal computer is going to have a much more powerfull proccessor then what you'd find in your average mainframe.. Of course if you have a few million dollars laying around you can find all sorts of stuff that is blazing fast.
The thing that Mainframes are good at are I/O. That is sorting and managing massive amounts of information. You'll have transactions and records being sorted that are numbered not in the thousands, but in the tens or hundreds of millions.
Also you have all these intellegent peripherals.
For instance in PC-land typically scsi drives are faster then SATA drives.
This isn't because SCSI is so much faster or using space age materials. (although they tend to last longer because they are simply better built to higher tolerances and also this allows them to spin faster.)
SCSI and SATA use pretty much the same technologies to do the same stuff. Same materials, same most anything. What makes them faster is the intellegent controllers and I/O bandwidth (although not so much anymore).
Mainframes are like that. Everything has a built in proccessor that does it's share of the workloads. All these intellegent controllers for all I/O. network offload. Disk activity offload. All this stuff.
Like a big brachiosaurus vs a monkey.
The dinosaur is the mainframe, the cat is the monkey.
So a beowolf cluster is like a thousand monkeys on a thousand typewriters.
A brachiosaurus supposadly had multiple special brains or nerve clusters around it's body. That way you have a controller in it's ass to control the back legs and the tail. A controller to control the multiple stomachs. etc etc.
That way it could live with a brain the size of a walnut.
So if you want to decode the genome, use the monkeys. If you want to move mountains, use the dinosaur.
It turns out to be a lot like mainframe computing in terms of physical infrastructure and administration, and in fact often takes over disused mainframe computing centres, at least in the university space.
Unlike the mainframe environment, anyone with Unix/Linux experience is already equipped to take full advantage of cluster and grid computing. Either enviroment provides specialized resources that you have to learn how to access, but to me, the advantage goes to whichever environment provides the most universal expression of those resources, and is least likely to lock my efforts into one particular architecture.
A mainframe is an especially proprietary architecture. Portability has never been its strong point. Conversely, most cluster computations that I've seen have been quite trivially ported from one cluster environment to another. And to some degree it's in every vendor's interest to make it so.
The exceptions are interesting but, at this point, surprisingly rare. Relatively few researchers are decomposing problems in a way which requires either MPI or shared memory. Perhaps the field is not mature enough for that yet, much less for the sorts of computation envisioned by the Grid community, though that day will eventually arrive.
What I mean is, the biggest market for massive computation is always going to be driven by ordinary computation which happens to operate at a massive scale. And for that, the plainer, more symmetric, and more standardized the architecture, the better, because development and testing costs are not going to go down in the face of massive computing resources, they're going to go up.
The perfect mainframe, in other words, is one node in a Beowulf cluster. And that's fine. Just don't go running MQ Series on it, okay?
Parity: What to do when the weekend comes.
Exactly - this is the wrong tool for the job. The hardware is fantastic - and yes, I've never seen a hardware problem, though crappy code (waves hand) can hork an instance. One of the machines I use is a eServer zSeries 900. Max of 16 CPU's, and this one has less than that - think it has 6, but not sure. Not that they would ever *allocate* all the MIPS my direction.
But lets say I had some stupid money. From the website, the latest greatest box...
If you need to crunch hard numbers - especially in parallel - there are much better options out there for the money. The folks a few miles down the road from me do a fantastic job with large Opteron clusters (waves to Malice). The mainframe is not the hardware you want when it comes to getting the math on.
IO and uptime... that is another story...
+++ UGUCAUCGUAUUUCU
If you open a Sun E10000, you'll find that it looks a lot like a mainframe on the inside.
Yeah, I've seen 'em. Sun E10Ks are practically mainframes. And they cost about as much, too.
Of course, when it comes to raw transactional throughput, your average E10K running Solaris and Oracle just doesn't hold a candle to, say, an IBM z9 Enterprise Class running z/OS and DB2.
My blog