| |||||||||||||
|
| |||||||||||||
Ballmer: Microsoft to go after Linux strongholds (ZDNet)Ballmer: Microsoft to go after Linux strongholds (ZDNet)Posted Oct 20, 2005 20:20 UTC (Thu) by drag (subscriber, #31333)In reply to: Ballmer: Microsoft to go after Linux strongholds (ZDNet) by jwb Parent article: Ballmer: Microsoft to go after Linux strongholds (ZDNet)
A while before this they were talking about the 'new' clustering technology that Windows was going to bring to the table.
The idea was was that businesses are going to want to have big easy-to-use clusters for various business applications. Big databases or accounting calculations, or something like that. They figured that you could use clustering with Microsoft's programming tools and .NET-based managed code so that businesses can produce HPC code easier and cheaper then with Linux. I wasn't sure that they are aiming at but they figure they could leaverage company's familarity with windows to open up completely new markets outside the very high end computing clusters that Linux dominates in.
The trouble is though that Microsoft has been trying to do HPC clustering for years. Years and years and years. This isn't something new for them. And still nobody cared. Thru working with Cornell university they've always have had a NT-based cluster in the top500 since the late 90's.
(Log in to post comments)
Ballmer: Microsoft to go after Linux strongholds (ZDNet) Posted Oct 20, 2005 20:27 UTC (Thu) by nix (subscriber, #2304) [Link] They're in trouble then: most accounting calculations parallelize really, really badly: they tend to have huge accumulations at their core, and, well, those depend on everything, so parallelizing them is pointless.There's a reason businesses tend to like mainframes for their accounting stuff --- and, for that matter, there's a reason mainframes have the performance characteristics they do: massive I/O bandwidth and really quite wimpy CPUs. If businesses needed massive parallelization to do accounting, IBM would already have sold it to them. They don't. (IBM have managed a clever trick here of course and also sell the same machines as massive virtual clusters running Linux under a hypervisor, but those aren't generally used to do accounting-crunching type jobs.)
Actually the opposite is true. Posted Oct 21, 2005 1:37 UTC (Fri) by xoddam (subscriber, #2322) [Link] > most accounting calculations parallelize really, really badly:> they tend to have huge accumulations at their core, and, well, > those depend on *everything*, so parallelizing them is pointless. I don't doubt that most accounting *software* parallelizes really badly. But that's an implementation detail. Commutative and associative operations like addition are easily made massively parallel. Google relies on it: http://labs.google.com/papers/sawzall.html
Actually the opposite is true. Posted Oct 21, 2005 12:34 UTC (Fri) by drag (subscriber, #31333) [Link] Makes sense.
So the questions that remains is:
1. Does it make sense to go to clustering for these applications when cheap muticore computers (4-8 cores) are getting cheaper, and 16+ cpu computers are more effective then they ever been before.
Mainframes I know are very good for certain types of data. Big 32+ cpu computers are better for others. Clusters are good for number crunching but are lousy for other things big iron and mainframes are good for, that's why IBM and sun still is able to sell million dollar boxes.
and if clustering for regular business makes sense, then..
2. Does it make more sense to rewrite all their applications to run on Linux or Windows? Since legacy windows apps won't work well on clusters, which platform is more likely to attract new developers/developments?
Peoplesoft/Oracle and the like are liking Linux (or so it seems to me) a lot more lately now that MS wants to compete with their own products. It's tough to fight a competing software vendor who owns the platform your software has to run on.
There still is a lot of noise about 'grid computing' and Linux is at the center of that. Whatever the hell 'grid computing' is suppose to mean.
But I won't dismiss the million billion dollars (just a rough estimate) MS is willing to throw at the problem.
Actually the opposite is true. Posted Oct 27, 2005 16:15 UTC (Thu) by jschrod (subscriber, #1646) [Link] The problem is not: "Does it makes sense?" but "Can it be financed?"
The last bank that I worked for had 40,000 mainframe programs alone for the share trading to end-customers (don't know the correct English term, equity management?). Rewriting these programs with the same functionality is just not worth the investment.
Since few specifications are available, rewrite them for new functionality is difficult as well. These programs literally evolved over decades and added all the special cases that must be handled. New software first have to re-discover all these special cases and re-program them anew. And believe me, it's no easy thing to analyse a twenty-year old COBOL program that has been actively changed all this time.
Oh yes, this bank tried that and failed. In the end, they had to write off 500 million Euros for the cancelled project. (And that was hard cash, not stock money.) It is a lesson for all those who think they can rewrite two decades of software assets on a whim.
Concerning compute clusters: These are not relevant for this realm. Most of the time, these programs don't compute, not even addition or multiplication. They just move data from one place to another, or update data. Performance advantages in database technology is the key to get them faster, not parallelization of applications.
Cheers, Joachim
Actually the opposite is true. Posted Oct 29, 2005 15:55 UTC (Sat) by rwmj (guest, #5474) [Link] You think for 500 bazillion euros they could have written anemulator with JIT optimizations. That would have allowed them to migrate the applications themselves slowly - or even not at all - but still have them run on newer, cheaper hardware.
I think the problem here is probably the banks themselves, who
Rich.
Ballmer: Microsoft to go after Linux strongholds (ZDNet) Posted Oct 27, 2005 7:48 UTC (Thu) by ekj (subscriber, #1524) [Link] Actually, a massive accumulation parallels trivially and effectively.In the non-parallell case you've got say one CPU adding up a billion numbers, one after another. This translates to a billion add-operations. With two CPUs you could let each add up half a billion items, and then at the end add the two results, so total steps is half a billion + 1 With say 1000 CPUs you could let each add up a million items, and then at the end add up all the results, theoretically that part goes fastest with a tree - that way there's only 10 sequential adds. In practice, due to latency and communication-issued simply adding them up at one node may be faster. So you end up with 1,001 million serial additions. That is really freaking close to the theoretical max speedup, 1,001 million is *fairly* close to 1 million. Now, I don't doubt that much accounting*software* scales badly, but that's a whole different ball of string.
|
|
||||||||||||