On Wed, Mar 19, 2003 at 11:57:42AM -0800, John Alvord wrote: > On Wed, 19 Mar 2003 19:34:26 +0800, John Summerfield > <[EMAIL PROTECTED]> wrote: > > >On Tue, 18 Mar 2003, Joseph Temple wrote: > > > >> I would point out that clustering makes hardware more available, not more > >> reliable. > > > >If the application stays up, ir's more reliable. > > > >> The things actually fail more often because there is more to > >> fail, > > > ..... > >We've discussed Google here before: would anyone notice if a few Google > >servers went missing for a while? Seems to me, probably not, and > >according to some in the discussion that is on low-cost hardware. > > You always have to take the application into account. A Google session > can drop out with little effect. A money transfer of a million dollars > is quite another story.
Google initially needed its big farm for number-crunching, not for for the interface to the queries: They have a big storage of pages, each with its own "google rank". This google rank needs to be re-calculated occasionally (daily?). This is a task that can be parallelised, and thus can be calculated by a multi-node cluster with relatively little overhead compared to one big iron. (Of course google claims that the real story is different, http://www.google.com/technology/pigeonrank.html :-)) Not all tasks are such. But anyway, using clustering for parallelising tasks is quite different from using clusters for high availability. -- Tzafrir Cohen +---------------------------+ http://www.technion.ac.il/~tzafrir/ |vim is a mutt's best friend| mailto:[EMAIL PROTECTED] +---------------------------+
