Re: Install and RAID
Quoting Adam McKenna (adam@debian.org):
> On Sun, Jan 28, 2001 at 09:42:30AM +1100, Glenn McGrath wrote:
> > Michael Neuffer wrote:
> > > Why do you think IBM is still successful with their Mainframes ?
> > > They do not have that much CPU power, but they can process an
> > > enormous amount of IO. How do you think they accomplish this ?
> > > By offloading the IO to IO processors.
> > >
> > So you think mainframes are a good long term growth industry ?
>
> No, but mainframe-like mega-IO systems such as EMC and NetApp are.
I'm not so sure they will die out in the near future, they have far
too many applications. Just look at the deal IBM just made consolidating
over 70 SUN servers into one mainframe. Most of what those machines were
doing was running web servers and guess what ? A perfect application
for a mainframe, especially if it is running lots of Linuxes. You just
need enormous amounts of IO. :-)
There is simply nothing that can compete with a mainframe on that field.
> > > The hardware RAID controller, at least one that is worthy of this name
> > > is just that (only smaller). An intelligent IO subsystem that offloads
> > > complexity from the main system and lets it do what it is supposed to.
> > >
> > > For me somebody that argues like you, is also somebody who thinks
> > > that Winmodems are the best thing since sliced bread.
> > >
> > hmm, nice analogy (but no i dont like hardware i cant use), i think what
> > it comes down to is computing power, the best value for money components
> > will always be those that are have the highest volumes.
> > There is a sweet spot, high powered stuff has a higher prcie/performance
> > ratio, but so does lower power stuff, hard drives especially (where
> > performance is size).
>
> No, size is only part of performance.
I would even venture to say that it is a small part. Many of my colleagues
would love to have their good old 2GB drives back instead of the 47GB
drives that we are using in the EMCs.
> > Maybe it would help you if you considered a server built with standard
> > consumer components to be one fancy hardware raid device.
>
> Maybe it would help if you considered the fact that when designing a storage
> subsystem or SAN, most people are thinking about the data that their business
> depends on, and not their MP3's.
Exactly. And in many such cases you are beeing monitored from customers
or some authority. Often it is like "Three strikes (within a certain period)
and you're out", which means you will have to close all or parts of your
company. This is the case for banks for example since nobody wants another
Berings (sp?) Bank desaster to happen.
> > Ever heard of google ? If mainframes and "special" hardware are so good,
> > given their high IO demands they should dump the thousands of "CHEAP"
> > computers that can successfully handle high IO and get a mainframe from
> > IBM.
>
> I'm sorry, I'm afraid I don't see how this fits into this discussion. You
> seem to imply that google is using distributed computing for its processing,
> but we are talking about storage here. I highly doubt that google is storing
> all of their data using Linux kernel RAID.
Actually that is indeed possible. They are farming out the requests
to about 4k machines IIRC. It is very possible that they have also
segmented the database and keep one segment on each machine. That way you
only have partial losses of your data when a machine or its drives goes
down. In their business model it does not cause any problems when they
lose parts of their data. The next index sweep of the net will bring it
back and the data outdates very quickly anyway.
It is unlike a banking environment or something similar where a dead
machine or some lost transactions can mean billions lost in seconds.
Cheers
Mike
Reply to: