Re: Larry Ellison comments on Microsoft's benchmark

"Darin McBride" <dmcbride_at_nospam.tower.to.org> wrote:
> >
> >Thank you for the info and the link to IBM documentation. From
> >what I understand, MSCS clusters are limited to 2 servers per
> >cluster. So if you have 12 machines you have 6 independent
> >clusters. Isn't this still 6 times less reliable than Oracle
> >Parallel Server?
>
> Let's use some "real" fake numbers. Let's pretend that each machine has a
> failure rate of 10%. This is awfully high, but it's a nice, round number.
> ;-)
>
> One machine: 10% failure rate.
>
> Twelve machines: 100% - ((100% - 10%)^12) = 71.75% failure rate.
>
> 1 2-way mutual-failure takeover cluster: 10% ^ 2 = 1% failure rate.
>
> 6 2-way mutual-failure takeover clusters: 100% - ((100% - 1%) ^ 6) = 5.85%
> failure rate.
>
> 12-way mutual-failure takeover cluster: 10% ^ 12 = 10^(-10)% failure rate.
>
> "6 times" is a misnomer. But even with a huge overstatement of the
 failure
> rates, it can become reasonable fairly quickly.

So in your example it is 0.0000000001% failure rate for OPS vs 5.85% failure rate for MSCS clusters. So OPS is almost infinitely more reliable than MSCS based solutions. (I realized 6 times was wrong immediately after I posted it. I should have said approx 6 times less reliable than a 2-machine cluster. In a previous post I had said that OPS is infinitely more reliable.)

MSCS-based solutions have other problems too. For example, if a machine fails, the load previously carried by that machine will have to be taken over by another already-heavily-loaded machine. Depending on how loaded the surviving machine is, this take over may overhelm the surviving machine and may cause a second failure, which would be fatal for the system as a whole. In Oracle Parallel Server on the other hand, the extra load is balanced by *all* remaining machines instead of putting it all on one machine and thereby killing that machine. Received on Fri Jul 07 2000 - 00:00:00 CDT