Tom,
In my opinion, the real problem w/NT is that it is not optimized
to work on a specific set of hardware. Those in marketing think
that is just great, but we're talking about OS software, here,
not applications.
NT scales poorly because of their locking scheme. In addition,
their thread libraries are not the exclusive property of user
space, so new pid's have to be allocated whenever a new thread
is opened. That requires a software trap into kernel space,
which is relatively expensive.
Check out Solaris if you want to know how locking should be
done in mp systems. One thing to remember: when locking data
you have to be absolutely certain that two separate threads do
not write to the same data at the same time. In computing
environments, latencies that are measured in micro-seconds are
too indeterminate. Most locking, that is really effective, like
mutex, relies on the hardware, which includes atomic instructions
that won't allow data corruption on a write request. MSFT can't
do that, because they don't write NT for any specific machine.
However, NT is written to run mostly on PC's. PC's aren't
powerful enough for enterprise systems.
The point is, the only companies that have any business writing
operating systems software for enterprise systems are those who
manufacture the hardware to go with it. The assemblers, compilers,
and OS's can be designed specifically to work with the hardware
it runs on, and nothing else. Hardware design and implementation
is the real key to optimal performance, not software, and as the
hardware gets more and more sophisticated, you need software
engineers to optimize your OS to take full advantage of it. Small
differences in kernel design can make huge price/performance
diferrences.
MSFT doesn't really have the engineering talent to bring it off.
That's why NT 5.0 completion dates have slipped again. They have
a significant learning curve up in Redmond and it will take some
time before they come up to speed.
cheers,
cherylw |
