Re: [TCLUG:5285] NTS4 vs. Linux study from Mindcraft.

[Randy Maas <randym@acm.org>]

At 08:57 AM 4/14/99 -0500, you wrote:
>I find it VERY difficult to believe 
>that NT is faster than Netware just because Netware has so little 
>overhead.

I kind of have to pipe up here.  Why?  I wrote and maintain my employers
Netware driver, in addition to our linux driver.  I work with the guy who
wrote our companies NT driver.  I also work with our characterization lab.

Short summary: NW blows.  big time.  Stick with Linux if you can.

NW can be configured faster than NT for random and some sequential file IO
test for similar hardware sets.  NT is faster than NW for every
configuration I have seen until the machines have 1GB of ram.  Then they
perform about the same: NW is marginally faster.  But not much.   For a
Xeon 4 CPU (333MHz), with 500GB of memory, it is about half the performance
of NT.

There are some cornerstone NW problems that are laughable:

1. NW will slow intensely down for "large" disks.  I made a RAID 10+
logical disk sized to about 70GB.  Then used a standard novell program to
fill it (dts).  It took about 36 hours to write the data out: starts off at
an okay speed, then nose dives.  But it took the next six days to verify
the data.  Dts is not written optimally (pure asynchronous IO), but it is
written in the same methods as nearly every other program out there.

2. Another basic test: create a few thousand files, then delete them.
Using netbasic or nwshell (on netware) with a "del *.*" takes all
afternoon.  "dir" takes several hours.  It has to do with the way the
filesystem is designed: just getting a descriptor from a filename is very
very slow in netware.  Open/close tests show similar results (no IO involved).

3. IO Queues.  NW is very proud of the buffering, which seems to work.  The
problem is that it bottlenecks all of the IO.  NW spends more time trying
to use its caches than it would cost to actually read IO from the disk.
And for any disk, NW only allows a maximum 16 outstanding IOs at a time. (A
single IO is basically a scatter gather list).  Modern controllers, and
disks can handle orders of magnitude more -- which would allow (if used)
higher performance.

4. More cpus shmore cpus.  NW for some, strange reason, likes to force
everything to one CPU.  This wouldn't be bad if you had one CPU, but it
reduces performance when there are three others sitting idle.

5. API bypass.  NT, like UNIX, has a mechanism to allow programs to
directly use the harddisk: no filesystem layer.  This is typically used to
by databases to allow higher performance; they also employ lots of
asynchronous IO.  NW officially claims to have one; it is undocumented.
And Oracle doesn't really use it.  Oracle performance on NW is far worse
than NT, in terms of transaction processing.


Yippee.  I could say lots of bad things about NT -- and have. But NW is not
too great.

Randy.