hardware

[Tim Bowden]

On Tue, 2008-01-08 at 10:17 -0800, rich.fromm wrote:
> Ed McNierney wrote:
> > 
> > Third, I do agree that fast RAID 5 disks and lots of RAM are always a good
> > idea!
> > 
> 
> Yes, for working with large data sets, more memory is in most cases likely
> to be more important than a faster CPU.
> 
> But to speak out against sweeping generalizations, I wouldn't necessarily
> agree that RAID 5 is always a good idea.  There are pros and cons, and it
> depends on your circumstances.  If I remember right (and it's been a while
> since I studied the details), the biggest downside of RAID 5 is that small
> writes are slower, because of the need to write both the data disk and the
> parity disk.  With most RAID controllers I believe (there might be
> exceptions) there are also limitations concerning incremental upgradability
> of the array.  (e.g. what happens if I have N disks and I just want to add 1
> or 2 more)
> 
> - Rich
> 

My experience here (raid) is quite limited (and possibly old?), but from
my understanding it also depends quite significantly on the quality of
the raid hardware.  Low end raid "hardware" actually does many of it's
operations in software drivers, meaning you pay a cpu price for every
operation, whilst "decent" raid hardware does it all in hardware without
the cpu hit.  In the case of linux and low end raid hardware, the advice
I've consistently been given is to just use linux software only raid, as
it's more efficient than hardware raid that implements it's functions in
software drivers.  I've never benchmarked that, but the advice from
different sources has been consistent. I assume someone else has
benchmarked it (I know, dangerous assumption).  I have no idea how that
compares to windows (does windows have software only raid drivers?) but
FWIW that's my understanding of low end raid at least (OK, so I'm too
much of a cheapskate to ever pay for high end raid!).

As far as the windows linux with mapserver question is concerned, the
only serious benchmarking I've been involved with (getting old now) the
details are:
decent hardware (Dell boxes: dual zeons, 4gb ram, fast raid arrays),
windows server 2003 [1] v RHEL v4, mapserver v 4.6 & 4.8, data in sde
9.1 on a separate box. They both perform about the same under light
load, but if you keep ramping up the load (requests per second) the
linux setup took twice the load before response times start to blow out
than the windows setup will.

As noted elsewhere numerous times, test results can be quite sensitive
to a number of variables, so I'd not put too much weight on any tests
except those using *your* hardware/software and *your* data.  The value
of other peoples tests is probably best realised by using them as guides
to what you should be looking for in *your* testing.

[1] In previous reports (long ago) I've quoted windows as being XP pro,
but that was incorrect.

Regards,
Tim Bowden