Linux vs. Solaris file IO performance

[Mark Komarinski]

On Mon, Oct 28, 2002 at 02:40:00PM -0500, Derek Martin wrote:
> At some point hitherto, Tom Varga hath spake thusly:
> >     I made the comparisons on an 8-way Sun-Fire-880 with 900Mhz processors
> > against my 700 Mhz IBM a20 linux laptop.  I'm fairly sure that the differences
> > that I'm seeing have nothing to do with hardware.
> 
> Despite the specs you quote, hardware could still be a factor.  What
> kind of disk(s)/controllers does the SunFire have, and how are you
> using them?  Is it in a RAID configuration, or no?  Certain RAID
> configurations can be slower than a single disk, in my experience.
 
Sun is standardizing on FC drives (yep, Fibre Channel).  And
you thought SCSI was expensive....

> You failed to mention anything about memory usage, which probably has
> the most impact here.  If there's a lot of free memory for caching,
> you'll see the kind of performance you're describing with the Linux
> machine.  If there's not, you'll see slower behavior like the Solaris
> machine.  As there won't be enough RAM to cache the I/O, all the I/O
> will be forced to go directly to disk.  Note that I said "free memory"
> above, not just memory.  Your server could have 10x as much memory as
> your laptop, but still have none free...
> 
> Of course, there could still be other factors, some of which could
> be hardware related, and some not.  And of course, there could be a
> combination of factors as well...

I think the real issue is that you're not looking at loaded systems.  One
user on a PC will always smoke (performance-wise) one user on a SPARC-based
box.

But what happens when you really start to load the system with users
and processes?  The PC will start to choke and sputter, while the SPARC
is probably plugging along at the same rate it was before.  Hence
the reason why you build clusters of PCs instead of buying a 32-way
PC (aside form cost).

-Mark