[H-GEN] Linux file server

[Russell Stuart]

On Fri, 2 Jul 2004 Edwards_Ewan_B at cat.com wrote:
> On Friday 02 July 2004 17:41, Russell Stuart wrote:
> I followed some of that discussion on the Linux Poweredge mail 
> list, and one point that I noted was to do with the application 
> load on the CPUs.  As I remember, if there were plenty of CPU 
> cycles available then those benchmarks certainly held up.  
> However, if the CPUs were very busy, then hardware RAID was 
> faster.

Yes.  That is why in my original post I said:
  "Linux software Raid is faster than hardware Raid,
  providing you have the spare CPU cycles."

Still, it is nice to see someone remembers the discussion
the same way I do :).

Looking back on it, CPU speeds have come close to
doubling since that discussion.  But you do need a beefy
motherboard.  David was talking about achieving 200MB/s
on his system.  It is an awesome figure.  PCI can only
manage 133 MB/s by comparision.  PCI-X manages 4 times
that, from memory.  So on a normal motherboard, you
aren't going to be limited by the CPU or the Raid card
- whichever you choose to use, as you will saturate the
PCI bus first.  But if you don't talk to your HDD's via
a card (ie you only use the raw ATA/SATA connectors
provided by your motherboard) you won't hit that
limitation.  This cuts out hardware Raid (at least until
it becomes common chip set feature), so again software
raid has the potential to be faster.  I don't know
whether it meets that potential or not, but in any
case its irrelevant for a file server that isn't
using gigabit ethernet - 100M bits/s ethernet can only
serve files at around 10M bytes/s or so.

Which take me around the full circle, I guess, because
this means my original point about software raid being
faster than hardware raid is irrelevant to question
posed in the original post.