Small update:

Being the thoroughly stubborn SOB I am , I decided to dig a little more on host-based RAID5 performance (a.k.a. Fake RAID). While W2K/2K3 performance is nowhere to be found, my guess is, since it will be using the same resources as the FakeRAID controllers, performance should be just about the same, perhaps sometimes even a litlle bit better (since there is one less driver layer to work with...).

That being said, I found out that a single-core Opteron 148 (here) managed to go as high as ~125MBps sequential, ~10MBps random read/write on a RocketRAID 2200 card (here, not bad considering the same controller only managed 18MBps random read/write on a RAID-0 array. My guess is I should expect something around those ballpark figures in W2K3 and similar configurations, right?

I also found out another review on the RR2300, with interesting results (here for relevant results), and yet another one (here), this time with the ICH7R in the mix, which should be more or less what I'll be able to get from the ICH7+software RAID (right?).

The odd thing is, unless these controller cards/ICHs have something that's helping the actual reading and writing of data to and from the disks (not talking about the XOR engine, which all lack), the CPU usage is incredibly low... On the last review I linked to, maximum CPU usage was below 3%... on READS! Writes were even lower....

Any coments? That's weird, I always though the CPU would be fighting with the XOR operations... Now it seems the bottleneck is somewhere else... I mean, if top-of-the-line single-core ~500MHz dedicated XOR engines are capable of 800+MBps, then if most of a general-purpose dual-core 2GHz+ CPU were to be used for that, one should be talking about at least 500+MBps throughput... Hell, if ~35MBps random writes only takes ~3% of a 955EE CPU, that CPU alone should be able to output at least 600MBps random writes, and still have room to do everything else the system needs...

I'm oficially ...

Cheers.

Miguel