[Nottingham] Big biiig problem regarding IDE-RAID

[Robert Davies]

On Tuesday 15 Jul 2003 15:56, Graeme Fowler wrote:
> On Tue, 15 Jul 2003, Robert Davies wrote:

> Notably, it's not reporting scads of activity through vmstat or top, so it
> appears that the driver is simply saying "invoke rebuild from drive 0 to
> drive 1" to the chipset, and not troubling itself with the data. Which is

The software RAID-1 mirroring is about the same, the transfers are DMA so the 
CPU just has to set them up, not copy data between buffers.  If the hardware 
controller is doing a lot of the work and not bothering the CPU, the 
interrupt rate should remain low when the sync occurs, I'd be pleasantly 
surprised if it does (base rate is 100 on i386).  IIRC CPU% was only a few % 
during those sync ups :)

> The system itself, for those who are interested, is an Intel S845WD1-E mobo
> with 512MB ECC DDR RAM (expandable to 2GB if you can afford 2*1GB DDR
> DIMMs) with 2.4GHz P4, dual onboard EtherExpress Pro 100+ NICs, onboard
> 64MB Rage graphics and onboard Promise IDE RAID.
>
> All in all, it's a very attractive proposition. For a specific job :)

Intel costs more.  Incidentally the MSI board lets you go up to 4GB with 1GB 
DIMMs.  The 2.6 kernel will improve the error reporting situation, it was one 
of the things worked on at 2.5 kernel summits, as part of the 'Enterprise' 
enhancements.  I've heard similar stories on software RAID-1 where drive 
power is deliberately failed, one reason for mentioning the 2 box solutions 
using HA project stuff and the preference for IDE enclosures with SCSI 
interfaces.

Another thing I remembered, really these RAID controllers tend to have fair 
amount of RAM cache, and unfortunately only the very expensive ones battery 
back it.  For ultra safety, using a 'Memory Technology' NVRAM device for the 
file system logs looks like a promising solution.

Rob