[ILUG] Highpoint Sata RAID

[Rick Moen]

Quoting Paul Jakma (paul at clubi.ie):

> - The Promise TX4 is *not* the same card as the Promise SX4[1].

/me checks online.

SATA150 SX4 appears to have the _same_ design limitation as the TX4:  It
has a chip devoted to XOR calculations only.  Striping, error-handling,
etc., are offloaded.  (In a further display of cheapness, I notice that
they also use a Marvell 88i8030-TBC PATA/SATA converter chip on each
channel, like a lot of the other first-generation jobs.)

This is not necessarily a bad thing:  It makes the cards a lot cheaper
than, say, 3Ware's and Areca's.  But the system load problem is real.

> - If you're using Linux MD, as you should be, (be it with a promise
>   TX4 or SX4 or Highpoint or whatever) then you're
>   obviously not using proprietary on-disk RAID formats..

Which, you will note, is my recommendation for those determined to save
a few extra Euros.  (Addendum:  People looking over _new_-generation
SATA-II cards might want to look at the Tekrams:  They closely imitate
the extremely good but high-priced Areca cards using the same chipsets
and basic designs.)

> - On-disk RAID superblock format is *not* the cause of performance
>   problems with soft-raid.

And of course I _didn't say it did_.

> - The primary reason soft-raid can be slower is because it requires
>   extra bus bandwidth - instead of writing a block once and letting
>   the controller write a copy to each disk you have to send each
>   block across the bus for each disk (for RAID-1). For RAID-5 the XOR
>   doesnt cost much on a modern CPU (which are *vastly* superior in
>   speed and RAM bandwidth to the CPUs and RAM used on "embedded
>   computer on a PCI card" RAID controllers), however it is an extra
>   block of data to pass across the bus.

That is correct, with exception noted below.

>   Ie, extra bus bandwidth of software RAID is the primary bottleneck
>   of soft RAID

Except during restriping, when the calculation overhead becomes
absolutely brutal, and the system is often effectively nonfunctional.