A modern RAID solution?

A modern RAID solution?

[Alexy Khrabrov]

Now that I got all those drives spinning, I'm eager to try out 
a RAID array.  I reckon it's a better way to hedge against disk
failures than backups on tape -- even 20/40 GB dds4 or DLT is
not enough those days, and loading/rotating is tedious.

I heard an opinion that software RAID on Linux with SCSI is
"almost" as good as a hardware controller.  What is the experience
here?  Also, since I'm running LVM throughout, including the root
partition, does RAID coexist well with LVM?

But if I go the way of the hardware controller, is it better
to get a separate one, or one of those new cards from Adaptec
which say they have a Host RAID 0/1 or some such?

I'm looking at a 3 drives RAID to begin with, perhaps 4.

-- 
Cheers,
Alexy Khrabrov :: www.setup.org :: Age Quod Agis

RE: A modern RAID solution?

[Eff Norwood]

> I heard an opinion that software RAID on Linux with SCSI is
> "almost" as good as a hardware controller.  What is the experience
> here?  Also, since I'm running LVM throughout, including the root
> partition, does RAID coexist well with LVM?

In my experience, software RAID is very flexible and good at the expense of
a lot of interrupts and some (5-15%) processor overhead depending on the
RAID level you choose. Hardware RAID is good at the expense of $ and doesn't
use as much processor overhead as software RAID. Ultimately, I chose
software RAID because it was vastly more flexible for me. I also abandoned
LVM and now use an IBM open source RAID/Volume manager called EVMS.
http://evms.sourceforge.net. The other thing about software RAID is that if
you move your drives to a new system, they'd still likely work. If you need
to move a hardware RAID card to a new system and it won't work in that
system, you might be out of luck.

Eff Norwood

Re: A modern RAID solution?

[Steven Dake]

Alexy,

As a previous RAID card engineer, I can definately say if your system is 
fast you won't see much improvement from an I/O card based RAID adaptor 
for either RAID 0 or RAID 1.  Keep in mind that the OS must still send 
messages to each adaptor over the PCI bus so the data transferred is 
nearly the same and with SCSI command qeueing each disk can accept 
multiple commands at once.  RAID 5 will see improvement because the I/O 
processors have special hardware to do 8 or 16 parallel XOR 
calculations.  Also generally I/O processors on the RAID cards run at 
much slower speeds then the 2 ghz standard PC these days.  One advantage 
of a hardware raid card is that is can accept more commands in an I/O 
queue (reducing read/write response time but improving performance). 
 There aren't alot of applications where this is useful unless you do 
alot of writing and not much reading or the raid card implements a 
special i/o channel for reads (which none do that I am aware of).

The adaptec "host raid 0/1" means that the host machine (your pc) is 
responsible for the striping or mirroring.  The adaptec adaptor just 
provides "bios mapping" to the disk layout specified by the design so 
the device can be booted.  This is equivalent to the MD driver except 
the MD driver is probably more optimized then Adaptec's driver if they 
have one for linux.  This is also how the "IDE RAID" cards work just 
providing BIOS mapping so int13s i/os match the layout of the disks.

Alexy Khrabrov wrote:

>Now that I got all those drives spinning, I'm eager to try out 
>a RAID array.  I reckon it's a better way to hedge against disk
>failures than backups on tape -- even 20/40 GB dds4 or DLT is
>not enough those days, and loading/rotating is tedious.
>
>I heard an opinion that software RAID on Linux with SCSI is
>"almost" as good as a hardware controller.  What is the experience
>here?  Also, since I'm running LVM throughout, including the root
>partition, does RAID coexist well with LVM?
>
>But if I go the way of the hardware controller, is it better
>to get a separate one, or one of those new cards from Adaptec
>which say they have a Host RAID 0/1 or some such?
>
>I'm looking at a 3 drives RAID to begin with, perhaps 4.
>
>  
>

Re: A modern RAID solution?

[Luben Tuikov]

Steven Dake wrote:
> 
> multiple commands at once.  RAID 5 will see improvement because the I/O
> processors have special hardware to do 8 or 16 parallel XOR
> calculations.  Also generally I/O processors on the RAID cards run at

I've long been thinking for a software RAID 5 which uses
the REGENERATE/XD*/XP* SCSI Block device commands.

For this reason I thought it useful for an application
client to know if the command made it to the medium as
I posted sometime ago (e.g. when SCSI core requests
through eh_abort_handler()).

-- 
Luben


P.S. (I.e. with the current queuing models for some transports
(and some of my drivers) it is possible to say that:
the cancelled command made it / didn't make it to the
medum (trivial) or that it is unknown whether it made
it to the medium (only in this case we have to re-request
to check).)