On Mon, 17 Jun 2002, Greg Oster wrote:

>Assuming your drives are sd0, sd1, and sd2, what happens when you do:
>
> csh
> foreach i (0 1 2)
> dd if=/dev/rsd{$i}c of=/dev/null bs=1m count=1000 &
> end

Damn that Linux background. I was using the block devices, rather than the 
raw devices. Using raw devices gets the speed back up again!

>> PS. As I've been writing this, I've just noticed that doing dd on a disk 
>> pushes lev3 interrupts through the roof! Before RAIDframe, I'd get ~200/s 
>> when doing dd, now I get more 1500/s! Would certainly explain the lack of 
>> performance. What gives?
>
>My guess is a very small stripe width... Try 64 or even 128.

Right, I've tried now 64 and 128, and even 256, but I'm still topping out
at ~4MB/s, much below the 5.5MB/s the disks can do by themselves.

Now I'm using the 'right' devices, the interrupts have come right down, so 
neither the SCSI bus nor the processor is overloaded.

Using bonnie to benchmark, I get interesting figures. The first is done on 
the raid set:
              -------Sequential Output-------- ---Sequential Input-- --Random--
              -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine    MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU  /sec %CPU
            0   732 88.2  4241 74.6  1957 56.3   637 89.1  3639 49.0  41.0 22.6


The second is done on a small partition on one of the raid disks:
              -------Sequential Output-------- ---Sequential Input-- --Random--
              -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine    MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU  /sec %CPU
            0   757 89.4  2938 44.3  1481 36.6   651 90.1  4750 57.7  43.3 13.7

Most notable from my point of view is the change between relative read and 
write speeds. The raid handles writes much faster than the non-raided, but 
suffers in the reads (as noted from the dd 'benchmark') Probably good for 
something like PostgreSQL, or maybe an LFS partition.

So, it appears that writing is the main benefit from using RAID-0, 
certainly on an IPX, though it uses much more CPU when writing, presumably 
due to the increased interrrupt usage.

I'll continue my play, and look at RAID-5 with the spare cheetah.

>
>Later...
>
>Greg Oster
>
>

-- 
    /"\
    \ /    ASCII RIBBON CAMPAIGN - AGAINST HTML MAIL 
     X                           - AGAINST MS ATTACHMENTS
    / \