readahead (blockdev --setra) with sw raid(5)?

readahead (blockdev --setra) with sw raid(5)?

[Stephan van Hienen]

Hi,

just saw a 3ware document with some hints for hardware raid5
they advice to use blockdev --setra 16384

since i'm not using 3ware hardware raid
(i'm using 2*7500 3ware's with sw raid5 over 14 disks)
i'm wondering if --setra can still improve my raid performance

any hints about this ?
or some other hints, improvements on filesystem ?
i'm currently using blocksize 128, maybe switching to 256 helps ? (with 
large files)

Re: readahead (blockdev --setra) with sw raid(5)?

[David Greaves]

I routinely use --setra 4096 on the underlying devices, the md array 
devices and the lvm devices
I am generally reading 1-2 gigabyte files.

I use XFS and it is often suggested as a good fs (though I think there 
may be some XFS/nfs issues with 2.6 at the moment - see lkml)

David

Stephan van Hienen wrote:

> Hi,
>
> just saw a 3ware document with some hints for hardware raid5
> they advice to use blockdev --setra 16384
>
> since i'm not using 3ware hardware raid
> (i'm using 2*7500 3ware's with sw raid5 over 14 disks)
> i'm wondering if --setra can still improve my raid performance
>
> any hints about this ?
> or some other hints, improvements on filesystem ?
> i'm currently using blocksize 128, maybe switching to 256 helps ? 
> (with large files)
> -
> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>

Re: readahead (blockdev --setra) with sw raid(5)?

[Stephan van Hienen]

On Sat, 27 Nov 2004, David Greaves wrote:

> I routinely use --setra 4096 on the underlying devices, the md array devices 
> and the lvm devices
> I am generally reading 1-2 gigabyte files.
>
> I use XFS and it is often suggested as a good fs (though I think there may be 
> some XFS/nfs issues with 2.6 at the moment - see lkml)
>

ok i'll try 4096 then on my devices
(why did you choose 4096 and not 2048 (or 8192 (or bigger?)

also i think i'll create my new raid5 with xfs
are there any special tuning options i can use with xfs ? (like stride 
with ext3)

i think i'll go with 256 stripe size (not sure if i want to go 512)

Re: readahead (blockdev --setra) with sw raid(5)?

[Neil Brown]

On Saturday November 27, raid@a2000.nu wrote:
> Hi,
> 
> just saw a 3ware document with some hints for hardware raid5
> they advice to use blockdev --setra 16384
> 
> since i'm not using 3ware hardware raid
> (i'm using 2*7500 3ware's with sw raid5 over 14 disks)
> i'm wondering if --setra can still improve my raid performance

--setra large can improve sequential read performance for any device,
possibly at the cost of reducing random-io bandwidth or using up
memory that other applications might want (I'm not sure what the exact
trade-off is).

md sets the read-ahead number for raid0/4/5/6 to be twice the stripe
size. i.e. chunksize * number of data disks * 2.

> 
> any hints about this ?
> or some other hints, improvements on filesystem ?
> i'm currently using blocksize 128, maybe switching to 256 helps ? (with 
> large files)

It might help large sequential reads, but might hurt smaller reads.

Some testing I did a couple of years ago (and so is probably
out-of-date) suggested that after about 128K, the returns from
doubling chunk size diminished very quickly.


This sort of tuning is very dependant on work load.  If you have just
one application that does lots of very large sequential reads, then
large chunk size and large 'ra' are good.  Otherwise ... it is hard to
know.

NeilBrown

> -
> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html