Hi, this might be more appropriate for lkml (or is there another place?)
because people with knowledge of how these layers interact might be here
and not on linux-raid-ml ?
If block cache is done on all levels and readahead is done on all levels,
then quite a lot of redundant block information is going to exist in
memory for all these layers? I can understand that it might make sense to
keep block cache for the fs and perhaps for the drive layer, but
md->dm(crypto)->lvm layers this might make less sense?
What about default readahead for these devices? Doing readahead on dm
device might be bad in some situations, perhaps good in others?
---------- Forwarded message ----------
Date: Thu, 29 Jul 2010 12:53:35 +0200 (CEST)
From: Mikael Abrahamsson <[email protected]>
To: Fabio Muzzi <[email protected]>
Cc: [email protected]
Subject: Re: MD raid and different elevators (disk i/o schedulers)
On Thu, 29 Jul 2010, Fabio Muzzi wrote:
> Is this true? Are there compatibility issues using different i/o schedulers
> with software raid?
I'd actually like to raise this one level further:
In the case of (drives)->md->dm(crypto)->lvm->fs, how do the schedulers,
readahead settings, blocksizes, barriers etc interact thru all these layers? Is
block caching done on all layers? Is readahead done on all layers?
--
Mikael Abrahamsson email: [email protected]