Re: [linux-lvm] Migrating LVM

[Phillip Susi <psusi cfl rr com>]

On 4/21/2010 11:36 AM, malahal us ibm com wrote:
> Interesting! You are doing direct I/O to avoid copying from cache to user
> buffer for read and vice-versa for write, but you are losing the ability
> to do them parallel! You are doing the next best, that is creating two
> "dd" threads -- one for reading and another for writing. Since the pipe
> is really implemented in memory, why should this be faster than normal
> "dd" that uses page cache? Likely that kswapd is not kicking early
> enough?

Oops, my eyes missed the pipe and second dd when I made my previous
comments.  That is pretty good for different disks then yes... not so
good for same physical disk.

> Enhancing "dd" to create a reader and a writer thread would really
> help, I believe.

I actually have some old hacked up dd code I made once to use 16
concurrent aio requests with O_DIRECT.  I need to clean it up a bit but
it showed great promise.