Re: [Aoetools-discuss] AoE much slower than iSCSI?

[Mark F. <ma...@fa...>]

On Wed, Aug 16, 2006 at 10:57:01AM -0400, Ed L. Cashin wrote:
> On Wed, Aug 16, 2006 at 03:52:00PM +0200, Tomasz Chmielewski wrote:
> > using userspace vblade will not explain 3x difference in write 
> > speeds (or would it)?
> 
> Yes, it would.  The vblade was designed to work but it wasn't ever
> intended to be used as a performance benchmark for AoE.
Indeed.

In my honest opinion, the vblade software as an excellent way to prototype
some AoE setups, but not much more. Performance as you point out wasn't a
design consideration and I guess neither was correctness with respect to a
running file system. We can solve the second one using direct I/O but it'll
slow things down further.

> There is already one in-kernel vblade that performs better than the
> userland vblade, and I expect that other, better performing vblades
> will be available in the coming months.
So, I have a theory that a userspace vblade doing direct asynchronous I/O
has a shot at performing very well, perhaps just as well as a kernel
solution. A userspace blade has the advtange of being easier to manage,
compile, package and change. And of course, when it crashes, it doesn't take
down your whole server :)

I've started prototyping an async I/O driven vblade in my spare time at
home. Assuming it actually goes anywhere - there's a fair chance it won't
:), I was wondering how receptive to such vblade patches you guys would be.
It adds a fair bit of complexity to the vblade source, so I'm not sure it
syncs up with all your design goals there.
	--Mark

-- 
"There's nothing like a cantankerous old man who takes a 'hey you kids, get
off my lawn!' approach to foreign policy" -Jon Stewart on the subject of
Donald Rumsfeld