Re: MIG->Corba (performance)

[OKUJI Yoshinori]

From: Marcus Brinkmann <Marcus.Brinkmann@ruhr-uni-bochum.de>
Subject: Re: MIG->Corba (performance)
Date: Fri, 9 Feb 2001 16:11:19 +0100

> Of course, but any message passing only adds a constant overhead to a single
> message. Performance increase can also be achieved by adding new interfaces,
> which remove the need to send several messages (by replacing them with a
> single message). I think this is what is planned for fork. Faster computers
> make the constant overhead for single messages negligible.

But the fact is that the cost of message packing is not
negligible. Some people (in particular parallel computer scientists)
assume that constants are not important to mesure speed performance,
but the real world is not so.

As a simple example, suppose that the cost of one IPC in A system is
three times higher than that in B system. Then, A system must reduce
the number of required IPC calls at least to 1/3, to achieve the
performance equivalent to B system. This is quite hard or impossible
in many cases.

> Also, AFAICS, there is no reason why any message interface generator should
> be significantly slower or faster than any other.

I'd recommend reading some papers about IDL performance, such as
"Stub-Code Performance is Becoming Important" which is avaiable from
<http://l4ka.org/projects/idl4/>.

Okuji