Message-ID: <37435F62.80A7941D@efortress.com>
Date: Wed, 19 May 1999 18:03:30 -0700
From: papageorgio <papageorgio@efortress.com>
MIME-Version: 1.0
To: new-httpd@apache.org
Subject: Re: Apache port to PVM
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Sender: new-httpd-owner@apache.org
Precedence: bulk
Reply-To: new-httpd@apache.org

----- Original Message -----
From: Paul J. Reder <rederpj@raleigh.ibm.com>
To: <new-httpd@apache.org>
Sent: Friday, May 14, 1999 7:30 PM
Subject: Re: Apache port to PVM(BeoWulf Clusters)?


> It is my understanding that PVM provides an underlying communications
> API that allows for "seamless" distribution of manually architected
> components. This provides a portable communications/distribution layer,
> but does not address other portability issues such as memory, file I/O,
> and mutex support.

I agree, for the most part that is left to the OS to resolve, but they
just
released a new PVM with a whole new set of features that could address
this,
but we have not had a chance to really look into it.

> There are many applications that benefit (or would benefit) from PVM.
> Other than getting load balancing for free what benefits do you see for
> Apache? Keep in mind that there is currently work being done to port
> Apache to use a portable runtime API set that addresses all of the
> portability issues, and that (as you pointed out) Apache can already
> be clustered.

If what your saying is that Apache is not the best application to port
to
the PVM api I agree. There are a couple of architectural issues that
would
need to be resolved. The most important is PVM is designed with a set
end
point. Data is broken up in chunks at the beginning of the application
and then passed out as needed, since there is a predetermined amount of
data, checking for slower process and redistributing load can be done at
a
task count level. Apache has no set endpoint nor does it have a fixed
data
amount.

If Apache was ported to PVM and the technical thingies were worked out.
The
cluster would act as one big server, tasks could be switched in
midstream to
even loads, there is an advanced messaging system in PVM and finally PVM
has
proven to be stable and scalable.

Running on top of a Linux box it could compare nicely to a single box
with
multiple cpus running anything you want to compare it to.

Linux is not the only OS that PVM runs on but Linux does have some very
serious benefits, because in essence PVM was designed to run on Linux
and in
a very real way the Beowulf project helped design Linux to run PVM.

Hmmmm.... that sort of looks like a Linux Rant, sorry.

> Those questions aside, I would certainly be curious about a port from
> a purely theoretical standpoint and might even be interested in playing
> with it some (but after my current thesis research is done). I am
> personally very interested in the Beowulf architecture.
>

Right now it is purely theoretical. All we have done is read Apache
source
and compared it to the now outdated PVM api. Many things will fit nicely
but some will be Ugly.

papageorgio