Re: alternate simplified proposal #1

Richard Frost (frost@sdsc.edu)
Tue, 11 Jun 1996 14:57:20 -0700 (PDT)

> I agree. However, can we be more specific? What kind of remote I/O
> servers do we want to support?

Scenario: user has accounts at a local and remote site and wants
local program to read data from remote server (suppose the sites
are Los Alamos and Pittsburg SC). The user needs a mechanism to
read the remote data from the archival storage server at PSC.

USENIX is developing "ticketed" access protocols. This will allow
users to manage their own security and still meet the stiff security
constraints of local sites.

So as a specific example, I would put my PGP or Kerberos ticket
along with my user name, the site name, and the server name into the
MPI_INFO argument -- then call MPI_CONNECT.

> What protocols do these servers use?

Various, and still emerging. For example: Unitree, HPSS. There are
custom daemons at some sites.

> Can info args support these protocols?

Yes, several schemas were discussed last week. I would like to
defer this until Bill Saphir posts his formal proposal.

Further, it is a quality of implementation issue whether a vendor's
MPI can support a "connection" to a remote XXX file server and treat
the I/O stream(s) as just another "file handle".

I think that the NASA Ames implementation of PMPIO is a specific
example of how to do this. Watch Parkson disagree ... :-)

====

Thanks for the interesting feedback. John May has also raised some
interesting points that I hope to respond to later today. At the
moment I'm enroute to New York.

Cheers,
Richard

P.S. Seek(fh=0) is just rewind.