[Mpi-forum] Good luck with MPI 3.0
treumann at us.ibm.com
Mon Dec 6 08:55:42 CST 2010
To MPI Forum members
I have decided to retire from IBM at the end of 2010 so I will no longer
be involved in the Forum.
I have been involved with MPI and with the MPI Forum since the days when
we were debating whether an MPI 2.0 effort was really needed. I attended
almost every MPI 2.0 meeting as IBM's representative from development,
working with Marc Snir and people from IBM Research that he brought in. In
the MPI 2.1 and 2.2 effort, I have been part of a broader group of IBM
representatives and only attended a few meetings.
I am proud of what the Forum accomplished over almost 2 decades and of my
small part in that. I have met a lot of amazingly creative people. I will
miss looking at the ideas people bring forward, thinking through the
implications and helping to refine them to enough clarity to be either
adopted or tabled.
I wish all of you well as you try to work out how the MPI model should
adapt to scales of parallelism that were barely imaginable in the early
Dealing with a million or more hardware threads without requiring every
thread to maintain local state about all million of them is really hard.
With such huge numbers of hardware threads, the failure of a thread
(processor, node, link) in the lifetime of an application run becomes
common and looking for ways to pick up and keep going becomes
irresistible. The trouble is, it is realy, really hard and is at odds
with the original MPI philosophy which built on the idea that applications
should be designed around algorithms with the freedom to assume
reliability. (If your algorithm was good, you could assume hardware
failures were too rare to matter)
Good wishes to MPI as a standard and to all of you personally
Dick Treumann - MPI Team
IBM Systems & Technology Group
Dept X2ZA / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
Tele (845) 433-7846 Fax (845) 433-8363
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the mpi-forum