ARSC T3D Users' Newsletter 71, January 26, 1996
Mike Ess's Remarks at the PSC Optimization of Codes Meeting
I don't often use this newsletter to explicitly editorialize, but I was asked to say a few words at the opening of the PSC meeting. Below is what I said:
Address to PSC Meeting on the Optimization of Codes for the CRAY MPP SystemsI am happy to say a few words on the importance of this meeting on code optimization.
As the youngest of the members of Pharoh, ARSC does not have a long history with CRI. In fact we received our Y-MP less than two and half years ago and our T3D has been on site only two years.
At ARSC I have had a chance to observe many attempts to port and optimize codes on to the T3D. It is still surprising to me that users attempt this without an understanding of vectorization. But because we are such a young site we do not have the long history of vectorization that an established CRI site would have.
But lack of a history in vectorization is a refreshing experience because these new users do not have the attachment to DO loops like some of us gray haired guys. And from our parallelization experience, we know that parallelization works best on a level higher than that of a DO loop.
In observing the many users and would-be users of ARSC's T3D I have tried to categorize them based on the following triangle:
/\ / \ I / \ E n / \ x f / \ p o / \ e r / \ r m / \ i a / \ e t / \ n i / \ c o / \ e n / \ ---------------------------- CommitmentBy evaluating a user's strength in each of these areas, I believe I can predict whether their project on the T3D will be a success or not. I evaluate a user in each of these areas by asking these questions:
- Experience: Have they run on other MPPs before?
- Commitment: How much time are they willing to invest in getting their code running?
- Information: Do they have access to specific T3D information?
ExperienceEach user is different and brings different skills and experience to the T3D. If a user has a working code on another parallel machine then porting it to the T3D is a short task. Beyond the syntax differences of different methods, the semantics of message passing are preserved on all MPP implementations. And CRI has one of the best implementations.
CommitmentIn giving the T3D classes at ARSC, it is typical that the students come in pairs: The professor/advisor who understands the code and the graduate student who will implement the code. The professor will provide the experience in this application and the student will provide the blood, sweat and time that it will take to get the code up and running on the T3D. And it will take lots of time.
InformationI have often seen a potential T3D user stuck at a point in development, unable to continue, not knowing of a library function, missing some NQS information, or just a overall view of how PEs communicate. Without this piece of information the progress stops. This is where I believe support people, like myself and CRI can do a great good. By getting this information out to the user we have a chance to keep the progress going.
There are two things that you can do with information, you can hoard it or you distribute it. I know in these times of downsizing and reduction that there is a tendency to hold onto information as it might somehow provide job security. But this tendency to hide information is easily discovered by others and then others do not share with you. I feel that we are still judged on what we produce and without the information that I receive from others, my own production would not be enough. So I have tried to share what I know in hopes that I will eventually receive in kind.
I see this conference as part of this opportunity to share information with other users. By sharing our own experience on the T3D we save others from the hard, time consuming effort that it takes to port and optimize codes on the T3D. And by making the effort less for everyone, we grow our own efforts and grow the business of parallel processing.
So to the presenters, I hope that we hold nothing back and give as much as we know. Mistakes and errors may not be flattering to ourselves, but we all learn from other's mistakes by not repeating them. And to the attendees, I hope that you take what you learn here and share it with those at your site.
Ed Kornkven ARSC HPC Specialist ph: 907-450-8669 Kate Hedstrom ARSC Oceanographic Specialist ph: 907-450-8678 Arctic Region Supercomputing Center University of Alaska Fairbanks PO Box 756020 Fairbanks AK 99775-6020
Subscribe to (or unsubscribe from) the e-mail edition of the
ARSC HPC Users' Newsletter.
Back issues of the ASCII e-mail edition of the ARSC T3D/T3E/HPC Users' Newsletter are available by request. Please contact the editors.