Chapter 8

SANDY - November 60

After Steve's call, I was quite curious about what he had in mind and what we would be working on. I knew I wanted to be involved, and told Paul about this when I briefed him on the call. The next day, after he had called Steve, we didn't really know any more about the technical aspects of the situation; Paul said to me, "I told Steve that engineers like us don't run this company, and that the project managers and business people would have to get involved. I talked to a couple of other people at the agency and gave them the names of the people here that they would need to contact. I think those discussions will be focused on business and contractual matters. When and if they come to some sort of agreement, I'm sure that we'll be involved in the technical discussions."

I replied, "OK, then. Steve mentioned that this was important from a national security viewpoint - maybe that will carry some weight with the managers."

I didn't think much more about this until the day when Paul said, "Looks like something may come of the joint project with Steve and the agency - there's going to be a meeting here in January. In the meantime, here's some paperwork for you to complete" as he handed me a long form with the heading 'Personnel Security Questionnaire'. He added, "If it's any consolation to you, I have to do one of them myself. We need to have them completed and turned in by the end of the week."

I took the form back to my office and studied it. They wanted to know almost everything about me for the past 20 years! Eventually, after I dug through some old letters back home that evening, I was able to come up with my college dorm addresses at Syracuse, and completed the form.

The week after Thanksgiving, Paul told me that the meeting was set for Monday, January 8, and would probably run for at least three days. I found myself looking forward to seeing Steve again and hearing more about what he wanted to do with the machine.

January 8 was cold and clear. Paul and I, along with two other engineers and his manager were in the meeting room when the group from the agency arrived. They were escorted from the lobby by two of the IBM business planners. Everyone introduced themselves, and I learned that, besides Steve, the other people from the agency were Dr. Winston Edwards, Sam Morris, and Murray Feldman. Our business people and Murray spent twenty minutes laying down some ground rules and reminding everyone that there were no commitments at this point, and then Murray reminded us that we would be discussing classified material, and finished by saying, "Any notes you take will need to be kept in a securely locked file cabinet."

Paul interrupted, saying, "Yes, you mentioned this to me last month. The secure file is being delivered to my office today, so there should be no issues about secure document storage."

"OK, good. Also you need to remember that you can't discuss any of this with your coworkers unless they also hold a security clearance, and then it must be on a 'need-to-know' basis."

With that, it was time for a coffee break, and after that, we finally got to the technical discussions.

Sam Morris began by saying, "I'm not going to say a great deal about the specifics of our problem for security reasons. Suffice it to say that every day we receive a large volume of data that we need to process on a timely basis. We had high hopes for this machine, but Steve and Dr. Edwards have come to the conclusion that it would need to be at least three times as fast in order to process our data in the required time. This conclusion is based on the timing numbers that Steve heard in the course, and an estimate of the most promising programming approaches that we have come up with. Steve?"

Steve continued, "First of all, our problem requires primarily character manipulation and logical operations for pattern recognition - there isn't much computation involved. On the basis of my limited understanding of the VFL unit, I don't see any way that it could be enhanced to come close to our requirements. One thing that we might be able to do with this data is to split it up into chunks and process the chunks in parallel, using independent data paths. I'm thinking in terms of an attached processor, with a number of parallel data paths."

I asked, "What information can you give us on the operations that would be needed for your application, and the nature of the data?" I wasn't completely comfortable with the parallel data paths.

It turned out, when Steve and Sam started talking, that although the individual operations involved in processing a single chunk were simple, the algorithm ended up being quite complex, and that the processing of a chunk needed input from the results for the previous chunk. However, there weren't that many independent data streams in the input data. Finally, after a bit of thought, I said, "It looks like you don't need the parallel paths as much as you need parallelism in the processing of the chunks as they come through the processor. Is that right?"

Steve said, "Let me think about this for a while. I think you may be onto something here,."

At that point, Paul interrupted the discussion and told us that if we wanted any lunch, we had to break now since the cafeteria was about to close. I found myself at a table with Steve and Dr. Edwards. I was reminded of my senior project, and asked Dr. Edwards, "By any chance, was your thesis about computational techniques for use with sparse matrices?"

"Why, yes. How in the world did you come across it?"

"Well, I graduated from Syracuse a year and a half back. In my senior year, I, along with two other engineering students, did an independent study project. We did a feasibility study to determine if a special purpose computer could be built to implement your methods. Once we got a good feel for the math, we came up with a reasonable approach to a hardware implementation. Unfortunately, we were way ahead of our time; I'm not sure that we could do it today even using the Stretch technology."

He seemed quite interested, and asked me, "Do you have a copy of your report? I'd love to look it over."

"I'll bring it in tomorrow and the repro room can make you a copy. To be honest, I hadn't thought of it much since I graduated."

Steve didn't say much, but I think he was impressed; he seemed to look up to Dr. Edwards.

We went back into the meeting; Paul took the lead and said, "While we're thinking about how we might process this data, can we get some feel for how much sustained bandwidth we're going to need to get the data in and out of the machine?"

Sam answered, "That's very sensitive information, but obviously you guys need to know this. We need to process data in five hours, max; otherwise the results are too old to be useful. We think we're looking at a minimum of twenty million characters for each five hour run."

I knew that this rate was well beyond the capabilities of the standard 729 tape drives that were attached to Stretch, and said so. Then I asked, "How does this data come in, and how would we get it into the machine in the first place?"

Sam said, "Well, I can't give you a full answer to that, but we would get it onto tape and then run the tapes through the program."

I was alarmed. I asked, "I suppose, then, that writing the data to the tapes counts as part of the five hour processing requirement?"

Sam allowed as how this was the case. Steve said, "Well, I have a thought. One of my MIT classmates, Chad Wilkinson, is working right here at IBM on tape drive development. The last time I saw him, he was talking about something called Hypertape. Can we get him cleared so we can kick this around with him and the tape people?"

Paul said, "Definitely. I'll talk to the project manager over there tomorrow, and get something going. However, if I remember the performance numbers, that tape drive is only about twice as fast as the 729. I'm afraid that isn't going to be good enough."

Steve said, "Maybe it would work. If we have this attached box, it can run its own program at full speed while the Stretch program is collecting data and writing it out to tape. Let's see how much interference and multiprogramming overhead we would have if we went this way."

By this time, it had gotten dark outside, and we were exhausted from the intense meeting. As we were breaking up, I asked Steve, "Do you guys have a car this trip, or are you stuck at the motel with nothing better than Howard Johnson's?"

Steve said, "Yes, fortunately, Murray sprung for the authorization for us to rent a car once we got here."

"I live at home with my mom, and she's expecting me for dinner tonight, but let me tell you where some of the decent restaurants are located. For simple German style, there's the Little Brauhaus on Main Street, right past Raymond Avenue in Arlington. For pizza and Mexican, just down the road, south of IBM, we have the Three Cabs - that's the Three Caballeros. For seafood, there's the Anchor Inn, almost opposite the Brauhaus in Arlington. Finally, if you want a really nice French meal and your expense account is up to it, there's Chateau Kaal Rock, on a bluff overlooking the Mid-Hudson Bridge. Here, I'll draw a map for you." I grabbed a pad of paper and quickly sketched out a map of Poughkeepsie with the major roads and indicated the locations of the restaurants.

Steve said, "Thanks - that really helps, Sometime when you come to Washington, I'll show you some of the really neat ethnic restaurants we have there. Any chance you'd be free for dinner tomorrow evening?"

"Steve, that would be very nice. I'd love to have dinner with you!" I didn't know what was happening to me, but I realized that I was really looking forward to the date with Steve.

The next morning the group reconvened in the conference room. Paul started out by suggesting that we defer the bandwidth issues for the time being and go back to the question of how to structure the data path (or paths). We spent most of the day going back and forth on this. By the middle of the afternoon, Steve said, "Dr. Edwards, Sam, and I have been thinking about this and the more we think about it, the more we like your idea of a long path with multiple logic units working on the data as it progresses through the path. This way, there would be many short cycles end-to-end in the processing of the chunk of data, but it could take a new unit of data every short cycle. It's a sort of streaming mode, I guess - the data streams through the logic chain and eventually comes out the far end."

This excited me, and I said, "Streaming is exactly the concept I had in mind, although I hadn't given it a name yet. Let's work through some details tomorrow and figure out how long the logic chain needs to be, and what the feedback paths look like. We also need to get a good handle on exactly what operations the individual elements of the chain must be able to perform."

Paul said, "I hope you guys realize that this isn't going to be cheap! I can't give you a price for it until I come up with an engineering estimate and our business guys add on their piece, but we're looking at something substantial."

Sam interrupted, saying, "Let us worry about that, right, Murray?"

Murray agreed with Sam. With that, we broke for the day. Steve and I walked out, and he said, "Sam and Murray have plans for the car. I'm embarrassed to have to ask, but can you drive us to the restaurant tonight?"

"Sure. You haven't seen my little car yet, have you? I think you'll like it."

We went out into the parking lot. When Steve saw my little Sprite, he grinned, saying, "Wow! This is so neat, I've always liked these little babies."

"Wait until you see what you have to do to get into it. You're almost tall enough to have real problems."

Well, Steve had to go through a few contortions, but eventually was able to fit in the passenger seat of my car. He said, "This isn't so bad. I've been looking at Volkswagen Beetles, although I almost don't need a car in the city."