Into the imperative depth

Imperative, I am talking about Turing machines, I am talking about Von Neumann machine, I am talking about read and write, but I am also talking about distributed systems. That has been ignored. Models of distributed systems have been ignored: as a result things are not working. At various levels, but especially at deep architectural levels, be it the Internet or multicore computing.

What are distributed systems? Perhaps one simple axiom gives you some idea:

The only way one can know the state of a remote system is to communicate with it.

And another (related) axiom, this time about a general idea to make your program run fast:

Let your program make the most of the available network bandwidth.

In other words it is a good idea to overlap computation and communication since without communication computation cannot proceed: in fact it is best to regard communication (the process of messages being sent and passed and delivered) as part of computation. In distributed systems we can see this in Internet: and in computer architecture we can see this in Transputers.

You may realise we can shorten these two axioms into one:

Communication and Concurrency

That is we cannot have concurrency without communication (in effect). Communication comes first. This is what Robin Milner said in 1980.

Imperative computation has something weird. The most weird thing is it runs. Another weird thing is it seems to become more easily cooked when it is sequential but by itself it has a strong need for asynchrony and concurrency (on this we shall come back in a later post). In other words its usual outlook may be a bit different from what it really desires in its instinct.

There are several things which have not been working: one of these things is models of distributed systems. That is why we cannot see very clearly what is taking place in CMP, chip-level multiprocessing, that is to say things are happening asynchronously and concurrently inside a chip so we say it is multiprocessing and it is happening on a chip so we say it is a CMP. And we seem to be struggling about understanding what we are struggling about. We feel there are difficulties but we do not know what they are.

A bad thing is the same thing seems to be true with the direction of hardware... When transputers were designed there was a clear understanding of what they wanted and that was deeply achieved... The history did not go in the direction of transputers, it is true. And ironically (but in fact inevitably) the other history, which is the actual history we have experienced, again leads to this idea of concurrency and asynchrony for imperative computing. But if you do not know what you wish to do then what will you do? If only so-called marketing is in your mind you see what you believe to be real which is something immediately possible and desirable or at least looks possible and desirable which means it is an answer interesting to almost nobody when it is finally introduced to this world (that is to say we do not see the potential since marketing or at least bad marketing only sees what is visibly desired now, not what is possible, since for the latter one needs true imagination which is different from both a fancy forgetting hard reality or a conformism which again is unrelated with reality in effect). Let me add in a hurry I am not criticising ignorance of others, far from it, since I confess that I do share that ignorance with everybody, in the sense that I do personally feel a severe lack of understanding, which I hope we can make clear together or at least investigate into it.

I think we have forgotten what the Von Neumann machine is about, what struck Alan Turing on one rainy day 80 years ago which started this science, this engineering, this adventure.

(To be continued.)

kohei