5 ms·

What are your current thoughts on objects and simulation in distributed environments? You have historically worked or advised on projects using approaches that

• alankay1 10y ago
  Good questions. To keep my first comment as short as I could make it, I didn't mention the next wave of researchers who got interested in this perspective on computing, for example: Dave Reed ("the slash in TCP/IP") in his 1978 MIT thesis, Dave Jefferson and Henry Sowizral at RAND ("Time Warp"), Leslie Lamport, etc. And the next round about 15 years ago -- for Croquet -- which included Dave Reed, David A Smith, Andreas Raab, myself, etc. There were three or four different versions of this mechanism, all of which worked, and the designs progressively got better. There are several sets of issues and tradeoffs. Reed's initial ideas in the 70s were about what an operating system for the whole Internet should look like. One of the many interesting ideas he proposed was to deal with the slow latencies and large numbers of potential users by distributing cloned computations organized by pseudo-time, and using the slow Internet for just input and synch deltas. This is what we first implemented in Croquet in the early 2000s when a typical good Internet round trip for a ping was about 80-100 milliseconds. This is good enough to create a massively parallel game like World of Warcraft (or even a flight simulator game) without any servers at all by just using the distributed machines that the game players are using to play on. Further versions did more, and more comprehensively. (A side note is that Croquet without the real-time graphics and interaction, etc. automatically provides a distributed object-oriented data base, etc. Pseudo-time is the larger idea underneath protected distributed atomic transaction data-bases, etc.) Your last question is a good one, and it has been discussed over the years. Can it be done? How much would need to be done, and where? Etc. The object work we did at Parc was a part of the network thinking going on in the ARPA/Parc community. And there was an eventual intent to think of higher-level network entities. A quite good intermediate real implementation of "net level entities" was the LOCUS system by Gerry Popek and his group at UCLA after he spent a year at Parc thinking about what network systems should be like (this was really good, and the first several chapters of his book are well worth reading). It was a migratory load balancing idea over heterogeneous machine types, which was quite complementary to the pseudo-time ideas. I would like to see a highly talented group take another pass at your last question. Resilience to scaling is quite difficult to predict ahead of time, and it's good to do real implementations. Still, the 9 orders of magnitude that the Internet needed was possible with the scheme that the ARPA/Parc community had, and Croquet worked perhaps better than it should have (especially with just a few people working on it). In any case, I think it is a bit crazy to not use pseudo-time if you know about it, and it's a bit amateurish not to be aware of it. In the middle are religions whose identities are built on various "one true way"s (and this is the opposite of the way any science should be conducted).
  • asolove 10y ago
    Thank you very much for this answer. I am glad to hear that you still expect pseudo-time approaches to yield fruit. And I appreciate some pointers to more recent work in this direction. I only recently read David Reed's thesis and did not know about this later work until you mentioned it. Almost all of the links I can find about Croquet no longer work. (OpenCroquet and OpenCobalt websites seem to have disappeared.) Can you recommend reading for someone who wanted to understand the design decisions made in the several versions of Croquet? I have been experimenting with layering pseudo-time on top of asynchronous message-passing in Erlang. One useful trick which I have not yet seen published is that you can have two-dimensional pseudo-times, so that one running system can track multiple chains of causation. (You would not want this happening all the time, since the whole point is to get a consistent history everywhere. But it is very useful for developers.) An example: Let's say I have an existing object and a clone of it where I have changed some of the methods' behavior, but I want to understand all the consequences of these differences across all the distributed objects they cooperate with. I can send them the same message with different y-coordinate pseudo-times. In a system that relied on wall clock time, the side-effects of these two message sends would merely step on each other non-deterministically. But with pseudo-time, their effects are causally distinct throughout the system and can be dynamically inspected, making it a lot easier to build a human understanding of the behavior of the distributed system.
    • alankay1 10y ago
      Yes, I think you are characterizing what is called "possible worlds reasoning" in AI. A notable system was "ART" (Automated Reasoning Tool) by Inference Corp some years ago. More recently you might be interested in the "Worlds" work by Alex Warth et al. http://www.vpri.org/pdf/tr2011001_final_worlds.pdf http://www.vpri.org/pdf/tr2011001_final_worlds.pdf. This is a highly efficient design and implementation of fine grain pseudo-time contexts for general programming. But what you are doing is good and important. We like a quotation by Goethe: "We should all share in the excitement of discovery without vain attempts to claim priority". In other words, it doesn't matter when good ideas happen -- it is enough that they do happen.
    • vanderZwan 10y ago