Re: Parallelism (Re: Introduction)

From: J. Andrew Rogers (andrew@ceruleansystems.com)
Date: Fri Sep 09 2005 - 01:34:56 MDT


On 9/8/05 11:30 PM, "Tennessee Leeuwenburg" <tennessee@tennessee.id.au>
wrote:
> Quantum theory allows us the possibility of using action at a distance
> to ensure consistency over a large area.

Last I checked, quantum theory does not allow information to move faster
than the speed of light. Being able to absolutely guarantee synchronicity
between any two points in space at any specific moment in time is the
computer scientist's wet dream. That we pretend we can do such things does
not mean that we actually can. All we can really do is reduce the
probability of an incorrect state with clever tricks, like error
detection/correction that can handle a subset of possible failures.

IIRC, the inability to guarantee correctness of state was discussed in GEB
albeit obliquely.

> If a brain was so large that one region could only produce output that
> would affect another region "tomorrow", it would be difficult to
> imagine a resulting consciousness which experienced a consistent
> history, but was able to experience "moments" lasting less than the
> time taken to achieve consistency between brain areas.

One of the basic mechanisms of distributed computing is having every locale
have a consistent view of history, even if it is on a different point of the
timeline than other locales in some absolute context. The mark of a
generally effective distributed processing system is that if you turned off
all external input there is a very high probability that it would eventually
settle into a globally consistent state. It is worth pointing out that
databases like Oracle or PostgreSQL do exactly this as well to guarantee
correctness on single processors with high-concurrency (virtual parallelism)
and an extremely high probability of correctness. The modern Multi-Version
Concurrency Control (MVCC) family of algorithms is a distributed processing
algorithm designed to approximately guarantee eventual global consistency
even in single address spaces, never mind networked systems. And in fact,
databases do give every user a consistent tidy view of the database, but
often two different users will have different inconsistent views that are
nonetheless internally consistent for each user. As is in evidence, this
usually works out very well in practice despite the fact that different
views in the system are operating on different points in the database's
history.

For many purposes, significant synchronization latencies will not impact the
correct functioning or consistency of the overall system as viewed from any
given point in the system. However, synchronization latencies will impact
the overall throughput and intelligence of the system.

J. Andrew Rogers



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:52 MDT