QuoteRef: hillWD_1985

topics > all references > ThesaHelp: references g-h

references g-h
problems with the von Neumann architecture
semantic networks
massively parallel processors
hardware for interprocess communication
data structures
data parallel processing
synchronized processing
concurrent operations
vector processing
parallel control statements
co-sequence operations
sequence reduction
interprocess communication
computer architecture
data flow machines
set operations
set construction
pointers to data
external search and sort
computer science
discrete vs. continuous
sensitivity of software to change
local vs. global
digital communication
Petri net
memory management
physics as computation


Hillis, W.D., The Connection Machine, Cambridge, Massachusetts, MIT Press, 1985. Google

4 ;;Quote: in a von Neumann computer, most transistors do nothing most the time; only for memory
10 ;;Quote: Connection Machine motivated by semantic networks; e.g., processor for 'apple' connected to processor for 'red'
14 ;;Quote: the Connection Machine has as many processors as are needed for a problem; each processor must be small
15 ;;Quote: the physical connectivity of the Connection Machine is controlled by software; in order to match problem's structure
18 ;;Quote: the Connection Machine represents and processes data as 'active data structures' of interconnected processor and memory cells
32 ;;Quote: both CmLisp and the Connection Machine achieves parallelism through simultaneous operations instead of concurrent control
33 ;;Quote: parallel operations in CmLisp via a xector--a set of processors with one value per processor; e.g., xector add
33 ;;Quote: a xector element consists of a index (processor name and memory address) and a value; includes sets, index sequence, and constants
37 ;;Quote: in CmLisp, use .alpha. to convert a value into a constant xector; i.e., the value is loaded into every processor
39 ;;Quote: apply a xector of functions to all tuples with common xector indices; e.g., (.alpha.+ '{a->1 b->2} '{b-3 c->2}) => {b->5}
40 ;;Quote: in CmLisp, use .bullet. to selectively cancel the constant meta-operation .alpha.
40 ;;Quote: think of CmLisp's .alpha. as a way to get a zillion of something; .bullet. marks subexpressions that already have a zillion
41 ;;Quote: in CmLisp, .beta. reduction applies a function to the values of a xector in logarithmic time; ignores indices
46 ;;Quote: the simplest .beta. reduces a xector to a value; general form reduces portions of xectors and produces a xector
46+;;Quote: CmLisp's .beta. operation corresponds to the message routers while its .alpha. operation corresponds to processor execution
46 ;;Quote: the generalized .beta. creates a xector from a value xector, an index xector, and a combining function for duplicates; gives example
47 ;;Quote: simple .beta. reduction is the generalized .beta. operation applied to one xector; every value is combined
47 ;;Quote: a Connection Machine is a hardware implementation of CmLisp and .alpha./.beta. reduction
47+;;Quote: xectors are the contents of Connection Machine memory cells; processors are .alpha. operations and routers are .beta. operations
54 ;;Quote: the communications network of the Connection Machine does most of the computation, limits the performance, and costs the most
59 ;;Quote: a hashnet is a random network topology; does well compared to proposed networks and easy to analyze
71 ;;Quote: description of the CM-1 Connection Machine; 64K cells, 4K bits memory each, 1-bit ALU, boolean n-cube topology, host computer
74 ;;Quote: description of wide microinstruction for CM-1 Connection Machine; truth table on 2 bits and 16 flags, N/E/S/W pins for I/O
80 ;;Quote: when message delivered to a router, corresponding address bit is cleared; all done when address is 0
82 ;;Quote: the Connection Machine handles message congestion by referring messages to adjoining routers; one step further from destination
84 ;;Quote: a CM-1 prototype achieved 10^10 bits/sec random-message bandwidth; for local traffic expect 4x larger bandwidth
91 ;;Quote: an active data structure is a machine; the host controls the Connection Machine by telling the data (processor/memory) what to do
92 ;;Quote: set union, intersection, and universal qualifier can be unit-time operations on the Connection Machine
92 ;;Quote: the domain of a xector; mark members of the set with bits, tags, or pointers
92 ;;Quote: in a Connection Machine can allocate one bit per cell to indicate set membership
93 ;;Quote: if sets are disjoint, can represent membership by a tag value identifying the cell's set
95 ;;Quote: can represent a set by a root cell pointing to a few fanout cells that point to the set's members
96 ;;Quote: on the Connection Machine, convert pointer-represented sets to bit representation by propagating a marker through the fanout cells
98 ;;Quote: the Connection Machine uses trees and butterflies for collecting, combining, and spreading information, e.g., sets; need two-way pointers
106 ;;Quote: a butterfly structure avoids the exponential behavior of trees by keeping the levels constant sized; omega network, perfect shuffle, FFT
109 ;;Quote: can implicitly represent regularly structured trees and butterflies by address
112 ;;Quote: the Connection Machine can shift arbitrarily large segments of data in unit time; with type codes can update pointers in unit time
112 ;;Quote: the Connection Machine can search for substrings in time proportional to the length of the substring
137 ;;Quote: computer science is messy because it lacks the locality, symmetry, and invariance to scale found in physics
138 ;;Quote: in physics, action has local effects (e.g., inverse square law); in computation, a tiny program can clear all of memory
138 ;;Quote: the old conception of computation treated wires as idealized, instantaneous connections
138+;;Quote: in computers, wires are much of the cost, space and delay times
138+;;Quote: memory is a wire turned sideways in time
139 ;;Quote: in classic computational theory, the wire is not considered but it is very important to engineering; mismatch with reality
140 ;;Quote: massive communication has properties of physics--distance important, congestion acts like mass
141 ;;Quote: in a Connection Machine, could migrate cells in the direction of most communication; groups of intercommunicating cells would cluster
143 ;;Quote: computational systems will become physics-like because of physical constraints such as 3-d and speed of light
143 ;;Quote: what will computation look like with a mole of processors?; physics

Related Topics up

ThesaHelp: references g-h (299 items)
Topic: problems with the von Neumann architecture (11 items)
Topic: semantic networks (42 items)
Topic: massively parallel processors (29 items)
Topic: hardware for interprocess communication (31 items)
Group: data structures   (12 topics, 275 quotes)
Topic: data parallel processing (12 items)
Topic: synchronized processing (35 items)
Topic: concurrent operations (22 items)
Topic: vector processing (14 items)
Topic: constants (21 items)
Topic: tuples (17 items)
Topic: parallel control statements (12 items)
Topic: co-sequence operations (8 items)
Topic: sequence reduction (10 items)
Topic: interprocess communication (29 items)
Topic: computer architecture (44 items)
Topic: data flow machines (14 items)
Topic: set operations (12 items)
Topic: set construction (20 items)
Topic: pointers to data (55 items)
Topic: trees (21 items)
Topic: external search and sort (23 items)
Group: computer science   (871 topics, 23143 quotes)
Topic: physics (48 items)
Topic: symmetry (11 items)
Topic: discrete vs. continuous (47 items)
Topic: sensitivity of software to change (44 items)
Topic: local vs. global (29 items)
Group: digital communication   (11 topics, 295 quotes)
Topic: Petri net (44 items)
Group: memory management   (11 topics, 346 quotes)
Topic: time (48 items)
Topic: physics as computation (31 items)

Collected barberCB 12/86
Copyright © 2002-2008 by C. Bradford Barber. All rights reserved.
Thesa is a trademark of C. Bradford Barber.