CM-5 Massively Parallel Supercomputer

1
CM-5 Massively Parallel Supercomputer
Thinking Machines Corporation
1993

• ALAN MOSER

2
CM-5 General Information

• Brain child of
• W. Daniel Hillis
• Lewis W. Tucker
• Founded Thinking Machines Corporation
• in the 1980s
• CM-5 was last in a line of successors to the
• original CM-1 Connection Machine

3
CM-5 Connection MachineHardware Overview

• 32-16384 processing nodes each of which contain
  a 32 MHz SPARC RISC processor
• 32 MB of distributed memory
• although the size may vary according to
  customer specifications
• 128 Mflops (per processing node) yielding a
  total performance of 1Teraflops. Tera240 or
  roughly 1012

4
Processing Node vs. Processor

• Processing Node not a single processor but a set
  of 5 chips
• Single 32 MHz SPARC RISC processor with
• 4 separate vector units capable of performing
  64 bit floating point and integer arithmetic

5
Diagram of Processing Node
6
CM-5 Operating System

• CM-5 runs the CMOST OS
• enhanced version of the UNIX OS
• Each processing node contains a microkernal of
  the OS

7
CM-5 MIMD or SIMD Machine?

• Referred to as synchronized MIMD machine
• somewhere between MIMD and SIMD
• best aspects of both types of machines

8
CM-5 Batch Processing or Timesharing?

• CM-5 allows both Batch processing and
  Timesharing
• Timesharing is provided by dividing processing
  nodes into a partition controlled by partition
  manager
• Protection is enforced by hardware so that one
  partition cannot interfere with another

9
Interconnection Network(s) What?

• CM-5 has not one but three overlapping
  interconnection networks
• Data Network
• Control Network
• Diagnostic Network

10
3 Overlapping Interconnection Networks
11
CM-5 Data Network

• Supports simultaneous sending of messages
  between processing nodes
• Solves several problems
• balancing message loads in network
• fetch-deadlock problem
• timesharing a parallel computer

12
CM-5 Data Network

Binary fat tree
13
CM-5 Data Network (cont.)

• Messages are passed between processing nodes
  using the least common ancestor
• Increasing bandwidth at each level up
• Avoids bottlenecks at the root node

14
CM-5 Control Network

• Designed a simple tree
• Provides synchronization so as to allow CM-5 to
  operate like SIMD computer
• In general control network provides for
• fast broadcasting of data
• barrier synchronization
• parallel prefix/postfix scan operations

15
CM-5 Diagnostic Network

• Organized as incomplete binary tree
• Able to map-out or ignore parts of the tree that
  are faulty
• Able to select and access groups of system chips
  in parallel including
• single chip
• single type of chip
• chips within a user partition
• chips associated within portion of the
  system such as board, cabinet, etc.

16
CM-5 Diagnostic Network Diagram
17
References

• The Network Architecture of the Connection
  Machine CM-5, Thinking Machines Corporation
  February 7, 1996
• Hillis, Daniel and Tucker, Lewis The CM-5
  Connection Machine A Scalable Supercomputer
• Communications of the ACM, November 1993
• Volume 36, Number 11