Bolt, Beranek and Newman (BBN)

Bolt, Beranek and Newman (BBN)

Status

Active and now a divison of GTE Internetworking.

Overview of Organisation

BBN was founded in 1948 as an acoustic consultancy company and has since diversified into computing and communications. Amongst other things BBN were involved in producing the ARPAnet communications system. BBN's first parallel processing platform was the Pluribus, a multi-bus machine using a number of interconnected Lockheed SUE minicomputers, which was designed as a node for ARPAnet. In 1978, BBN was commissioned by DARPA to produce a new parallel computer system, with special emphasis on the communications switching technology upon which the machine would be based. This switch became known as the butterfly because its wiring topology is similar in structure to the FFT butterfly layout. The butterfly still forms the basis for BBN computers.

The original butterfly machine was redesigned to be aimed at the time-critical systems market and as such was redesignated the TC2000 (TC standing for Time Critical). At the same time, some upgrades to the original machine were made and this became known as the GP1000, (GP standing for General Purpose).

Platforms Documented

Contact Address 

Bolt, Beranek and Newman Advanced Computers Inc.
10 Fawcett Street
Cambridge MA 02138, USA.
Tel (617) 873 6000
Fax (617) 873 3315

BBN Inc.
Heriot-Watt Research Park
Riccarton
Edinburgh EH14 4AP
Scotland
031-449-5488

Company's homepage.

See Also

BBN have a commercial server entry in the Global Network Navigator and also support a server for BBN Systems and Technologies DIS Dept . .

BBN Butterfly

Overview of Platform

The Butterfly was commissioned specifically by DARPA to investigate the switching technology upon which the machine would be based.

This machine was upgraded to the Butterfly-Plus, which was later redesignated as the GP1000.

Parallel Butterfly Network Architecture.

Compute Hardware

Motorola 68000-series microprocessors.

Interconnect / Communications System

Butterfly switch consisting of a set of 4 by 4 crossbar switch modules. This is inferior (but cheaper) than a full crossbar, and gives better performance for a practical number of processors than would a bus based communications system.

Memory System

500kBytes of memory per processor.

Benchmarks / Compute and data transfer performance

N/A

Operating System Software and Environment

N/A

Networkability/ I/O System / Integrability / Reliability / Scalability

The Butterfly was configured as a compute engine with a front-end host machine on which user code was developed.

Notable Applications / Customers / Market Sectors

Between 30 and 40 systems were sold, mostly to universities, interested in the machine's novel architecture.

Overall Comments

The butterfly switch is undoubtedly the most noteworthy feature of this machine.

BBN GP1000

Overview of Platform

The Butterfly evolved into the Butterfly-plus, which no longer had a fron-end host and provided its own operating system - namely the MACH UNIX developed by Carnegie Mellon University

The Butterfly was also redesignated as the general purpose GP1000.

The Butterfly GP1000 is a tightly coupled, shared memory multiprocessor housing up to 256 processor boards, each with an MC68020 microprocessor and an IEEE-complying MC68882 floating-point coprocessor. Every processor board includes 4 Mbytes of globally shared memory. Any processor can access any memory location through the Butterfly switch, a fast, modular, multi-stage interconnect. Processors also have direct access to their own 4 Mbyte share of the global memory pool. Providing true parallel access to memory, the Butterfly performs up to 256 simultaneous reads or writes and automatically resolves contention for memory.

Other architectural features include: 

Multiple instruction, multiple data (MIMD) architecture.
Up to 600 mips of processing power in 2.5 mip increments.
All processors have access to as much as 1024 Mbytes
(one Gbyte) of main memory.
Memory bandwidth up to 1024 Mbytes/sec (one Gbyte/sec).
Memory access time is typically less than 1 microsecond,
  (4 microseconds worst case --- without contention. )
Distributed I/O system supports RS-232, RS-449, Ethernet, Multibus,
and VME bus.
Field expandable in single processor increments.

Compute Hardware

Motorola 68020 series microprocessors.

Interconnect / Communications System

Butterfly switch consisting of a set of 4 by 4 crossbar switch modules. This is inferior (but cheaper) than a full crossbar, and gives better performance for a practical number of processors than would a bus based communications system.

Memory System

4Mbytes of memory per node.

Benchmarks / Compute and data transfer performance

N/A

Operating System Software and Environment

MACH Unix developed by Carnegie Mellon University was used on the GP1000.

Software includes: Mach 1000, the GP1000 operating system, is based on Berkeley 4.3bsd UNIX, with extensions for parallel processing. The GP1000 supports C, Fortran 77, Common Lisp, and Scheme (a Lisp dialect). Ada is being developed. All languages are extended naturally to support parallel structures. A rich, graphically-oriented debugging environment is provided.

Networkability/ I/O System / Integrability / Reliability / Scalability

No front end host was used on the GP1000, the operating system was implemented directly on node processors.

The GP1000 is a standalone system supporting a full range of peripherals including 500 Mbyte and 850 Mbyte disk drives; 1/4" cartridge and 1/2" reel-to-reel tape drives; a flexible terminal control system; and an Ethernet interface.

Notable Applications / Customers / Market Sectors

Approximately 60 GP1000's are believed to have been sold, some 75% to universities.

Overall Comments

This machine seems to have been succesful in attracting sales, albeit to universities and other organisations interested in experimenting with the architecture rather than as a standard compute engine.

BBN TC2000

Overview of Platform

The TC2000 was introduced in July 1989 and by 1991 was the only BBN machine being actively marketed.

The TC2000 is a shared memory, medium grained, MIMD computer on whose individual processors the user runs different tasks. The machine may be partitioned into autonomous clusters, to allow multiple applications to be run concurrently. Cluster allocation and management is dynamic so extra processors from the pool can be allocated to especially intensive jobs.

The multiprocessing architecture of the BBN TC2000 allows field-expansion from eight to 504 processors, with corresponding increases in memory, memory-access bandwidth, and I/O capabilities. TC2000 system supports two operating systems concurrently. At the same time as some processors are running the pSOS+m real-time executive for time-critical applications, others can be using the nX operating system (based on UNIX 4.3 BSD) for either analysis or time-shared program development.

A major feature unique to the TC2000 system is its software- controlled clustering capability. Processors can be assigned to groups or clusters, which are then designated for either nX or pSOS+m operation. Different sections of an application can be run concurrently on each one. In addition, data can be shared within and between clusters, so a TC2000 system can integrate various segments of a complex application traditionally dispersed among a number of loosely-coupled computers. Processor allocation is dynamic, meaning that resources can be reallocated during an actual run.

To reduce the time and cost of applications development, the TC2000 system includes the only graphical development tools specifically designed for a multiprocessor environment. Based on the X Window System standard, the Xtra (X Tools for Runtime Analysis) environment makes it easier for programmers to handle the complexities inherent in multiprocessor programming. Included within the Xtra environment is the TotalView source- level, multi-processing debugger and the Gist graphics- oriented performance analyzer. Optimized Ada, Fortran-77 and C compilers are also available for the TC2000 system.

Compute Hardware

Motorola 88000-series RISC chips. A node is made up ofan 81000 processor with three 88200 cache/memory management chips and the Butterfly switch interface.

Interconnect / Communications System

Butterfly switch consisting of a set of 8 by 8 crossbar switch modules. This is inferior (but cheaper) than a full crossbar, and gives better performance for a practical number of processors than would a bus based communications system. A two stage 8 by 8 crossbar allows a maximum of 64 nodes in the machine, but by using a 3 stage switch this can be increased to 512 nodes.

Memory System

Each node has either 4 or 16MBytes of memory, and the three 88200 chips all have 16kBytes of four-way associative cache memory.

Benchmarks / Compute and data transfer performance

The peak I/O rate for a 64 node machine is determined by that of the VME interfaces and is quoted as 320MBytes per second. 

MAXIMUM  SYSTEM PERFORMANCE (504 processors)
Integer                    9,576 Dhrystone MIPS
Whetstone                  6,552 Whetstones
Floating Point            10,080 MFLOPS
Memory                    16,096 MBytes
I/O Bandwidth              2,560 MBytes/sec

Operating System Software and Environment

Message routing through the butterfly switch is based on a network packet routing system. The operating systems provided are based either on UNIX or on a proprietary system designed for real time applications.

Performance monitoring and debugging facilities toolsets are also provided.

Disk access is through a SCSI system connected to the VME bus on each node.

Sustainable rates of 7MFLOPS per node have been reported independently of BBN.

Networkability/ I/O System / Integrability / Reliability / Scalability

Each TC2000 has at least one VME interface to support attached graphics devices and disks, and up to 5 VME interfaces can be used on one switch module, thus increasing the available bandwidth which is important for time critical applications.

Notable Applications / Customers / Market Sectors

The real time or time-critical applications market is what the TC2000 was targetted at. At least 12 systems are known to have been sold.

Overall Comments

This machine embodies some interesting ideas in switch networking but it is difficult to see how the butterfly switch can be scaled to very large sizes in a cost effective manner.

hawick@npac.syr.edu
saleh@npac.syr.edu