next up previous contents
Next: The MasPar MP-2. Up: Distributed-memory SIMD systems Previous: The Digital Equipment Corp.

The MasPar MP-1.

Machine type Processor array
Models MP1101, MP1102, MP1104, MP1208, MP1216
Front-end DECstation 5000 or DEC VAX
Operating system Internal OS transparent to the user, Ultrix or VMS on front-end
Connection structure 2-D mesh, crossbar (see remarks)
Compilers MPL: (C with extensions), MPF: (Fortran 90-like with extensions)

System parameters:

Model MP1101 MP1102 MP1104 MP1208 MP1216
Clock cycle 83 ns 83 ns 83 ns 83 ns 83 ns
No. of processors 1024 2048 4096 8192 16384
Theor. peak performance
Per proc. (Mflop/s ) 0.034 0.034 0.034 0.034 0.034
Mop/s (32-bit) 1600 3200 6400 13000 26000
Mop/s (64-bit) 800 1600 3200 6400 13000
Mflop/s (32-bit) 75 150 300 600 1200
Mflop/s (64-bit) 34 69 138 275 550
Program memory 1-4MB 1-4MB 1-4MB 1-4MB 1-4MB
Data memory 16-64MB 32-128MB 64-256MB 128-512MB 256-1GB
Int. comm. speed
Via Xnet (n. neighbour) 1.4 GB/s 2.8 GB/s 5.7 GB/s 11.5 GB/s 23.0 GB/s
Via global router 80 MB/s 160 MB/s 320 MB/s 640 MB/s 1.28 GB/s

Remarks:

The Processing Elements (PEs) of the MP-1 are more intricate than those from the DAP Gamma II (3.2.2). Each PE contains a 4-bit parallel Arithmetic/Logic Unit together with a 1-bit functional unit, a 16-bit exponent unit and a 64-bit mantissa unit. These units may be operated separately or in concord (e.g., for floating-point calculations). Because of the hardware implementation of the PEs, only 1-, 8-, 16-, 32-, and 64-bit data types are allowed.

Unlike on the DAP Gamma II, on the MP-1 it is possible to address data items in the data memories indirectly. This greatly facilitates manipulation of matrix objects indexed through an index matrix.

One type of interconnection of the PEs is a 2-D rectangular mesh (with wrap-around). This is however implemented by connecting the PEs diagonally via 3-way switches. As the setting of the switches only takes 1 cycle, this means that every PE can reach it 8 surrounding neighbours in 1 cycle. For more general routing schemes a Global Router is available. This acts, in principle, as a three-stage crossbar. PEs are arranged in clusters of tex2html_wrap_inline1148 , which connect to other clusters through the first level of the crossbar. All clusters connect via an intermediate stage to the target stage (again at cluster level). The ports from the clusters are multiplexed to the individual PEs within a cluster. As this type of communication is fairly intricate, it is much slower than via the Xnet (see system parameter list above).

As with the DAP Gamma II, there are provisions for connecting a frame buffer and/or disks directly to the MP-1. Also like the DAP, the MP-1 is essentially a single-user machine, that is, only one user at a time can have a task on the MP-1. Of course, tasks can be scheduled via a multi-user interface on the front-end system.

The MP-1 features a very nice X-window based programming environment, MPPE, which integrates an interactive source debugger, a profiler, and output windows in one environment.

Measured Performances: In [#linpackbm##1#] the solution of a full linear system was reported on a 16384 PE machine with a speed of 440 Mflop/s. The same report estimated the peak performance to be 580 Mflop/s in 64-bit precision.



next up previous contents
Next: The MasPar MP-2. Up: Distributed-memory SIMD systems Previous: The Digital Equipment Corp.



Jack Dongarra
Sat Feb 10 15:12:38 EST 1996