This paper describes the architecture and operating system, and gives
an evaluation of NEC's new parallel computer Cenju-4 Major features of
Cenju-4 are: a) parallel memory architecture which encompasses distri
buted shared memory and user-level inter-processor communication. b) S
calable system from 8 nodes to 1,024 nodes. Using the powerful RISC pr
ocessor VR10000 (200 MHz) from MIPS II Technologies, Inc., Cenju-4 sys
tem can be configured from 8 nodes to 1,024 nodes, flexibly extending
the system as the demand arises. c) Utilization of a flexible micro-ke
rnel operating system. Since the system adopts a micro-kernel based op
erating system (MACH), it can be configured into several software envi
ronments such as UNIX double dagger server systems and, single system
image systems. The key components of the system are two 1 M gate array
s which implement memory control, inter processor communication contro
l and network communication controls. The programming environment prov
ided are de-facto standard libraries, high-level programming languages
such as MPI (Message Passing Interface), PVM (Parallel Virtual Machin
e) and HPF (High Performance Fortran). The operating system and the in
ter-processor communication libraries fully exploit the functionality
of the hardware to realize an inter-processor communication latency of
4.5 mu s and the throughput of 169 MB/s at user program level.