K. Suzuki et K. Yabuki, DESIGN AND IMPLEMENTATION OF GANG SCHEDULER ON MULTI-NODE SX-4 AND SX-5 SUPERCOMPUTER SYSTEMS, NEC research & development, 39(4), 1998, pp. 396-401
SX-4 and SX-5 series multi-node models have distributed shared memory
architecture with 16 or 32 nodes connected together through an IXS (In
ternode Crossbar Switch). Each node is a complete 32 or 16 processor s
hared memory parallel vector supercomputer. Parallelization of applica
tions, both within a node and across multiple nodes, can take full adv
antage of the performance of the series. However, to achieve the highe
st possible performance, careful attention to job and processor schedu
ling is required so as to ensure that sufficient resources are always
available precisely when required. The general requirement is to minim
ize waiting time for any system resource such as memory, processor, or
storage. Ln the case of processor resource, this is sometimes difficu
lt under general multiprocessing conditions with multiple parallelized
applications simultaneously contending for processor resource. This p
aper presents the design of the Gang scheduler in the SUPER-UX operati
ng system as used on all SX series systems. The Gang scheduler automat
ically and efficiently synchronizes the allocation of processors even
in a multiple parallel job environment where total available processor
s are oversubscribed.