Portal de Periódicos da CAPES

The sto//mann data flow machine

1989; Springer Science+Business Media; Linguagem: Inglês

10.1007/3540512845_55

1611-3349

Elvira Glück-Hiltrop, Matthias Ramlow, Ute Schürfeld,

Distributed systems and fault tolerance

Most of the data flow architectures proposed in the last years, were designed to run in the area of number crunching and to exploit fine grain parallelism. Fine grain seems to lead naturally to (expensive) special hardware. Our aim is to adapt the data flow concept to commercial applications as for example from the database area and to make use of variable respectively coarse granularity. For this reason and because of market acceptance we built a machine using off-the-shelf hardware components and lifted the data flow mechanism from hardware onto software level. To avoid the classical bottleneck of "firing" we introduced distributed parallel firing. The dataflow mechanism is implemented on a shared memory architecture. The chosen hardware configuration is based on Motorola 68020 processors — each equipped with 4 MB dual ported RAM — connected to a VMEbus. The host processor runs the UNIX V.3 operating system whereas the other nodes run SRTX, a realtime operating system kernel. Subjects of investigation on the prototype are among others the correct choice of grain size, the appropriate ratio of firing control units and execution units and questions related to the topic of load balance.