Load Balancing and Resource Management in the ADAM Machine
Olivier Maquelin · 1992
The dataflow architectures are well-known for their ability to expose large amounts of parallelism. However, when running highly parallel programs this property can become a serious drawback. Due to the aggressive exposure of parallelism, a large amount of storage may be needed to hold intermediate results of the computation. Diverse throttling mechanisms have been proposed so far, both in hardware and in software, but none of them seems to offer a definitive solution. For such a mechanism to work, it should not only be able to limit the exploitation of further parallelism depending on the load of the machine, but it should also be able to guess with sufficient accuracy which portions of the code should be executed next. This goal can be very difficult to achieve, especially in conjunction with a nonstrict execution model. Load balancing mechanisms were incorporated from the beginning into the ADAM architecture. In contrast to most other dataflow designs, this machine has a strict exec...