We present an implementation of a stiff ODE solver on the API2OB, and discuss design considerations which should be generally applicable to ODE software on attached processors of this type.
1. Aims of the project
together with two banks of 32 fast registers and a separate integer arithmetic and logic unit which Attached processors (AP's), which run as slaves permits efficient handling of array indexing. The of a general-purpose host machine, are a cheap separate program source memory, interleaved main way to increase substantially the raw computing data memory, and a large number of data paths power of an existing computer system. They have between units which can be reconfigured at each proven successful in specialized applications where machine cycle all make the AP 1 20B a powerful a limited number of algorithms, such as FFT's and processor for intensive array computations, lackdigital filtering, are used very intensively. How-ing only the ability for direct I/O. This I/O ever, the use of AP's for general numerical work facility is provided instead by the host computer runs into algorithmic and data management prob-which communicates with the AP via a fast lems. A group of us at Bristol have been examin-Direct-Memory-Access (DMA) channel. ing a variety of numerical computational problems in the context of one particular attached processor, the Floating Point Systems API2OB, in order to 3. Fortran user-interface discover the limitations imposed by the hardware organisation and the support software. This article A given problem is performed on the host-AP describes work we have done on writing a solver pair by partitioning the algorithm between the for stiff systems of ODE's. This was intended to machines with the most computationally intensive form the lowest level of a package for solving sections performed on the AP. The data structure parabolic PDE's by collocation and the method of on which the algorithm operates must of course be lines, but the difficulties of providing a user-divided correspondingly. The AP essentially exfriendly package have been greater than hoped ecutes a sequence of subroutine calls invoked by and a lot of work remains to be done.
the host Fortran program, and each such call carries a sizeable overhead due to transfer of data and control across the host-AP interface. What 2. Hardware overview we seek is the partitioning of algorithm and data which maximises computing efficiency and mini-The speed of the AP 1 20B (a potential 12 mizes these communication overheads. Ideally one MFLOPS) comes from the pipelined floating point would perform all the computation on the AP but adder and multiplier units which can operate in the AP's program memory is generally too small parallel each producing a result every 167 ns, for this (and we found the manufacturer's overlay