HALLAM@vax1.physics.oxford.ac.uk ("Phillip M. Hallam-Baker") (03/25/90)
Re Geant on transputer, I did some work at CERN recently for the DEC-CERN joint project, looking into the possible uses of MIMD machines at CERN. My report "Applying Parallel Processing at CERN is I beleive avalible form someone there. The basic problem of implementing GEANT is that any detector simulation of resonable size requires HUGE quantities of memory. This is a problem when it is run on the CRAY. The detector description database is divided into segments to allow it to be squeezed into memory. In this circumstance the CPU price becomes insignificant compared with the cost of 16-32 Mb of RAM which is the minimum required for a resonable simulation. GEANT is coded in CERN-FORTRAN a variant which depreciates any hint of program structure. The insistance on protability to the IBM means that all variable names are 6 characters or less. The upshot is that parallel decomposition is not ecconomic. If you wish to run GEANT the cheapest way is a RISC workstation such as the Apollo Domain 10000 or a VAX 9000 series if you want an operating system rather than just UNIX. There is a group trying to paralleise the code to run on the CRAY. They expect the work to take 3 man years and double the speed. The automatic parallel decomposition built into the CRAY compiler does not provide any speedup over the scalar code. This does not however imply that the CRAY compiler is no good - the CERN code is too worn for any hint of structure to be found. Phillip Hallam-Baker Oxford University ZEUS group. These are personal opinions, nothing in this letter should be taken to be a statement by any of the parties mentioned above.