Aspects of Using Multiprocessors for Meteorological Modelling

Since the introduction of computers for scientific research, meteorological modelling has always been one of the first disciplines to use the most advanced computers available at any given time. Today, meteorological institutions throughout the world operate class VI supercomputers like the CYBER 205, CRAY-1, or CRAY X-MP. Average processing speeds in excess of 200 Mflops (106 floating operations per second) are achievable for global weather prediction models with a horizontal resolution of about 100 km in grid-point space. As the results of the operational meteorological models are only useful during a very short time--the forecast for today is history tomorrow--the demands for greater processing power of computers are obvious, especially when one considers that a doubling in horizontal resolution extends the forecast range but in turn increases the computing requirements by the factor 8 = 2 3. The next logical step envisaged by meteorologists for the end of the decade is to double the model resolution without changing the elapsed time for the computation. Therefore, a computer capable of delivering around 1600 Mflops on average and in excess of 3000 Mflops as peak rate is required. This estimate of the computing speed assumes that the complexity of the calculations does not increase as well.