11.1 Why memory management is important

Productivity is related to total throughput of a computer system over time. The throughput is limited by how well a mix of jobs fits into available storage (memory and disk) and how fast the mix executes. At GFDL, the job mix is a combination of production models, analysis, development and interactive work. It is time dependent and is determined by guidelines intended to optimize total throughput. All jobs tie up a certain amount of memory^11.2. The amount of memory consumed by high resolution models has a cruicial impact on overall productivity because it can easily exceed the computational storage capacity of the system. How much storage is enough? To admit most eddy structures would require a resolution of about ${1/12}^\circ$ which would take about 930MW for one time level of one variable in a global model assuming 100 vertical levels^11.3. Even low resolution models that use memory wastefully limit the number of jobs in the system and therefore also impact overall throughput. The dataflow scheme described below utilizes a memory window to minimize model memory requirements. Without this memory window, memory requirements could easily increase by one or two orders of magnitude on a single processor! This is the scenerio for large jobs where the bulk of data must stored on a rotating or preferably solid state disk. For smaller jobs, the bulk of data may be stored on a ramdrive^11.4 and for this scenerio, memory requirements on a single processor would increase by at least 50% if a memory window were not employed.