Realizing Massive Parallel Computing in the Earth Simulator

Current situation on model development and points to consider when conducting numerical calculation to be utilized on the Earth Simulator will be explained by Dr. Keiko Takahashi.

Dr. Keiko Takahashi (Integrated Modeling Research Program)

The Earth Simulator is a distributed memory parallel system that consists of 640 processor nodes, each with shared memory vector processors, connected in a high-speed network. Two key factors are involved in the execution of numerical calculations that fully realize the massive parallel computing capacity of this supercomputer. One of these is the overall superior vector performance capability for large-scale data processing. The other factor is the Earth Simulator's ability to perform parallel computing, whereby the level of communications between processors is relatively small, and processing can be performed with each processor subjected to an equivalent load.

Towards its goal of elucidating the mechanisms of climate change, the Frontier Research System for Global Change (FRSGC) is endeavoring to use the Earth Simulator-the largest and fastest computer of its kind in the world-to conduct unprecedented large-scale numerical experimenta-tion, which could not be realized previously due to the requirement of overwhelmingly long processing time. Numerical calculations using general circulation models of the atmosphere and ocean, and coupled ocean-atmosphere models are regarded as suitable for vector processors. Merely altering the parameters of such models that have been used to date, however, does not make them actually capable of being used without any modification to execute large-scale calculations in the Earth Simulator.

Generally, the higher the number of processors used, the lesser the processing time required by parallel computing. In reality, however, the whole computing performance is affected by the method of communications between the processors and the quality of the computation sequence. The impact of changes in computation sequencing and the balance of communications between processors, which are not relevant when using a single processor, must be factored into the performance equation. The higher the number of processing elements used for parallel computing, the more often the parallel effect cannot be obtained, depending on the method of parallel computing used. Therefore, these problems need to be carefully examined when developing models.

Fig.1 Relationship of parallelization ratio and the number of processing elements (PEs) with the speedup ratio (Amdahl's Law)

The FRSGC is addressing the above issues, and with the cooperation of the Center for Climate System Research, University of Tokyo and the collaboration among many of the FRSGC Research Programs, we are able to vigorously promote the simulation and verification of climate change phenomena using general circulation models for the atmosphere and ocean, and coupled ocean-atmosphere models. At the same time, cooperation with the Earth Simulator Research and Development Center, which has detailed knowledge of the Earth Simulator specifications, allows the simulation of various climate change scenarios and the development of models that incorporates the maximum benefit from the Earth Simulator's capabilities.

The ultimate goal is to develop a coupled ocean-atmosphere model that would be able to realize a high-resolution coupled model for the atmosphere (T213L50: horizontal resolution of approximately 0.5 degrees) and ocean (horizontal resolution of 0.1 degrees, 50 vertical levels) for 1000 years of integration in three to four months, which would require some 600 years to calculate with current single vector processors.

Fig.2 Percentage of contributing to parallelization

Back Cover Next

FRSGC Website