4:00 PM - 4:15 PM
[AOS14-08] On the speeding up of OGCMs by using a single precision advection scheme
Keywords:Ocean model, Single-precision computation, Advection scheme
The use of single-precision computation instead of double-precision computation in current supercomputer systems is expected to increase speed by up to a factor of two, as twice as many vector operations and communications can be packed in the registry. (The same applies to the use of GPUs). However, since it is a factor of 2 at best, it was not a priority to work on it at a time when the computing power of supercomputers was growing exponentially year by year, because it is not an easy task to deal with side effects such as digit drops due to single-precision computation. However, in the absence of significant speed increases in supercomputers in recent years, a doubling of speed has become attractive.
As a first step in speeding up the Meteorological Research Institute's Common Ocean Model (MRI.COM) by increasing the use of single-precision computations, we attempted to use single-precision computation for the Second-Order Moment (SOM) scheme, a highly accurate but computationally expensive tracer advection scheme in which the distribution of tracers within a grid is approximated by a polynomial of order 0 to 2 and advection is performed. Simply converting all the variables of the SOM scheme to single precision reduced the computational cost of the advection scheme and the overall scheme to 55% and 86% respectively, but a non-negligible error of a few degrees appeared in the temperature of the deep layer when the global model was integrated over a long period (three centuries). Ideal experiments studies have shown that this temperature increase is due to the growth of errors in the evaluation of the continuity equation. To improve this, only the zeroth-order terms in the tracer advection scheme, which are also used in the evaluation of the continuity equation, were changed to be calculated with double precision, and the error was significantly reduced to less than 0.01 degrees in the 300-year integration. The advection scheme and overall calculation costs were 57% and 87%, respectively, which were not much different from those of the single-precision case. The devised single-precision advection scheme was also tested in an eddy-resolving model, and no systematic differences were found between the double-precision and the advection scheme in terms of climate values and error development over several decades.
All this suggests that this devised single-precision version of the SOM is a useful option that can be used regularly.
As a first step in speeding up the Meteorological Research Institute's Common Ocean Model (MRI.COM) by increasing the use of single-precision computations, we attempted to use single-precision computation for the Second-Order Moment (SOM) scheme, a highly accurate but computationally expensive tracer advection scheme in which the distribution of tracers within a grid is approximated by a polynomial of order 0 to 2 and advection is performed. Simply converting all the variables of the SOM scheme to single precision reduced the computational cost of the advection scheme and the overall scheme to 55% and 86% respectively, but a non-negligible error of a few degrees appeared in the temperature of the deep layer when the global model was integrated over a long period (three centuries). Ideal experiments studies have shown that this temperature increase is due to the growth of errors in the evaluation of the continuity equation. To improve this, only the zeroth-order terms in the tracer advection scheme, which are also used in the evaluation of the continuity equation, were changed to be calculated with double precision, and the error was significantly reduced to less than 0.01 degrees in the 300-year integration. The advection scheme and overall calculation costs were 57% and 87%, respectively, which were not much different from those of the single-precision case. The devised single-precision advection scheme was also tested in an eddy-resolving model, and no systematic differences were found between the double-precision and the advection scheme in terms of climate values and error development over several decades.
All this suggests that this devised single-precision version of the SOM is a useful option that can be used regularly.