A case study of large-scale parallel I/O analysis and optimization for numerical weather prediction system

• We implement two parallel I/O strategies based on MPI-IO and ADIOS in GRAPES.
• “MPI_AMR” in ADIOS is ideal if aggregator number is twice or four times that of OSTs.
• 114 MB per OST is the optimal size for each I/O on the Tianhe-1A system.
• 8 time-step aggregation by “MPI_AMR” reaches 7.69 GB/s, 68.4% of theoretical bandwidth.
• There is an optimal number of OSTs to distribute one MPI file for good I/O performance.

Numerical weather forecast is a most efficient means to reduce the effects of unexpected weather events. With the increasing prediction precision and the time-critical requirement, technologies of high performance computing have been improved much. However, I/O has become a significant performance bottleneck when scaling up to thousands of processes. In this paper we analyze the I/O access patterns of GRAPES (Global/Regional Assimilation and Prediction System) for numerical weather prediction system as a case of regular multi-dimensional data access. And we implement two parallel I/O strategies based on MPI-IO and ADIOS (Adaptive I/O System), making full use of efficient synchronous I/O schemes. For ADIOS, the “MPI_AMR” method is employed to improve the parallel output bandwidth, which uses aggregator processes to execute I/O operations and write to one subfile on one OST for each aggregators, reducing I/O conflicts. Experiments show that the two optimizations outperform the original sequential I/O access, achieving very impressive improvements on Tianhe-1A system and Subway Bluelight system in China. The I/O cost based on ADIOS only accounts for no more than 9% scaling up to 2K processes on Tianhe-1A system, while the sequential I/O costs more than 50% of total time when scaling to 1K or more processes. It is also found that the aggregate output based on ADIOS achieves better output performance improvements, whose peak reaches 3.84 GB/s with one time-step output on Tianhe-1A system. On the contrary, MPI-IO has obtained good input performance improvements, whose peak reaches 4.55 GB/s.We use the GRAPES’s I/O component as a benchmark to make a further study on I/O performance using ADIOS. From the rules found, we can design an efficient scheme of using “MPI_AMR” for ADIOS on Tianhe-1A system. We take 15-km horizontal resolution for instance. Since the maximum number of OSTs available for our test on Tianhe-1A system is no more than 80, 32 or 64 OSTs are chosen to facilitate parallel I/O. Then the number of aggregators should be set as 64 or 128. The optimal data size of 114 MB on one OST on Tianhe-1A system can be tested by simple cases. If we use 32 OSTs with 1024 processes, then 4 time-step aggregation can be calculated out, which obtains optimal I/O performance under such number of OSTs. It is true of the situation of 64 OSTs used. Hence, time-step aggregation is useful for output optimization based on “MPI_AMR”, whose peak reaches 7.69 GB/s on 2K processes with 64 OSTs and 128 aggregators if 8 time-step aggregation is used.We also examine the performance effects of data layout in the Lustre file system based on MPI-IO, which implies that data distribution on more OSTs outperforms a limited number of OSTs used, while the I/O performance is more likely to be disturbed with data distributed on most of all the OSTs. This influence is more apparent based on MPI-IO compared with ADIOS.