Parallel IO in the Community Earth System Model Jim Edwards John Dennis (NCAR) Ray Loy(ANL) Pat Worley (ORNL)
CPL7 COUPLER CAM ATMOSPHERIC MODEL CLM LAND MODEL POP2 OCEAN MODEL CICE OCEAN ICE MODEL CISL LAND ICE MODEL
Some CESM 1.1 Capabilities: Ensemble configurations with multiple instances of each component Highly scalable capability proven to 100K+ tasks Regionally refined grids Data assimilation with DART
Prior to PIO Each model component was independent with it’s own IO interface Mix of file formats NetCDF Binary (POSIX) Binary (Fortran) Gather-Scatter method to interface serial IO
Steps toward PIO Converge on a single file format NetCDF selected Self describing Lossless with lossy capability (netcdf4 only) Works with the current postprocessing tool chain
Extension to parallel Reduce single task memory profile Maintain single file decomposition independent format Performance (secondary issue)
Parallel IO from all compute tasks is not the best strategy Data rearrangement is complicated leading to numerous small and inefficient IO operations MPI-IO aggregation alone cannot overcome this problem
Parallel I/O library (PIO) Goals: Reduce per MPI task memory usage Easy to use Improve performance Write/read a single file from parallel application Multiple backend libraries: MPI-IO,NetCDF3, NetCDF4, pNetCDF, NetCDF+VDC Meta-IO library: potential interface to other general libraries
PIO VDC netcdf4 pnetcdf HDF5 netcdf3 MPI-IO CPL7 COUPLER CISL LAND ICE MODEL CAM ATMOSPHERIC MODEL CLM LAND MODEL CICE OCEAN ICE MODEL POP2 OCEAN MODEL PIO VDC netcdf4 pnetcdf HDF5 netcdf3 MPI-IO
PIO design principles Separation of Concerns Separate computational and I/O decomposition Flexible user-level rearrangement Encapsulate expert knowledge
Separation of concerns What versus How Concern of the user: What to write/read to/from disk? eg: “I want to write T,V, PS.” Concern of the library developer: How to efficiently access the disk? eq: “How do I construct I/O operations so that write bandwidth is maximized?” Improves ease of use Improves robustness Enables better reuse
Separate computational and I/O decomposition Rearrangement between computational and I/O decompositions
Flexible user-level rearrangement A single technical solution is not suitable for the entire user community: User A: Linux cluster, 32 core job, 200 MB files, NFS file system User B: Cray XE6, 115,000 core job, 100 GB files, Lustre file system Different compute environment requires different technical solution!
Writing distributed data (I) I/O decomposition Computational decomposition Rearrangement + Maximize size of individual io-op’s to disk - Non-scalable user space buffering Very large fan-in large MPI buffer allocations Correct solution for User A
Writing distributed data (II) I/O decomposition Computational decomposition Rearrangement + Scalable user space memory + Relatively large individual io-op’s to disk Very large fan-in large MPI buffer allocations
Writing distributed data (III) I/O decomposition Computational decomposition Rearrangement + Scalable user space memory + Smaller fan-in -> modest MPI buffer allocations Smaller individual io-op’s to disk Correct solution for User B
Encapsulate Expert knowledge Insert images here Flow-control algorithm Match size of I/O operations to stripe size Cray XT5/XE6 + Lustre file system Minimize message passing traffic at MPI-IO layer Load balance disk traffic over all I/O nodes IBM Blue Gene/{L,P}+ GPFS file system Utilizes Blue Gene specific topology information
Experimental setup Did we achieve our design goals? Impact of PIO features Flow-control Vary number of IO-tasks Different general I/O backends Read/write 3D POP sized variable [3600x2400x40] 10 files, 10 variables per file, [max bandwidth] Using Kraken (Cray XT5) + Lustre filesystem Used 16 of 336 OST
3D POP arrays [3600x2400x40]
3D POP arrays [3600x2400x40]
3D POP arrays [3600x2400x40]
3D POP arrays [3600x2400x40]
3D POP arrays [3600x2400x40]
PIOVDC Parallel output to a VAPOR Data Collection (VDC) A wavelet-based, gridded data format supporting both progressive access and efficient data subsetting Data may be progressively accessed (read back) at different levels of detail, permitting the application to trade off speed and accuracy Think GoogleEarth: less detail when the viewer is far away, progressively more detail as the viewer zooms in Enables rapid (interactive) exploration and hypothesis testing that can subsequently be validated with full fidelity data as needed Subsetting Arrays are decomposed into smaller blocks that significantly improve extraction of arbitrarily oriented sub arrays Wavelet transform Similar to Fourier transforms Computationally efficient: order O(n) Basis for many multimedia compression technologies (e.g. mpeg4, jpeg2000)
Other PIO Users Earth System Modeling Framework (ESMF) Model for Prediction Across Scales (MPAS) Geophysical High Order Suite for Turbulence (GHOST) Data Assimilation Research Testbed (DART)
Write performance on BG/L Update slide with new data Write performance on BG/L April 26, 2010 Penn State University
Read performance on BG/L Update slide with new data Read performance on BG/L April 26, 2010 Penn State University
Coefficient prioritization (VDC2) 100:1 Compression with coefficient prioritization 10243 Taylor-Green turbulence (enstrophy field) [P. Mininni, 2006] No compression Coefficient prioritization (VDC2)
800:1 compressed: 0.34GBs/field Original: 275GBs/field 40963 Homogenous turbulence simulation Volume rendering of original enstrophy field and 800:1 compressed field 800:1 compressed: 0.34GBs/field Original: 275GBs/field Data provided by P.K. Yeung at Georgia Tech and Diego Donzis at Texas A&M
F90 code generation interface PIO_write_darray ! TYPE real,int ! DIMS 1,2,3 module procedure write_darray_{DIMS}d_{TYPE} end interface genf90.pl
# 1 "tmp.F90.in" interface PIO_write_darray module procedure dosomething_1d_real module procedure dosomething_2d_real module procedure dosomething_3d_real module procedure dosomething_1d_int module procedure dosomething_2d_int module procedure dosomething_3d_int end interface
PIO is opensource http://code.google.com/p/parallelio/ Documentation using doxygen http://web.ncar.teragrid.org/~dennis/pio_doc/html/
Thank you
Existing I/O libraries netCDF3 Serial Easy to implement Limited flexibility HDF5 Serial and Parallel Very flexible Difficult to implement Difficult to achieve good performance netCDF4 Based on HDF5
Existing I/O libraries (con’t) Parallel-netCDF Parallel Easy to implement Limited flexibility Difficult to achieve good performance MPI-IO Very difficult to implement Very flexible ADIOS Serial and parallel BP file format Easy to achieve good performance All other file formats