I/O Resources at NERSC¶
NERSC provides a range of online resources to assist users developing, deploying, understanding, and tuning their scientific I/O workloads, supplemented by direct support from the NERSC Consultants. Here, we provide a consolidated summary of these resources, along with pointers to relevant online documentation.
Libraries and Tools available at NERSC¶
NERSC provides a number of I/O middleware libraries, as well as tools for profiling I/O performed by your jobs and monitoring system status. These resources include:
- High-level I/O libraries available at NERSC, including the popular HDF5 and NetCDF libraries
- The Darshan job-level I/O profiling tool may be used to examine the I/O activity of your own jobs.
Users should keep in mind the entire HPC I/O stack in order to design the best I/O strategy for the long term.
|I/O Layer||I/O Libraries|
|Productive Interface||H5py, Python, Spark, Tensorflow, PyTorch|
|High-Level I/O Library||HDF5, NetCDF, PnetCDF, Root|
|I/O Middleware||MPIIO, POSIX|
|Parallel File System||Lustre, Datawarp, GPFS, HPSS|
Please refer to the resources below which present more detailed introductions to some of these topics in tutorial form.
Best Practices for Scientific I/O¶
While there is clearly a wide range of I/O workloads associated with the many scientific applications deployed at NERSC, there are a number of general guidelines for achieving good performance when accessing our file systems from parallel codes. Some of the most important guidelines include:
- Use file systems for their intended use-case; for example, don't use your home directory for production I/O (more details on intended use case may be found on the NERSC file systems page).
- Know what fraction of your wall-clock time is spent in I/O; for example, with estimates provided by Darshan, profiling of critical I/O routines (such as with Craypat's trace groups), or explicit timing / instrumentation.
- When algorithmically possible:
- Avoid workflows that produce large numbers of small files (e.g. a "file-per-process" access model at high levels of concurrency).
- Avoid random-access I/O workloads in favor of contiguous access.
- Prefer I/O workloads that perform large transfers that are similar in size or larger than, and are also aligned with, the underlying file system storage granularity (e.g. blocksize on GPFS-based file systems, stripe width on Lustre-based file systems). If using Fortran list-directed I/O, consider using runtime environment variables to increase the read buffer size (e.g.
FORT_FMT_RECLfor the Intel Fortran runtime).
- Use high-level libraries for data management and parallel I/O operations (as these will often apply optimizations in line with the above suggestions, such as MPI-IO collective buffering to improve aggregate transfer size, alignment, and contiguous access).
With these suggestions in mind, there are also file system-specific tuning parameters which may be used to enable high-performance I/O on Lustre. Details can be found on the Lustre file system tuning page for Perlmutter
Tutorials, Support, and Resource Allocation¶
Here, we list additional support resources for NERSC users, as well as pointers to previous and ongoing research projects associated with NERSC staff and LBL researchers to support high-performance scientific I/O.
Online tutorials at NERSC¶
- A detailed Introduction to Scientific I/O, discussing access patterns, Lustre, and examples using the (parallel) HDF5 library
- A brief overview of I/O Formats at in use NERSC (focused on HDF5 and NetCDF)
User support at NERSC¶
- Consulting services provided by NERSC Consultants
Previous and ongoing I/O research projects contributed to by NERSC and LBL researchers¶
- The ExaHDF5 group is working to develop next-generation I/O libraries and middleware to support scientific I/O (focused in particular on the HDF5 data format)