Using Burst Buffers with PnetCDF

Figure 1. Design of the Burst Buffer Driver. The green lines show the data flow for write operations when the Burst Buffer Driver is disabled. The blue lines indicate the write requests are first stored on the burst buffer. The red line shows the data flow when flushing the data from the burst buffer to the parallel file system.

Enable the Burst Buffer Driver

    ./configure --prefix=/path/to/install --enable-burst-buffering

    make

    make install

Use Burst Buffer Driver

• One is by setting MPI file info object and passing it through the call to file creating API ncmpi_create or file open API ncmpi_open.
• The other is by setting the environment variable PNETCDF_HINTS.

    MPI_Info_set(info, "nc_burst_buf", "enable");

    export PNETCDF_HINTS="nc_burst_buf=enable"

PnetCDF Hints for Controlling Burst Buffer Driver

Example Program

More example programs can be found under example/burst_buffer folder.

    #include<stdio.h>
    #include<stdlib.h>
    #include<pnetcdf.h>

    int main(int argc, char *argv[]){
        int ncid, err;
        MPI_Info info;

        MPI_Info_create(&info);

        /* Enable the Burst Buffer Driver.
        * The hint is not required if it is set in the environment variable PNETCDF_HINTS
        */
        MPI_Info_set(info, "nc_burst_buf", "enable");

        /* Create a NetCDF file with hint to enable Burst Buffer Driver*/
        ncmpi_create(MPI_COMM_WORLD, filename, NC_CLOBBER, info, &ncid);

        MPI_Info_free(&info);

        /* For doing other IO operations, the code is the same as usual
         * No actions required after file creation
         */

        /* Data stored in the burst buffer will be flushed automatically to the PFS when the file is closed */
        mpi_close(ncid);

        return 0
    }

Running Jobs Using the Burst Buffer Driver

Requesting Burst Buffer Access

Here is an example command of requesting a burst buffer space in sm_pool with stripe count at least 64.

#DW jobdw capacity=1289GiB access_mode=private type=scratch

For more information, please refer to "How to use the Burst Buffer" and "Performance Tuning" page on NERSC website.

Configuring the Parallel File System

Here's an example of setting the stripe count to 64 and stripe size to 8 MiB on current working directory that is using Lustre on Cori:

lfs setstripe -c 64 -s 8m .

For more information, please refer to "Optimizing I/O performance on the Lustre file system" page on NERSC website and "Configuring Lustre File Striping" on Lustre wiki site.

Example Job Script

    #!/bin/bash
    #SBATCH -p debug
    #SBATCH -N 1
    #SBATCH -C haswell
    #SBATCH -t 00:01:00
    #SBATCH -o burst buffer_FLASH_example.txt
    #SBATCH -L scratch
    #DW jobdw capacity=1289GiB access_mode=private
    NNodes=${SLURM_NNODES}
    NProc=NNodes*32

    export PNETCDF_HINTS="nc_burst_buf=enable;nc_burst_buf_del_on_close=disable;nc_burst_buf_dirname=${DW_JOB_PRIVATE}"

    srun -n ${NProc} ./flash_benchmark_io ${SCRATCH}/flash_

Known Issues

• Error reporting: The error codes returned from write related APIs are the errors encountered when writing to the burst buffer system. When flushing data from burst buffer to the parallel file system, the error code returned will be the first error encountered.
• Non-blocking IO: Non-blocking write requests are written to burst buffer as well. However, for perofrmance consideration, non-blocking requests may be flushed to the parallel file system before the wait API, ncmpi_wait or ncmpi_wait_all, is called. In such case, users are not able to cancel those non-blocking requests and calling ncmpi_cancel will receive error code NC_EFLUSHED.
• Unsupported API type: Currently, the Burst Buffer Driver does not support buffering on ncmpi_put_vard and ncmpi_put_vard_all. When these APIs are called, the data will be written to the parallel file system directly, bypassing the burst buffer.

Questions and Answers

Q: Can I run the Burst Buffer Driver on a different machine with a different burst buffer configuration?
A: We currently have only tested the Burst Buffer Driver on Cori at NERSC and the Cray burst buffer system installed there. It may also work on other burst buffer implementations where the burst buffer is mounted as a file system.

Q: Can I run the Burst Buffer Driver without the burst buffer?
A: Yes, as long as the hint "nc_burst_buf_dirname" is a valid directory, PnetCDF will use it as a burst buffer. However, the performance will be affected depending on the performance of the devices.

Q: Do I need to enable DataWarp module on Cori to build PnetCDF with Burst Buffer Driver support?
A: No, we currently do not use DataWarp APIs in the implementation.

Q: Can I leave the data on burst buffer without flushing it to the parallel file system?
A: No, in this version, the data will always be flushed when the file is closed. We do not recommand terminating the program without closing the file.

Q: What is the file name of intermediate files Burst Buffer Driver used to buffer I/O data? Will it overwrite my existing file if there is a name conflict?
A: The intermediate files are named as <NetCDF file name>_<ncid>_<rank>.meta and <NetCDF file name>_<ncid>_<rank>.data. To prevent overwriting other files by accident, Burst Buffer Driver will not proceed when there is a file name conflict. In such case, ncmpi_create and ncmpi_open will fail with error code NC_EEXISTS.

Q: Does the directory used by burst buffer to store I/O data need to be an empty folder?
A: No, there's no such restriction. However, we strongly recommend the use of an empty folder to prevent errors caused by file name conflict.

Q: Is there a way to trigger a flush without closing the NetCDF file?
A: Yes, calling ncmpi_sync triggers a flsuh. A flush can also be triggered if the user waits on a non-blocking write request that is still buffered on burst buffer.

Source Code Checkout

    git clone https://github.com/Parallel-NetCDF/PnetCDF.git

    autoreconf -i