NERSCPowering Scientific Discovery Since 1974

Introduction to Scientific I/O

Optimizations for HDF5 on Lustre

The Lustre file system performs best when writes are aligned to stripe boundaries. This minimizes the overhead of the distributed lock manager which has to serialize access when a stripe is accessed by more than one client. Two methods to achieve stripe alignment on writes are described below.

Chunking and Alignment for Balanced 3D Grids

When every MPI task has the same amount of data to write into a 3D grid, a data layout technique called chunking can be enabled in HDF5. This layout maps the 3D block on each task to its own contiguous 1D section of the file on disk. Figure 6.1 shows an example in 2D where 4 MPI tasks each write their own 4x4 block contiguously into a file, and an index in the HDF5 metadata keeps track of where each block resides in the file. A second HDF5 tuning parameter called the alignment can be set to pad out the size of each chunk to a multiple of the stripe size. Although padding wastes space in the file on disk, the gain in write bandwidth often outweighs this, especially on HPC systems like Hopper and Edison with large scratch file systems. A more problematic restriction of the chunked layout, however, is that future parallel accesses to the file (for instance with an analysis or visualization tool) are optimized for reads that are multiples of the chunk size.

Effect of chunking on grid layout

Figure 6.1. Four tasks write to a contiguous 2D grid on the left versus a chunked grid on the right.


In our 3D code example, chunking and alignment can be enabled using the following:

fapl = H5Pcreate(H5P_FILE_ACCESS);
H5Pset_alignment(fapl, 0, stripe_size);
file = H5Fcreate("myparfile.h5", H5F_ACC_TRUNC, H5P_DEFAULT, fapl);
dcpl = H5Pcreate(H5P_DATASET_CREATE);
H5Pset_chunk(dcpl, 3, chunk_dims);
H5Dcreate(file, "mydataset", type, filespace, H5P_DEFAULT, dcpl, H5P_DEFAULT);

HDF5 uses a B-tree to index the location of chunks inside the file. The default size for this B-tree occupies only a few kilobytes in the HDF5 metadata entry. When this space is exceeded, additional B-trees are created, leading to many small metadata writes when thousands of chunks are written to a file. The small metadata writes, even when they are padded and aligned using the alignment parameter still adversely affect write performance. HDF5 provides the following mechanism for increasing the default size of the B-tree so that it is roughly the same size as a stripe:

btree_ik = (stripe_size - 4096) / 96;
fcpl = H5Pcreate(H5P_FILE_CREATE);
H5Pset_istore_k(fcpl, btree_ik);
file = H5Fcreate("myparfile.h5", H5F_ACC_TRUNC, fcpl, fapl);

HDF5 also maintains a metadata cache for open files, and if enough metadata fills the cache, it can cause an eviction that interrupts other file accesses. Usually, it is better to disable these evictions

Metadata in HDF5 files is cached by the HDF5 library to improve access times for frequently accessed items. When operating in a sequential application, individual metadata items are flushed to the file (if dirty) and evicted from the metadata cache. However, when operating in a parallel application, these operations are deferred and batched together into eviction epochs, to reduce communication and synchronization overhead. At the end of an eviction epoch (measured by the amount of dirty metadata produced), the processes in the application are synchronized and the oldest dirty metadata items are flushed to the file.

To reduce the frequency of performing small I/O operations, it is possible to put the eviction of items from the HDF5 library's metadata cache entirely under the application's control with the following:

mdc_config.version = H5AC__CURR_CACHE_CONFIG_VERSION;
H5Pget_mdc_config(file, &mdc_config)
mdc_config.evictions_enabled = FALSE;
mdc_config.incr_mode = H5C_incr__off;
mdc_config.decr_mode = H5C_decr__off;
H5Pset_mdc_config(file, &mdc_config);

This sequence of calls disables evictions from the metadata cache, unless H5Fflush is called or the file is closed. Suspending automatic eviction of cached metadata items also prevents frequently dirtied items from being written to the file repeatedly. Suspending metadata evictions may not be appropriate for all applications however, because if the application crashes before the cached metadata is written to the file, the HDF5 file will be unusable.

Collective Buffering

In the case where each task writes different amounts of data, chunking cannot be used. Instead, a different optimization called Collective buffering (or two-phase IO can be used if collective mode has been enabled, as described in Section 4. In fact, collective buffering can also be used in the balanced 3D grid case discussed above instead of chunking, and the optimization in HDF5 for disabling cache evictions also applies to collective buffering.

Collective buffering works by breaking file accesses down into to stages. For a collective read, the first stage uses a subset of MPI tasks (called aggregators) to communicate with the IO servers (OSTs in Lustre) and read a large chunk of data into a temporary buffer. In the second stage, the aggregators ship the data from the buffer to its destination among the remaining MPI tasks using point-to-point MPI calls. A collective write does the reverse, aggregating the data through MPI into buffers on the aggregator nodes, then writing from the aggregator nodes to the IO servers. The advantage of collective buffering is that fewer nodes are communicating with the IO servers, which reduces contention. In fact, Lustre prefers a one-to-one mapping of aggregator nodes to OSTs.

Since the release of mpt/3.3, Cray has included a Lustre-aware implementation of the MPI-IO collective buffering algorithm. This implementation is able to buffer data on the aggregator nodes into stripe-sized chunks so that all read and writes to the Lustre filesystem are automatically stripe aligned without requiring any padding or manual alignment from the developer. Because of the way Lustre is designed, alignment is a key factor in achieving optimal performance.

Several environment variables can be used to control the behavior of collective buffering on Hopper and Edison. The MPIIO_MPICH_HINTS variable specifies hints to the MPI-IO library that can, for instance, override the built-in heuristic and force collective buffering on (in csh):

% setenv MPIIO_MPICH_HINTS "*:romio_cb_write=enable:romio_ds_write=disable"

Placing this command in your batch file before calling aprun will cause your program to use these hints. The * indicates that the hint applies to any file opened by MPI-IO, while romio_cb_write controls collective buffering for writes and romio_ds_write controls data sieving for writes, an older collective mode optimization that is no longer used and can interfere with collective buffering. The options for these hints are enabled, disabled, or automatic (the default value, which uses the built-in heuristic).

It is also possible to control the number of aggregator nodes using the cb_nodes hint, although the MPI-IO library will automatically set this to the stripe count of your file.

When set to 1, the MPICH_MPIIO_HINTS_DISPLAY variable causes your program to dump a summary of the current MPI-IO hints to stderr each time a file is opened. This is useful for debugging and as a sanity check against spelling errors in your hints.

The MPICH_MPIIO_XSTATS variable enables profiling of the collective buffering algorithm, including binning of read/write calls and timings for the two phases. Setting it to 1 provides summary data, while setting it to 2 or 3 provides more detail.

More detail on MPICH runtime environment variables, including a full list and description of MPI-IO hints, is available from the intro_mpi man page on Hopper and Edison. Please also see our documentation on tuning Lustre I/O on both machines.