PACT Benchmarks

PDB and HDF5 Compared

Gregory J. Smethells

PACT Copyright

1 Introduction

Both the portable binary database (PDB) and the hierarchical data format revision 5 (HDF5) libraries offer an interface to binary data storage and retrieval in an architecture independent manner. That is, enough meta-data is stored within the files to completely describe the format of the raw, user data such that the library can properly read back that data on any other platform. This allows these file formats to be portable and circumvents such hazards as endianess and machine-dependent data-type format differences. Other features include, associating names with the user data stored within PDB and HDF5 files so that the data can be easily referenced, the ability to access some subpart of any data set, being able to associate attributes with the data, and mechanisms to group data sets into directory type structures to provide organization. Both libraries provide a variety of APIs. The PDB library can be called from C, C++, and FORTRAN and has bindings in Scheme and Python. The HDF5 library has C, C++, Java, and FORTRAN APIs and bindings in Python, Scheme, and LabVIEW. Each library provides similar data and meta-data storage capabilities.

With respect to the user-level API, we argue that the PDB interface is simpler and has a gentler learning curve than the HDF5 APIs. More lines of code are required to setup and perform an operation such as a restart dump using HDF5 than are required to do the same operation using PDBLib. In terms of performance characteristics, our benchmarks show that the PDB and HDF5 libraries provide comparable native performance; however, when data conversions are required, HDF5 performs slightly better than PDB for similar operations. Lastly, PDB provides support for managing indirections (pointers) within data structures while HDF5 currently does not.

2 Feature Comparison

2.1 Utilities

Each file format has utilities that support basic and specialized operations on their files. PDB has pdbview, pdbdiff, and ultra. A UNIX like diff tool, pdbdiff does an intelligent difference between two files and is capable of handling any file format that it has an input spoke for including PDB and HDF5. No HDF5 tool is currently available to do a diff on HDF5 files, though one is under development. Ultra can be used for analysis and display of one-dimensional data sets and pdbview has capabilities to view, plot, analyze, and work with data in PDB and any other file format for which an input spoke is present. HDF5 is packaged with h5ls, h5repart, h5dump, among others, and has other specialized tools available. H5ls lists information about file objects in an HDF5 file. H5repart can repartition an HDF5 file or family of files. H5dump provides a means to examine and dump a HDF5 file's contents.

2.2 Data types

PDB and HDF5 both handle essentially the same set of data types. Atomic types (such as int and double) are managed through the use of meta-data to describe basic memory layout characteristics including the byte size of different primitive types, endianess, alignment, and floating point format standards so that data written on one machine can always be read on another. In addition, both file formats provide ways to manage compound data structures to handle more complicated data types by building upon primitive types or any previously defined compound data types. Lastly, arrays of fairly arbitrary dimensionality for any data type can be created and handled by both libraries. Taken as a whole, these provide a basic level of data structure capability and pass responsibility of any more complex, higher-level data structure management to the user-level code.

The main difference between the capability of these libraries to handle data types is in the use of indirections (pointers). PACT provides its own memory manager, SCORE, and leveraging its added functionality, PDB provides a mechanism to do I/O over variables containing any arbitrary level of indirection that terminates in a primitive or a structure containing no indirections. HDF5 currently has no similar internal capability, though the user-level program may provide a mechanism to recursively walk indirections in its data structures and insert the proper data into the HDF5 file by hand. Note that the user-level program would also be responsible for properly allocating memory and reading the data back into the correct data structures when re-initializing its data from the file.

2.3 Data spaces and Data sets

HDF5 breaks down a data set into a data space, a data type, and the raw, user data. The user refers to a data set by name and treats it as a single object containing both the data and meta-data. The user of HDF5 is fully aware of this collection, as a data space and data type must be created and associated with a data set along with the actual user data itself when preparing to write to a file. Most of this seems somewhat more transparent in PDB where much of the same information is collected during the write call. For instance, in the basic PD_write call, data space and type information is described in the syntax of the name and type argument strings, respectively. Dimension descriptions are encoded as in a FORTRAN dimension statement within the name. How this space and type information is managed internally is not the concern of the user. PDB sticks closely to the data model already present within programming languages, while HDF5 steps back and abstracts these into objects within its own data model.

2.4 Data Conversions and Targeting

Neither file format requires user data be converted to the data standards of an architecture before the data can be read there, rather, there is always sufficient meta-data provided within the file to do any necessary data conversions on-the-fly. Both HDF5 and PDB read operations automatically convert to the local host's native formats if user data isn't already in that form. However, targeting write operations to a destination architecture has proven to be a useful ability, as that prevents conversions from being necessary on subsequent reads of the file on a target machine where the file may be read multiple times. The more often reads are done non-natively on that file, the more important the benefit of pre-converting data becomes. PDB and HDF5 both provide mechanisms to accomplish pre-converting/targeting data, though the approach taken by each differs.

All the necessary information to describe the standards PDB requires to do conversions are conveyed through the PD_target call and type argument string of a write operation. User code passes references to the target architecture's data standards and alignment during a call to PD_target. From that point on, conversions are automatically applied to all data handled by a PDB write operation. What kinds of conversions are necessary are determined internally by the library by comparing the data standards and alignments set during the PD_target call with the data standards and alignment detected on the local host for a given type.

A similar interface is provided by the H5T API in HDF5; however, a different approach is taken to describe the format of the data in memory and describe what format the data should have when it is written out to disk. A user initializes the data types of each of their data sets with the proper type of the destination machine (say, a 64 bit, big endian, floating-point: H5T_IEEE_F64BE). This only describes half the picture, that is, the format of the data once in the file. Then, as an argument to HD5write, the user must pass a data type descriptor for the format of the data resident in main memory on the local host (say, H5T_NATIVE_DOUBLE). PDB requires the latter as well, of course; however, when targeting one can apply a global setting for the format of data on disk using PD_target. In HDF5, if the user wishes to target all writes to the file for a specific machine, the format of each data set being written out to disk must be handled individually, rather than collectively. Overall, this approach, from the point of view of the library, is quite flexible, but may be seen as more cumbersome at the user-level.

2.5 Data Organization

Both libraries provide the user with a method of hierarchically grouping variables and other data in a manner reminiscent of a file system's directory structure. PDB refers to this organization in much the same manner as one might at a UNIX command line interface, providing functions for changing directory (PD_CD), making directories (PD_MKDIR), listing members within a directory (PD_LS), and the like. HDF5 uses language that refers to directories as groups and has capabilities for creating groups (H5Gcreate) and getting a handle to an existing group (H5Gopen) at which point the user may iterate some function across its members (H5Giterate) who may be other groups or just datasets. Essentially, both libraries provide the same mechanisms. Though the iterator pattern is not seen within PDB, the PDB user could loop over the variables within a directory in their user-level code.

3 Code Comparisons

In the general case, code to write out data is more succinct when written using PDBLib. The following snippets of code show examples of using PDB and HDF5 to perform the same operations. The reason less code is required when writing PBDLib seems to be due to how much PDB hides data space and data type information within the parameters to its functions. For example, storing an array of one-hundred doubles named f64Primitive to a PDB file would be written in C as:

status = PD_write(tempFile, "f64Primitive(0:99)", "double", f64Primitive);

The same user-level data space and type information would be conveyed to HDF5 as:

dim[0] = 100;
arraySpace = H5Screate_simple(1, dim, NULL);
dataset = H5Dcreate(tempFile, "f64Primitive", H5T_NATIVE_DOUBLE, arraySpace, H5P_DEFAULT);
status = H5Dwrite(dataset, H5T_NATIVE_DOUBLE, H5S_ALL, H5S_ALL, H5P_DEFAULT, f64Primitive);

Here H5S refers to the data space API and H5D refers to the data set API. As can be seen above, the meta-data gathered by PDB is taken in during the call to write the user data. The name is associated with the user data and the type and space given by the parameters to the function call are used to construct the necessary meta-data structures on disk. The meta-data is spread out in the HDF5 case, into a data space (arraySpace), and data type (H5T_NATIVE_DOUBLE), and a data set, which groups together the user data, data space, and data type and associates a name with it. Once the information is gathered into a data set, the data set can be written out to disk.

When using the PDB API, the following code would suffice for writing a simple struct array out to disk:

PD_defstr( tempFile, "testStruct",
                 "float x",
                 "float y",
                 "float z",
                 "float t",
                 "double val",
                 LAST );

status = PD_write(tempFile, "myComp(0:99)", "testStruct", myComp);

For HDF5 this would be:

comp_t = H5Tcreate(H5T_COMPOUND, sizeof(testStruct));

H5Tinsert(comp_t, "x", HOFFSET(testStruct, x), H5T_NATIVE_FLOAT);
H5Tinsert(comp_t, "y", HOFFSET(testStruct, y), H5T_NATIVE_FLOAT);
H5Tinsert(comp_t, "z", HOFFSET(testStruct, z), H5T_NATIVE_FLOAT);
H5Tinsert(comp_t, "t", HOFFSET(testStruct, t), H5T_NATIVE_FLOAT);
H5Tinsert(comp_t, "val", HOFFSET(testStruct, val), H5T_NATIVE_DOUBLE);

dim[0] = 100;
arraySpace = H5Screate_simple(1, dim, NULL);
dataset = H5Dcreate(tempFile, "myComp", comp_t, arraySpace, H5P_DEFAULT);
status = H5Dwrite(dataset, comp_t, H5S_ALL, H5S_ALL, H5P_DEFAULT, myComp);

Each of these HDF5 examples show only native data types being written out. When the data format of the variable in memory differs from the desired format of the variable when written to file, the code grows. PDB only necessitates a call to PD_target. Assume for a moment we are targeting for an Intel cluster from a big endian system.


snprintf( alphaType, BUFFER_SIZE, "double alpha(%d)", SMALL_ARRAY_SIZE );
snprintf( betaType,  BUFFER_SIZE, "double beta(%d)",  SMALL_ARRAY_SIZE );
snprintf( gammaType, BUFFER_SIZE, "double gamma(%d)", SMALL_ARRAY_SIZE );

PD_defstr( tempFile, "arrayStruct",
               "float xCoord",
               "float yCoord",
               "float zCoord",
               "float t",
               LAST );

status = PD_write(tempFile, "theStruct(0:99)", "arrayStruct", theStruct);

And again in HDF5:

targetFloat_t  = H5Tcopy(H5T_IEEE_F32LE);
targetDouble_t = H5Tcopy(H5T_IEEE_F64LE);

compound_t   = H5Tcreate(H5T_COMPOUND, sizeof(arrayStruct));
small_dim[0] = SMALL_ARRAY_SIZE;
array_t      = H5Tarray_create(H5T_NATIVE_DOUBLE, 1, small_dim, NULL);

H5Tinsert(compound_t, "xCoord", HOFFSET(testStruct, xCoord), H5T_NATIVE_FLOAT);
H5Tinsert(compound_t, "yCoord", HOFFSET(testStruct, yCoord), H5T_NATIVE_FLOAT);
H5Tinsert(compound_t, "zCoord", HOFFSET(testStruct, zCoord), H5T_NATIVE_FLOAT);
H5Tinsert(compound_t, "t",      HOFFSET(testStruct, t),      H5T_NATIVE_FLOAT);
H5Tinsert(compound_t, "alpha",  HOFFSET(testStruct, alpha),  array_t);
H5Tinsert(compound_t, "beta",   HOFFSET(testStruct, beta),   array_t);
H5Tinsert(compound_t, "gamma",  HOFFSET(testStruct, gamma),  array_t);

compoundTarget_t = H5Tcreate(H5T_COMPOUND,
    (H5Tget_size(targetFloat_t)*4 + H5Tget_size(targetDouble_t)*SMALL_ARRAY_SIZE*3));
small_dim[0] = SMALL_ARRAY_SIZE;
arrayTarget_t = H5Tarray_create(targetDouble_t, 1, small_dim, NULL);

H5Tinsert(compoundTarget_t, "xCoord", 0,                            targetFloat_t);
H5Tinsert(compoundTarget_t, "yCoord", H5Tget_size(targetFloat_t),   targetFloat_t);
H5Tinsert(compoundTarget_t, "zCoord", H5Tget_size(targetFloat_t)*2, targetFloat_t);
H5Tinsert(compoundTarget_t, "t",      H5Tget_size(targetFloat_t)*3, targetFloat_t);
H5Tinsert(compoundTarget_t, "alpha",  H5Tget_size(targetFloat_t)*4, arrayTarget_t);
H5Tinsert(compoundTarget_t, "beta",
    (H5Tget_size(targetFloat_t)*4 + H5Tget_size(targetDouble_t)*SMALL_ARRAY_SIZE*1),
H5Tinsert(compoundTarget_t, "gamma",
    (H5Tget_size(targetFloat_t)*4 + H5Tget_size(targetDouble_t)*SMALL_ARRAY_SIZE*2),

dim[0] = 100;
arraySpace = H5Screate_simple(1, dim, NULL);
dataset = H5Dcreate(tempFile, "theStruct",
                      compoundTarget_t, arraySpace, H5P_DEFAULT);

status = H5Dwrite(dataset, compound_t,
                      H5S_ALL, H5S_ALL, H5P_DEFAULT, theStruct);

Note the use of two compound type variables. One describes the format of the struct in the local host's memory and those descriptors with the "target" modifier describe the format of data on the target machine as it should appear in the file.

In fairness to HDF5, we discuss three important issues. First, the code used in the creation and insertion into compoundTarget_t could be changed to use HOFFSET. However, in the general case, it may not be true that the size of the converted types (those variables containing "target" in their name) and the local host's native types are the same on both machines. Hence, this type of code may be necessary so members within the struct do not overlap within the file. Second, HDF5 handles conversions on-the-fly fairly quickly, hence, targeting to a particular architecture may not be necessary depending on user needs and the benefits from targeting the file would only be seen if the file was read more often than written. Third, an HDF5 Lite API is under development and being distributed separately from the basic HDF5 library. It currently provides a C API only. This higher-level interface allows the user to create and read datasets using a simpler interface. However, the API does not appear to provide a means to do hyperslab copies, nor would it simplify the number of data type descriptors necessary in the example above. The HDF5 Lite API is certainly a step in the right direction as far as simplifying user-level code, but depending on user constraints, one may still need to use the basic HDF5 API.

4 Benchmark Results

4.1 System Setup

For the benchmark tests performed, the latest version of PACT from CVS was used, as well as the latest version of HDF5 (1.4.4) available on the web. The libraries were compiled on each platform in production mode with default optimizations on and debugging off. The benchmarks that we will be describing were run on the following systems:

SystemProcessorsRAMOperating System
IBM SP(16) 375 MHz Power 316 GBAIX 5.1
Compaq ES45(4) 1 GHz EV6.8 Alpha4 GBTru64 Unix 5.1
Generic Intel(2) 2.2 GHz Pentium 4 Xeon2 GBCHAOS 1.0 (based on RH 7.3)

The benchmarks were setup with a flexible interface to allow the user to flag several options including the number of variables, data size, and what machine, if any, to target the file for. The benchmark simulated performing a restart dump for a scientific application by doing I/O over a range of variable types using both PDB and HDF5, and then timing the operation. Data types included large arrays of: int, long, float, double, simple structs, nested structs, and (if flagged) structs containing pointers. For the latter, code was written to follow the indirections on behalf of HDF5 and write out the data by hand, and similarly when reading the data back, code was written to handle allocating memory and restoring the original state of the data structure. All benchmarks were run serially, on a single node, reading and writing to the local disk. The benchmarks were not run in single-user mode on a quiet job queue, hence some timing error must be assumed due to the process waiting in the run queue to be rescheduled on the processor. Several iterations of the same restart dump simulation were performed to determine a mean duration. All data written out was compared to that read back in as a correctness check.

We classify our benchmarks into three different cases. For Native benchmarks, all data was written to disk over the course of several PDB or HDF5 calls, and then the file was closed. Several iterations of this data dump were done and a mean duration was taken. Then the same file was opened and either PDB or HDF5 calls were made to read all data back into their respective variables. Again several iterations were used during this restart to determine a mean duration. The number of variables and the data size were varied to exhibit a range of I/O conditions. Native benchmarks with indirections consisted of the same methodology as Native benchmarks with the addition of an extra variable, a struct containing pointers to other structures. Lastly, Cross-platform benchmarks were run by targeting data formats in the file to a particular architecture, writing the file to the local disk, then moving that file across the network to the target machine's local disk, where the file was finally read back in. Each of these scenarios will be discussed in turn.

4.2 Native Performance

For these graphs, the number of variables was held constant as the amount of data per variable was varied. The datasize on the x-axis represents the sum of the byte size of all user variables in the restart dump. Time on the y-axis is either the sum of the time to write the restart dump file (Dump) or the time to read back all data from the restart dump file (Restart). The first benchmark graphs shown were performed on the Intel Xeon node. In the general case, the native performance exhibited by both libraries is fairly similar. PDB appears to scale more smoothly to larger I/O sizes in these first two graphs, taking less time than HDF5, though the percent difference is fairly small and may be due to noise introduced by other jobs on the system.

Graph I -- Performed on the Intel Xeon node.
Graph II -- Performed on the Intel Xeon node.

The following graphs show native performance on the IBM SP node. The graphs again indicate very similar performance characteristics though the results show the opposite case with HDF5 scaling better. These four graphs are representative of the performance shown in other cases tried on both machines. More fine-tuned investigation would be necessary to determine how the results are affected by process scheduling noise and I/O architecture differences. The high end of the datasize scale is more sensitive to scheduling noise as many more I/O syscalls are placed and more time is spent within the system giving more opportunity for other processes to affect the timing of the restart dump simulation. Overall, the native performance appears to be fairly similar.

Graph III -- Performed on the IBM SP node.
Graph IV -- Performed on the IBM SP node.

These last graphs were performed on the Compaq ES45 node. The datasize per variable is based off of the system pagesize and datasizes have doubled due to the pagesize on Tru64 UNIX being twice that of the other systems. Here the I/O performance is again comparable assuming some error due to scheduling or other noise within the system.

Graph V -- Performed on the Compaq ES45 node.
Graph VI -- Performed on the Compaq ES45 node.

4.3 Native Performance with Indirections

These benchmarks focus on performance when an extra array of structs containing pointer members are added to the variable set. The inefficiencies that can be seen on HDF5's behalf are likely due to one of, or a combination of, two issues. The methodology of the HDF5 indirection handling code was to loop over the array of structs containing indirections, dereferencing and writing out each pointer as a separate data set. This greatly increases the number of datasets within the HDF5 file and it is feasible that HDF5 does not scale well with respect to the number of variables to be managed. The other possibility is that due to the large increase in the number of H5Dwrite calls, the overhead due to many more function calls and many, smaller writes causes the decrease in performance. More efficient code may be feasible than that used during the benchmarks, but the most straightforward code would also likely be the first approximation of user code that would be written to handle such a data structure by any given scientific application code base.

Graph VII -- Performed on the Intel Xeon node
Graph VIII -- Performed on the Compaq ES45 node

Further investigation shows that HDF5 does not scale as well as PDB when the number of data sets to be managed increases to very large values. The reasons behind the lack of efficiency for HDF5 in this case are likely due to B-tree data structures that must be maintained as more data sets are added to the file. Graph VIII is a native benchmark without any indirection variables where the datasize per variable is held constant at an eighth of the system pagesize while the number of user variables varies. This suggests the likely cause of the decreased performance seen in the native indirections benchmarks are at least partially the result of this situation. Approaches that could continue to store the data associated with the indirect structs in the previous benchmarks while maintaining fewer data sets would likely increase the efficiency of the HDF5 code, however it would remain that the user-level code would be left responsible for creating this efficient code base in order to handle any situation where the structure members were not statically allocated.

4.4 Cross-platform Performance

On the Intel Xeon node, benchmarks were run targeting the format of the restart dump data for the Power 3 on the IBM SP node. The differences of note between the two architectures is that the Intel chip is little endian and the Power 3 is big endian hence byte-order reversals will be performed. No alignment nor primitive datatype byte size differences exist.

Graph IX -- Target files to Power3 architecture
Graph X -- Create files without targeting

All restart dump files are written on the Intel Xeon system, then transfered to and read by the IBM SP node. The first graph shows the performance when the file's data formats are targeted to the IBM SP platform, while the second graph illustrates performance when no targeting is applied. Comparing the restart times of the two graphs, we see both libraries were better off taking the extra time to write a file meant for a Power 3 processor. We also see that PDB and HDF5 restart performance is comparable in Graph IX where both libraries target their files though a little extra cost is taken during the creation of the PDB file to obtain similar restart times. It appears that PDB is somewhat less efficient here and this cost must be taken at either the time of creation of the file (Graph IX) or when the file is read (Graph X), if it is the case that the file is transfered between machines of a different architecture. Otherwise, if there is no change in platform, the native benchmarks are the ones that apply.

5 Conclusions

We have discussed the features provided by both libraries and have found that, in general, they provide a very similar interface to binary data storage and retrieval. The major exception being PDB's ability to handle data structures containing indirections. HDF5 is unable to handle this situation unless the user takes the responsibility upon themselves to write the code necessary to traverse the data structures. Otherwise, available data types, attributes, file organization features, as well as conversion abilities are nearly identical.

Native performance exhibited by both codes was very comparable. The mean duration to do large restart dump simulations was similar across a large spectrum of I/O conditions. If the user's needs keep the file usage local, then the performance will be fairly similar, though if the user's needs require that the file be moved across non-homogeneous systems, then the performance will favor HDF5 slightly. Also, when scaling the number of variables to larger values it was shown that PDB was more efficient than HDF5. This may be due to the internal data structures chosen to manage the meta-data within the file.

PDB code was found to be more succinct than HDF5 for similar operations, though HDF5 Lite has taken on the task of developing a simpler interface to HDF5. The basic HDF5 interface requires the user to manipulate many more handles than PDB and perhaps some of this could be hidden away. PDB manages much of the data-type and data-space complexity by allowing the user to embed the information in the read or write call itself. PDB handles reads and writes as essentially added functionality to the operating system level read and write calls. HDF5 handles the same situation by breaking out each component of their I/O model into its own API.

The features provided by PDB and HDF5 are very similar; however, upon closer inspection we have shown that some differences do exist. The main similarities appear in the native performance and basic library features, while the main differences show themselves in cross-platform performance and the handling of data types with indirections. Overall, the functionality provided by both libraries is well-developed and either would be a good choice for a project's binary storage requirements.


For questions and comments, please contact the PACT Development Team.
Last Updated: 12/19/2005
llnl logo  
  LLNL-CODE-422942| Privacy & Legal Notice
Last modified: January 28, 2010