Modularized I/O characterization using Darshan 3.x

Introduction

Darshan is a lightweight toolkit for characterizing the I/O performance of instrumented HPC applications.

Starting with version 3.0.0, the Darshan runtime environment and log file format have been redesigned such that new “instrumentation modules” can be added without breaking existing tools. Developers are given a framework to implement arbitrary instrumentation modules, which are responsible for gathering I/O data from a specific system component (which could be from an I/O library, platform-specific data, etc.). Darshan can then manage these modules at runtime and create a valid Darshan log regardless of how many or what types of modules are used.

Overview of Darshan’s modularized architecture

The Darshan source tree is organized into two primary components:

• darshan-runtime: Darshan runtime framework necessary for instrumenting MPI applications and generating I/O characterization logs.

• darshan-util: Darshan utilities for analyzing the contents of a given Darshan I/O characterization log.

The following subsections provide detailed overviews of each of these components to give a better understanding of the architecture of the modularized version of Darshan. In Adding new instrumentation modules, we actually outline the necessary steps for integrating new instrumentation modules into Darshan.

Darshan-runtime

The primary responsibilities of the darshan-runtime component are:

• intercepting I/O functions of interest from a target application;

• extracting statistics, timing information, and other data characterizing the application’s I/O workload;

• compressing I/O characterization data and corresponding metadata;

• logging the compressed I/O characterization to file for future evaluation

The first two responsibilities are the burden of module developers, while the last two are handled automatically by Darshan.

In general, instrumentation modules are composed of:

• wrapper functions for intercepting I/O functions;

• internal functions for initializing and maintaining internal data structures and module-specific I/O characterization data;

• a set of functions for interfacing with the Darshan runtime environment

A block diagram illustrating the interaction of an example POSIX instrumentation module and the Darshan runtime environment is given below in Figure 1.

Figure 1. Darshan runtime environment

As shown in Figure 1, the Darshan runtime environment is just a library (libdarshan) which intercepts and instruments functions of interest made by an application to existing system libraries. Two primary components of this library are darshan-core and darshan-common. darshan-core is the central component which manages the initialization/shutdown of Darshan, coordinates with active instrumentation modules, and writes I/O characterization logs to disk, among other things. darshan-core intercepts MPI_Init() to initialize key internal data structures and intercepts MPI_Finalize() to initiate Darshan’s shutdown process. darshan-common simply provides module developers with functionality that is likely to be reused across modules to minimize development and maintenance costs. Instrumentation modules must utilize darshan-core to register themselves and corresponding I/O records with Darshan so they can be added to the output I/O characterization. While not shown in Figure 1, numerous modules can be registered with Darshan at any given time and Darshan is capable of correlating records between these modules.

In the next three subsections, we describe instrumentation modules, the darshan-core component, and the darshan-common component in more detail.

Instrumentation modules

The new modularized version of Darshan allows for the generation of I/O characterizations composed from numerous instrumentation modules, where an instrumentation module is simply a Darshan component responsible for capturing I/O data from some arbitrary source. For example, distinct instrumentation modules may be defined for different I/O interfaces or to gather system-specific I/O parameters from a given computing system. Each instrumentation module interfaces with the darshan-core component to coordinate its initialization and shutdown and to provide output I/O characterization data to be written to log.

In general, there are two different methods an instrumentation module can use to initialize itself: static initialization at Darshan startup time or dynamic initialization within intercepted function calls during application execution. The initialization process should initialize module-specific data structures and register the module with the darshan-core component so it is included in the output I/O characterization.

The static initialization approach is useful for modules that do not have function calls that can be intercepted and instead can just grab all I/O characterization data at Darshan startup or shutdown time. A module can be statically initialized at Darshan startup time by adding its initialization routine to the mod_static_init_fns array at the top of the lib/darshan-core.c source file.

Note

Modules may wish to add a corresponding configure option to disable the module from attempting to gather I/O data. The ability to disable a module using a configure option is especially necessary for system-specific modules which can not be built or used on other systems.

Most instrumentation modules can just bootstrap themselves within wrapper functions during normal application execution. Each of Darshan’s current I/O library instrumentation modules (POSIX, MPI-IO, stdio, HDF5, PnetCDF) follow this approach. Each wrapper function should just include logic to initialize data structures and register with darshan-core if this initialization has not already occurred. Darshan intercepts function calls of interest by inserting these wrappers at compile time for statically linked executables (e.g., using the linkers --wrap mechanism) and at runtime for dynamically linked executables (using LD_PRELOAD).

Note

Modules should not perform any I/O or communication within wrapper functions. Darshan records I/O data independently on each application process, then merges the data from all processes when the job is shutting down. This defers expensive I/O and communication operations to the shutdown process, minimizing Darshan’s impact on application I/O performance.

When the instrumented application terminates and Darshan begins its shutdown procedure, it requires a way to interface with any active modules that have data to contribute to the output I/O characterization. The following function is implemented by each module to finalize (and perhaps reorganize) module records before returning the record memory back to darshan-core to be compressed and written to file.

typedef void (*darshan_module_shutdown)(
    MPI_Comm mod_comm,
    darshan_record_id *shared_recs,
    int shared_rec_count,
    void** mod_buf,
    int* mod_buf_sz
);

This function can be used to run collective MPI operations on module data; for instance, Darshan typically tries to reduce file records which are shared across all application processes into a single data record (more details on the shared file reduction mechanism are given in Adding new instrumentation modules). This function also serves as a final opportunity for modules to cleanup and free any allocated data structures, etc.

• mod_comm is the MPI communicator to use for collective communication

• shared_recs is a list of Darshan record identifiers that are shared across all application processes

• shared_rec_count is the size of the shared record list

• mod_buf is a pointer to the buffer address of the module’s contiguous set of data records

• mod_buf_sz is a pointer to a variable storing the aggregate size of the module’s records. On input, the pointed to value indicates the aggregate size of the module’s registered records; on output, the value may be updated if, for instance, certain records are discarded

darshan-core

Within darshan-runtime, the darshan-core component manages the initialization and shutdown of the Darshan environment, provides an interface for modules to register themselves and their data records with Darshan, and manages the compressing and the writing of the resultant I/O characterization. As illustrated in Figure 1, the darshan-core runtime environment intercepts MPI_Init and MPI_Finalize routines to initialize and shutdown the Darshan runtime environment, respectively.

Each of the functions provided by darshan-core to interface with instrumentation modules are described in detail below.

void darshan_core_register_module(
    darshan_module_id mod_id,
    darshan_module_shutdown mod_shutdown_func,
    int *mod_mem_limit,
    int *rank,
    int *sys_mem_alignment);

The darshan_core_register_module function registers Darshan instrumentation modules with the darshan-core runtime environment. This function needs to be called once for any module that will contribute data to Darshan’s final I/O characterization.

• mod_id is a unique identifier for the given module, which is defined in the Darshan log format header file (darshan-log-format.h).

• mod_shutdown_func is the function pointer to the module shutdown function described in the previous section.

• inout_mod_buf_size is an input/output argument that stores the amount of module memory being requested when calling the function and the amount of memory actually reserved by darshan-core when returning.

• rank is a pointer to an integer to store the calling process’s application MPI rank in. NULL may be passed in to ignore this value.

• sys_mem_alignment is a pointer to an integer which will store the system memory alignment value Darshan was configured with. NULL may be passed in to ignore this value.

void darshan_core_unregister_module(darshan_module_id mod_id);

The darshan_core_unregister_module function disassociates the given module from the darshan-core runtime. Consequentially, Darshan does not interface with the given module at shutdown time and will not log any I/O data from the module. This function should only be used if a module registers itself with darshan-core but later decides it does not want to contribute any I/O data. Note that, in the current implementation, Darshan does not have the ability to reclaim the record memory allocated to the calling module to assign to other modules.

• mod_id is the unique identifier for the module being unregistered.

darshan_record_id darshan_core_gen_record_id(const char *name);

The darshan_core_gen_record_id function simply generates a unique record identifier for a given record name. This function is generally called to convert a name string to a unique record identifier that is needed to register a data record with darshan-core. The generation of IDs is consistent, such that modules which reference records with the same names will store these records using the same unique IDs, simplifying the correlation of these records for analysis.

• name is the name of the corresponding data record (often times this is just a file name).

void *darshan_core_register_record(
    darshan_record_id rec_id,
    const char *name,
    darshan_module_id mod_id,
    int rec_len,
    int *fs_info);

The darshan_core_register_record function registers a data record with the darshan-core runtime, allocating memory for the record so that it is persisted in the output log file. This record could reference a POSIX file or perhaps an object identifier for an object storage system, for instance. This function should only be called once for each record being tracked by a module to avoid duplicating record memory. This function returns the address which the record should be stored at or NULL if there is insufficient memory for storing the record.

• rec_id is a unique integer identifier for this record (generally generated using the darshan_core_gen_record_id function).

• name is the string name of the data record, which could be a file path, object ID, etc. If given, darshan-core will associate the given name with the record identifier and store this mapping in the log file so it can be retrieved for analysis. NULL may be passed in to generate an anonymous (unnamed) record.

• mod_id is the identifier for the module attempting to register this record.

• rec_len is the length of the record.

• fs_info is a pointer to a structure of relevant info for the file system associated with the given record – this structure is defined in the darshan.h header. Note that this functionality only works for record names that are absolute file paths, since we determine the file system by matching the file path to the list of mount points Darshan is aware of. NULL may be passed in to ignore this value.

double darshan_core_wtime(void);

The darshan_core_wtime function simply returns a floating point number of seconds since Darshan was initialized. This functionality can be used to time the duration of application I/O calls or to store timestamps of when functions of interest were called.

double darshan_core_excluded_path(const char *path);

The darshan_core_excluded_path function checks to see if a given file path is in Darshan’s list of excluded file paths (i.e., paths that we don’t instrument I/O to/from, such as /etc, /dev, /usr, etc.).

• path is the absolute file path we are checking.

darshan-common

darshan-common is a utility component of darshan-runtime, providing module developers with general functions that are likely to be reused across multiple modules. These functions are distinct from darshan-core functions since they do not require access to internal Darshan state.

char* darshan_clean_file_path(const char* path);

The darshan_clean_file_path function just cleans up the input path string, converting relative paths to absolute paths and suppressing any potential noise within the string. The address of the new string is returned and should be freed by the user.

• path_ is the input path string to be cleaned up.

darshan-common also currently includes functions for maintaining counters that store common I/O values (such as common I/O access sizes or strides used by an application), as well as functions for calculating the variance of a given counter across all processes. As more modules are contributed, it is likely that more functionality can be refactored out of module implementations and maintained in darshan-common, facilitating code reuse and simplifying maintenance.

Darshan-util

The darshan-util component is composed of a helper library for accessing log file data records (libdarshan-util) and a set of utilities that use this library to analyze application I/O behavior. libdarhan-util includes a generic interface (darshan-logutils) for retrieving specific components of a given log file. Specifically, this interface allows utilities to retrieve a log’s header metadata, job details, record ID to name mapping, and any module-specific data contained within the log.

libdarshan-util additionally includes the definition of a generic module interface (darshan-mod-logutils) that may be implemented by modules to provide a consistent way for Darshan utilities to interact with module data stored in log files. This interface is necessary since each module has records of varying size and format, so module-specific code is needed to interact with the records in a generic manner. This interface is used by the darshan-parser utility, for instance, to extract data records from all modules contained in a log file and to print these records in a consistent format that is amenable to further analysis by other tools.

darshan-logutils

Here we define each function in the darshan-logutils interface, which can be used to create new log utilities and to implement module-specific interfaces into log files.

darshan_fd darshan_log_open(const char *name);

Opens Darshan log file stored at path name. The log file must already exist and is opened for reading only. As part of the open routine, the log file header is read to set internal file descriptor data structures. Returns a Darshan file descriptor on success or NULL on error.

darshan_fd darshan_log_create(const char *name, enum darshan_comp_type comp_type, int partial_flag);

Creates a new darshan log file for writing only at path name. comp_type denotes the underlying compression type used on the log file (currently either libz or bzip2) and partial_flag denotes whether the log is storing partial data (that is, all possible application file records were not tracked by darshan). Returns a Darshan file descriptor on success or NULL on error.

int darshan_log_get_job(darshan_fd fd, struct darshan_job *job);
int darshan_log_put_job(darshan_fd fd, struct darshan_job *job);

Reads/writes job structure from/to the log file referenced by descriptor fd. The darshan_job structure is defined in darshan-log-format.h. Returns 0 on success, -1 on failure.

int darshan_log_get_exe(darshan_fd fd, char *buf);
int darshan_log_put_exe(darshan_fd fd, char *buf);

Reads/writes the corresponding executable string (exe name and command line arguments) from/to the Darshan log referenced by fd. Returns 0 on success, -1 on failure.

int darshan_log_get_mounts(darshan_fd fd, char*** mnt_pts, char*** fs_types, int* count);
int darshan_log_put_mounts(darshan_fd fd, char** mnt_pts, char** fs_types, int count);

Reads/writes mounted file system information for the Darshan log referenced by fd. mnt_pnts points to an array of strings storing mount points, fs_types points to an array of strings storing file system types (e.g., ext4, nfs, etc.), and count points to an integer storing the total number of mounted file systems recorded by Darshan. Returns 0 on success, -1 on failure.

int darshan_log_get_namehash(darshan_fd fd, struct darshan_name_record_ref **hash);
int darshan_log_put_namehash(darshan_fd fd, struct darshan_name_record_ref *hash);

Reads/writes the hash table of Darshan record identifiers to full names for all records contained in the Darshan log referenced by fd. hash is a pointer to the hash table (of type struct darshan_name_record_ref *), which should be initialized to NULL for reading. This hash table is defined by the uthash hash table implementation and includes corresponding macros for searching, iterating, and deleting records from the hash. For detailed documentation on using this hash table, consult uthash documentation in darshan-util/uthash-1.9.2/doc/txt/userguide.txt. The darshan-parser utility (for parsing module information out of a Darshan log) provides an example of how this hash table may be used. Returns 0 on success, -1 on failure.

int darshan_log_get_mod(darshan_fd fd, darshan_module_id mod_id, void *mod_buf, int mod_buf_sz);
int darshan_log_put_mod(darshan_fd fd, darshan_module_id mod_id, void *mod_buf, int mod_buf_sz, int ver);

Reads/writes a chunk of (uncompressed) module data for the module identified by mod_id from/to the Darshan log referenced by fd. mod_buf is the buffer to read data into or write data from, and mod_buf_sz is the corresponding size of the buffer. The darshan_log_getmod routine can be repeatedly called to retrieve chunks of uncompressed data from a specific module region of the log file given by fd. The darshan_log_putmod routine just continually appends data to a specific module region in the log file given by fd and accepts an additional ver parameter indicating the version number for the module data records being written. These functions return the number of bytes read/written on success, -1 on failure.

Note

Darshan use a “reader makes right” conversion strategy to rectify Endianness issues between the machine a log was generated on and a machine analyzing the log. Accordingly, module-specific log utility functions will need to check the swap_flag variable of the Darshan file descriptor to determine if byte swapping is necessary. 32-bit and 64-bit byte swapping macros (DARSHAN_BSWAP32/DARSHAN_BSWAP64) are provided in darshan-logutils.h.

void darshan_log_close(darshan_fd fd);

Close Darshan file descriptor fd. This routine must be called for newly created log files, as it flushes pending writes and writes a corresponding log file header before closing.

Note

For newly created Darshan log files, care must be taken to write log file data in the correct order, since the log file write routines basically are appending data to the log file. The correct order for writing all log file data to file is: (1) job data, (2) exe string, (3) mount data, (4) record id -> file name map, (5) each module’s data, in increasing order of module identifiers.

darshan-mod-logutils

The darshan-mod-logutils interface provides a convenient way to implement new log functionality across all Darshan instrumentation modules, which can potentially greatly simplify the development of new Darshan log utilities. These functions are defined in the darshan_mod_logutil_funcs structure in darshan-logutils.h – instrumentation modules simply provide their own implementation of each function, then utilities can leverage this functionality using the mod_logutils array defined in darshan-logutils.c. A description of some of the currently implemented functions are provided below.

int log_get_record(darshan_fd fd, void **buf);
int log_put_record(darshan_fd fd, void *buf);

Reads/writes the module record stored in buf to the log referenced by fd. Notice that a size parameter is not needed since the utilities calling this interface will likely not know the record size – the module-specific log utility code can determine the corresponding size before reading/writing the record from/to file.

Note

log_get_record takes a pointer to a buffer address rather than just the buffer address. If the pointed to address is equal to NULL, then record memory should be allocated instead. This functionality helps optimize memory usage, since utilities often don’t know the size of records being accessed but still must provide a buffer to read them into.

void log_print_record(void *rec, char *name, char *mnt_pt, char *fs_type);

Prints all data associated with the record pointed to by rec. name holds the corresponding name string for this record. mnt_pt and fs_type hold the corresponding mount point path and file system type strings associated with the record (only valid for records with names that are absolute file paths).

void log_print_description(int ver);

Prints a description of the data stored within records for this module (with version number ver).

Adding new instrumentation modules

In this section we outline each step necessary for adding a module to Darshan. To assist module developers, we have provided the example “NULL” module as part of the Darshan source tree (darshan-null-log-format.h, darshan-runtime/lib/darshan-null.c, and darshan-util/darshan-null-logutils.*) This example can be used as a minimal stubbed out module implementation that is heavily annotated to further clarify how modules interact with Darshan and to provide best practices to future module developers. For full-fledged module implementation examples, developers are encouraged to examine the POSIX and MPI-IO modules.

Log format headers

The following modifications to Darshan log format headers are required for defining the module’s record structure:

• Add a module identifier to the DARSHAN_MODULE_IDS macro at the top of the darshan-log-format.h header. In this macro, the first field is a corresponding enum value that can be used to identify the module, the second field is a string name for the module, the third field is the current version number of the given module’s log format, and the fourth field is a corresponding pointer to a Darshan log utility implementation for this module (which can be set to NULL until the module has its own log utility implementation).

• Add a top-level header that defines an I/O data record structure for the module. Consider the “NULL” module and POSIX module log format headers for examples (darshan-null-log-format.h and darshan-posix-log-format.h, respectively).

These log format headers are defined at the top level of the Darshan source tree, since both the darshan-runtime and darshan-util repositories depend on their definitions.

Darshan-runtime

Build modifications

The following modifications to the darshan-runtime build system are necessary to integrate new instrumentation modules:

• Necessary linker flags for inserting this module’s wrapper functions need to be added to a module-specific file which is used when linking applications with Darshan. For an example, consider darshan-runtime/share/ld-opts/darshan-posix-ld-opts, the required linker options for the POSIX module. The base linker options file for Darshan (darshan-runtime/share/ld-opts/darshan-base-ld-opts.in) must also be updated to point to the new module-specific linker options file.

• Targets must be added to Makefile.in to build static and shared objects for the module’s source files, which will be stored in the darshan-runtime/lib/ directory. The prerequisites to building static and dynamic versions of libdarshan must be updated to include these objects, as well.

  • If the module defines a linker options file, a rule must also be added to install this file with libdarshan.

Instrumentation module implementation

In addtion to the development notes from above and the exemplar “NULL” and POSIX modules, we provide the following notes to assist module developers:

• Modules only need to include the darshan.h header to interface with darshan-core.

• The file record identifier given when registering a record with darshan-core should be used to store the record structure in a hash table or some other structure.

  • Subsequent calls that need to modify this record can then use the corresponding record identifier to lookup the record in this local hash table.

  • It may be necessary to maintain a separate hash table for other handles which the module may use to refer to a given record. For instance, the POSIX module may need to look up a file record based on a given file descriptor, rather than a path name.

Darshan-util

Build modifications

The following modifications to the darshan-util build system are necessary to integrate new instrumentation modules:

• Update Makefile.in with new targets necessary for building module-specific logutil source.

  • Make sure to add the module’s logutil implementation objects as a prerequisite for building libdarshan-util.

  • Make sure to update all, clean, and install rules to reference updates.

Module-specific logutils and utilities

For a straightforward reference implementation of module-specific log utility functions, consider the implementations for the NULL module (darshan-util/darshan-null-logutils.*) and the POSIX module (darshan-util/darshan-posix-logutils.*). These module-specific log utility implementations are built on top of the darshan_log_getmod() and darshan_log_putmod() functions, and are used to read/write complete module records from/to file.

Also, consider the darshan-parser source code for an example of a utility which can leverage libdarshan-util for analyzing the contents of a Darshan I/O characterization log with data from arbitrary instrumentation modules.

Shared record reductions

Since Darshan prefers to aggregate data records which are shared across all processes into a single data record, module developers should consider implementing this functionality eventually, though it is not strictly required.

Module developers should implement the shared record reduction mechanism within the module’s darshan_module_shutdown() function, as it provides an MPI communicator for the module to use for collective communication and a list of record identifiers which are shared globally by the module (as described in Darshan-runtime).

In general, implementing a shared record reduction involves the following steps:

• reorganizing shared records into a contiguous region in the buffer of module records

• allocating a record buffer to store the reduction output on application rank 0

• creating an MPI reduction operation using the MPI_Op_create() function (see more in MPI_Op_create manpage).

• reducing all shared records using the created MPI reduction operation and the send and receive buffers described above

For a more in-depth example of how to use the shared record reduction mechanism, consider the implementations of this in the POSIX or MPI-IO modules.

Other resources

• Darshan GitLab page

• Darshan project website

• Darshan Runtime

• Darshan Utilities