STXXL
1.4.1
|
Version 1.4.1 (29 October 2014) * support kernel based asynchronous I/O on Linux (new file type "linuxaio"), which exploits Native Command Queuing (NCQ) if available. disable/enable with the define STXXL_FILE_LINUXAIO 0/1 via cmake * adding new disk_config entry device_id, which specifies the physical device id of the "disk" used during prefetching sequence calculations. This used to be identical with the queue id, however, for linuxaio there is only one queue; thus the distinction had to be made. In a default config, no changes are necessary, as the device_id parameter is automatically enumerated. * adding stxxl::binary_buffer which can be used for compact serialization and reading via a stxxl::binary_reader cursor interface. * stxxl::unordered_map is a hash map, backed by external memory. It probably only works well when lots of internal memory is used to buffer access to a relatively small working set. Then, however, fast direct in-memory item access can be used. * stxxl::external_shared_ptr is a proxy class to allow use of shared_ptr classes inside stxxl containers. Reference counts are kept in memory, while data may be swapped out to disk. * removing struct default_completion_handler, using a NULL pointer in default complete handler instead, since otherwise a default_completion_handler objects is cloned for _each_io_request_! Using a NULL pointer avoids superfluous new/delete work on the heap. * minor changes: - disable TPIE benchmarks by default, removing a warning. - compilation and tests work under MinGW 64-bit with special threads. - fixed compilation on 32-bit systems, everything is -Wconversion safe. - adding 32-bit and 64-bit cmdline_parser::add_bytes() variants. - removing all double underscores. - use atomic_counted_object in class file for request reference counting. - adding local/test2.cpp containing a stxxl::sorter example. Version 1.4.0 (12 December 2013) * Reorganized Directory Hierarchy - separating old component based directories into lib/ (libstxxl source files), tests/ (for tests cases), examples/ (for tutorial and example code) and tools/ (for stxxl_tool and more benchmarks). * CMake build system for cross-platform compilation - replaced adhoc Makefiles with CMake build system - CMake scripts can also output MSVC project files and more. - all component tests must register with CMake's test system -> run make test - generates pkg-config and CMake's project config files * Greatly Improved Doxygen Documentation - integrated old LaTeX tutorial - newly written tutorials for almost all STXXL containers and algorithms - embedded design documentation from Roman Dementiev's PhD thesis - new compilation instructions for CMake build system * Improved .stxxl Config Options for File I/O Implementations - looking for .stxxl in more places than just the current directory (see the docs for which places exactly). - parse file size with SI and IEC units, like "10GiB" in disk= config lines. - parse additional file I/O implementation options. - support for "unlink_on_open" (keep file descriptor but unlink file) and "delete_at_exit" options. - support for "direct=off/try/on" to turn off O_DIRECT for tmpfs and similar. - the default open mode is now direct=try -> prints a warning if O_DIRECT is unavailable on the system. * .stxxl Config Options Represented as disk_config Class - all disk= config lines are parsed into disk_config objects. - allow API calls to configure disks via stxxl::config::add_disk() - delay config file load till config::initialize() is called by block_manager * Command line parser stxxl::cmdline_parser - stand-alone command line parser with ability to parse SI and IEC units for "bytes sizes". - automatically output nicely formatted help for all options and flags. - using the cmdline_parser for all stxxl_tool subtools. * Created stxxl_tool from multiple small tools and benchmarks - stxxl_tool is a collection of subtools, previous scattered around in the directory hierarchy. - goal is to have one tool installed in binary distribution together with libstxxl library and includes. - primary tool for standardized platform speed tests. - provide good command line help on these subtools via cmdline_parser. * Adding skew3 as "real-world" STXXL application to examples subdirectory. - enhanced integrated stxxl::tuple construction for this application. * Enable use of raw block devices on Linux - these do provide more stable I/O throughput than even XFS. - disable growing of these devices * Most platform configuration is now done by CMake and saved into the <stxxl/bits/config.h> header file * Enhanced stxxl::block_manager to count maximum, current and total number of bytes allocated in external memory. * Support Visual Studio 2012 and 2013 _without_ Boost by using std::thread - VS 2012 and 2013 support STL threads, mutexes, condition_variables and more, which are basically modeled after the Boost libraries. Thus we can simply uses these included libraries instead of Boost. - Fixed bad runtime deadlock bug occurring when joining std::threads after main() exits. The bug persists in VS 2012 and 2013. * Implementing counting_ptr reference counted objects - used like boost::intrusive_ptr with a base class. - reimplementing request_ptr using just counting_ptr<request>. - (optionally) use atomic operations for thread-safe reference counting - applying counting_ptr to stream::sorted_runs to avoid extra copying * New STXXL container "stxxl::sequence" - basically a deque container without random access to blocks. - efficient {push/pop}_{front/back} at both ends. - access to block sequence via stream and reverse_stream constructions * New STXXL container "stxxl::sorter" - combination of stream::runs_creator and stream::runs_merger into a convenient sorting container. * Added vector_bufreader and vector_bufwriter to overlapping I/O and computation when accessing a stxxl::vector in scanning mode. - documented problems of scanning in a vector - C++11 range loop access support to stxxl::vector via vector_bufreader. * Added uint40 and uint48 types for integers with five and six bytes. * Adding STXXL_CHECK macros to always check outcomes in tests. * Renaming mng/mng.{h,cpp} -> mng/block_manager.{h,cpp}. * Reworking version.h - add integer version like 10400 for 1.4.0 - add get_version_string() for longer version string with features - check_library_version() for mismatch between headers and libraries * Reworking error_handling.h to throw much more useful exceptions with errno information and more. * Various file renames (which are only possible at major version changes). * Fixing bug in shrinking vector with vector::resize(n,true). ------------------------------------------ Version 1.3.2 (unreleased) * New Features - Add defines STXXL_VERSION_{MAJOR,MINOR,PATCHLEVEL} and library routines stxxl::version_{major,minor,patchlevel}() for getting STXXL version information and an inline function check_library_version() to check at runtime for mismatching header and library versions. (Bz:#6) Available via <stxxl/version.h> * Merged matrix branch which allows external memory matrix operations. - merged dsr::argument_helper for command line processing - also required an internal memory addressable_priority_queue, a shared_object_pointer (instrusive pointer) - adds block_scheduler as submodule to prefetch matrix blocks in a pre-recorded order - contains a test and benchmark. * Fixed bug in stxxl::priority_queue - added dump_sizes() - The priority queue may become fragmented after a significant number of insertions and deletes so that all slots in all groups are being used, but the last external group is nearly empty (counting elements). In that case we merge the whole last external group into a temporary ext_merger and swap this overflow merger and the (now empty) last external merger afterwards. Increasing the number of groups beyond 4 is not possible without significant changes elsewhere. * Fixed bug in stxxl::queue - added special case when first block has space at the beginning - added another test N * (push,pop,push), N * (pop, push, pop) * Using LIKELY and UNLIKEY to prioritize branchs in gcc. * Platform Support - add GCC 4.6 - add clang++ 2.9 - update ICPC 12.0 (Intel C++ Composer XE 2011 Update 4): drop workarounds for bugs fixed in Update 4 - Boost 1.46: Boost Filesystem Version 3 ------------------------------------------ Version 1.3.1 (10 March 2011) * Possibly breaking changes - No longer open syscall files with O_SYNC|O_DSYNC|O_RSYNC by default when doing direct I/O, to avoid a write performance decrease starting with Linux kernel version 2.6.32. A stxxl::file::SYNC flag has been added to allow forcing the O_*SYNC flags when opening files. - Lock files by default (if implemented for this file type), can by disabled via stxxl::file::NO_LOCK. - block_prefetcher now takes a completion_handler instead of a raw function pointer. Furthermore, the completion_handler is already executed after the read operation completes, not only when the issuer waits for it to complete. The I/O thread issues the call to the completion_handler, so for multiple disks, this may happen concurrently. * Internal changes, not user-visible - Change handling of (empty) filler elements in typed_block to fix mismatching struct size in nested blocks. - Removed debugmon which was very limited and disabled for a long time. * Bugfixes - Allow prefetch schedule computation of vectors that are bound to a file. - Fix race growing files during concurrent block allocation. - Allow reading a block that spans over end-of-file, fill remainder with zeroes. - Crash at program termination when using global stxxl containers. - Enable syscall_file to read/write >=2GiB of data in a single I/O operation. * New public headers: stxxl/stats, stxxl/request * Parallel mode can be switched on for internal STXXL computation selectively (i.e. without enabling it for the non-stxxl part of an application), by setting USE_PARALLEL_MODE=yes in make.settings.local. * Platform Support - add ICPC 12.0, works with both MCSTL (needs libstdc++ from GCC 4.2) and parallel mode (needs libstdc++ from GCC 4.3 (not 4.4/4.5)) - add CLANG++ 2.8 - add MSVC 2010/Windows 7 - allow the library name generated by MSVC to be changed by setting LIBNAME and redefining STXXL_LIBNAME (defaults to "stxxl") * Under-the-hood improvements - code cleanup - I/O-layer: renaming and reorganization of the request implementation * Documentation updates * Several new test programs * Several test programs improved ------------------------------------------ Version 1.3.0 (12 August 2010) * New file types - fileperblock_syscall/fileperblock_mmap/fileperblock_boostfd/fileperblock_wincall: Use a separate file for each block, which is accessed by means of the underlying file type. The "filename" given is used as a prefix of the block file names. The advantage of these file types is that unused disk space is freed earlier on the file system level. - wbtl_file: Do library-based write-combining (good for writing small blocks onto SSDs) * I/O subsystem - separate the disk number of a file (which queue to put requests in) from the fact that blocks for this file are allocated via the block manager (disk number -1 otherwise). - separate wait time counting for read and write I/Os - wait times can be logged to a separate log file (compile with -DSTXXL_WAIT_LOG_ENABLED and set STXXLWAITLOGFILE in the environment) * Parallel PQ - priority_queue now supports parallelized operations utilizing the libstdc++ parallel mode * Other new functionality - file requests can now be canceled (success not guaranteed), completion handlers are called anyway. - log file locations are now configurable by the environment variables STXXLLOGFILE and STXXLERRLOGFILE - single queue I/O scheduling - reverse_iterator added to containers types vector, deque and map - autogrow files (specified in .stxxl by a size of 0), will be deleted on normal program termination - add infrastructure to build multiple binaries with different CXXFLAGS from a single source - overwriting deleted memfile regions with uninitialized data is now optional (STXXL_CLEAR_FREE_MEMFILE_MEM) - read_write_pool that combines prefetch_pool with write_pool and ensures cache coherency - add a replaceable exit handler implementation, can be overwritten e.g. for library inclusion * Many, many bug fixes, in particular concerning - priority queue - optimal prefetch schedule - possible race condition while creating singleton instances - random_shuffle() * Platform Support - add GCC 4.4 (parallel mode features not working in GCC 4.3 can now be used) - add GCC 4.5 - initial support for GCC 4.6 - add ICPC 11.x - add Boost 1.42 - add FreeBSD - drop Solaris * New benchmarks - io/benchmark_disks: more command line paramaters for finer testing granularity - measure hard-disk and flash-disk combined performance and determine best block size pairs (io/benchmark_disk_and_flash) - benchmark using regular STXXL configuration (benchmark_configured_disks) * Possibly breaking changes - API name changes vector::touch -> vector::block_externally_updated - Template parameter changes stream::sorted_runs, trigger_entry, trigger_entry_cmp, trigger_entry_iterator - Priority of write changes - Other name changes (considered internal) typed_block<>::has_filler -> !typed_block<>::has_only_data file::delete_region -> file::discard vector::alloc_strategy -> vector::alloc_strategy_type - stxxl::sort(..., M) and stxxl::stream::sort(..., M) now adhere to to the memory limit M more strictly and throw errors instead of overallocating - Execute completion handler before notifying waiters, so far after. * Deprecated methods: - stxxl::stats::reset(), stxxl::stats::_reset_io_wait_time(): use stxxl::stats_data instead to store snapshots of the counters and compute differences * Under-the-hood improvements - I/O layer cleanup - aligned_alloc * Doxy now also lists undocumented members * stop requiring STXXL_ROOT to be a hard coded absolute path set in make.settings.local, a default of CURDIR works fine * document #defines in defines.h - switch on/off file types with STXXL_HAVE_xxx_FILE (to avoid platform compatibility problems) ------------------------------------------ Version 1.2.1 (14 August 2008) * add support for the libstdc++ parallel_mode (successor of MCSTL), new make targets: library_g++_pmode, tests_g++_pmode, clean_g++_pmode (requires g++ 4.3.2 or later) * new stxxl file type stxxl::mem_file (for debugging purposes only), helps debugging with valgrind because no memory cell ever leaves valgrind's control * properly destroy the singleton objects (block manager, disk queue threads, logger, stats, config, ...) at program exit * fixed a bug (recursive locking) in recursive block allocation * added a test program for recursive block allocation * sensible dependencies for nmake (MSVC): only rebuild files as necessary * improve performance of createdisks * human-readable output for I/O stats * replace hard coded min/max values by std::numeric_limits<>::min/max in examples * fix a case where only half of the available memory was used during recursive merging * verify stxxl::set_seed() functionality in tests * remove stxxl disk files created from default configuration (no .stxxl file found) and other temporary files at program end * stop using deprecated functions, switch to the replacements instead * unify the mutex lock handling by introducing stxxl::scoped_mutex_lock * unify the I/O wait time counting to work like read/write time counting * simplify I/O time counting with scoped_{read,write,wait}_timer * add some more tests * more code cleanup + reformatting * move some more code to the library * reduce some include dependencies * build system tuning * propagate region deletion (when blocks are freed) to the file level * fix problem in sorters where blocks were released too early * specialize is_sorted() to use const_vector_iterators, no extra writes * add c++0x style const_iterator cbegin()/cend() to all containers ------------------------------------------ Version 1.2.0 (05 July 2008) * made the block management layer thread-safe * made all size_types unsigned * stxxl::priority_queue - fixed a rare assertion - fixed a race condition by using proper block hinting - insert_queue: replaced std::priority_queue with a special implementation internal_priority_queue that allows for fast exporting of all elements - even more bugs and inefficiencies fixed - significant speed improvements * random number generators are now all seedable, should allow redoing identical program runs for debugging purposes * stxxl::noncopyable, inspired by boost::noncopyable: inheriting from this class forbids copying of objects when this is undesirable - used in all classes that had implemented their own variants previously * stxxl::vector, all sorting functions: replaced two2one_dim_array_row_adapter with array_of_sequences_iterator which is much faster, especially if blocks have padding * if required, verify that the sentinels satisfy a strict weak ordering * stxxl::vector: several operations sped up, several more implemented * fix existing allocation strategies and add experimental support for distinguishing between regular disks and flash devices * stxxl::stable_ksort - add experimental warning, some parts are not yet implemented - fixed an off-by-one error in memory allocation vs. usage * btree: fuse last two nodes/leaves if possible, rebalancing can fail * btree tests: ensure uniqueness of test data if required * reduce function call overhead of stxxl::debugmon if it's not activated * add public interface headers: stxxl/types, stxxl/aligned_alloc * add compatibility wrappers for standard extensions hash_map, hash_set, auto_ptr * MCSTL is only supported with g++ 4.2 and icpc 10.x * lots of small bugfixes * made the TPIE, BDB and LEDA_SM tests compile again * general code cleanup - fixed most compiler warnings - elimination of duplicate and unused code - cleaned up and sanitized debugging macros - no more 'using namespace std' and 'using namespace stxxl' - fixed ambiguities noted by g++ 4.3 - unify the #include directives - add/unify/cleanup copyright headers * general code reformatting (uncrustify) * add support for new compiler releases * portability fixes for different platforms * implemented file truncation on windows platforms * build system - lots of small modifications - now only requires GNU make 3.78 or later - check whether STXXL_ROOT has been set correctly and if unset, try autoconfiguration by creating make.settings.local with defaults - improved and simplified boost support for posix systems - Mac OS X support * changed all tests so that they behave well in their default parameters, system assumptions and return values and can be run from a script - use aligned_alloc/aligned_dealloc appropriately * added some more test programs * add misc/run_all_tests that runs all tests with sensible parameters, optionally via valgrind * checked all tests with valgrind - fixed use of uninitialized memory when writing to normal files - (optionally) initialize typed_blocks and payload data in test structs to suppress (most) uninitialized write errors when writing to disk files - fix mismatched new/delete in mcstl * update install and usage instructions * spell checked sources and documentation ------------------------------------------ Version 1.1.0 (31 July 2007) * stxxl is now hosted on SourceForge: http://stxxl.sourceforge.net/ * Restructured the source layout: - includes moves to include/ - introduced some public headers: stxxl.h, stxxl/algorithm, stxxl/bits, stxxl/deque, stxxl/io, stxxl/ksort, stxxl/mallocstats, stxxl/map, stxxl/mng, stxxl/priority_queue, stxxl/queue, stxxl/random, stxxl/scan, stxxl/sort, stxxl/stable_ksort, stxxl/stack, stxxl/stream, stxxl/timer, stxxl/vector - the header "stxxl" is no longer available, please use "stxxl.h" instead - the use of any other (internal) header is discouraged, additional public headers can be added as the need arises * Overhauled the build system: - merged configuration files, simplified option tuning - support object files and binaries with and without mcstl support existing in parallel - the library build creates stxxl.mk which can be included in an applications Makefile to set the correct compiler/linker switches for stxxl - similarly mcstxxl.mk is created for a stxxl compiled with mcstl support - add dependency tracking and improve parallelism during build * compiler support matrix: (for an up-to-date list, please see the doxygen documentation) compiler | stxxl stxxl + mcstl --------------+------------------------ GCC 4.3 | x x GCC 4.2 | x x GCC 4.1 | x n/a GCC 4.0 | x n/a GCC 3.4 | x n/a GCC 3.3 | o n/a GCC 2.95 | - n/a ICPC 9.1.051 | x x¹ ICPC 10.0.025 | x x¹ MSVC 2005 8.0 | x n/a x = full support o = partial support - = unsupported n/a = compiler does not support OpenMP which is needed by MCSTL ¹ = does not work with STL GCC 4.2.0 (ICPC bug), workaround: the first include in the program must be "stxxl/bits/common/intel_compatibility.h" * pipelined stream::sort, stream::runs_creator and stream::runs_merger are parallelized using MCSTL * obsolete files removed * fixed include dependencies * fixed lots of warnings * fixed lots of small bugs * add header compilation tests * stxxl::vector: implemented some more functions * const char* stxxl::get_version_string(): new function * comparators inherit from std::binary_function<> * cleanup, reformat the code to a consistent format (Thanks, uncrustify!) ------------------------------------------ Version 1.0e (4 June 2007) * Bugfix: for objects with the destructors, memory block were allocated in a non-aligned fashion * Internal CPU efficiency of Stxxl priority queue is improved (loser trees in external mergers are integrated) * Fixed bug with conversion of const/non-const iterators, implemented comparison and difference operations on const/non-const iterators (for vector, map, and deque) * Added operator[] to vector and deque iterators * stxxl::random_shuffle added * improved Makefile structure (thanx to Andreas Beckmann) * Allow parallel build for g++ ------------------------------------------ Version 1.0d (16 January 2007) * Support of Visual Studio 2005 Express (VC++ 8.0) * Algorithms and data structures of Stxxl can now use more than 4GB of main memory on 64-bit processors/compilers * Support of error reporting using the C++ exception mechanism ------------------------------------------- Version 1.0c (21 September 2006) * An implementation of an I/O-efficient deque * STXXL uses MCSTL library (optional) to improve the performance of stxxl::sort and pipelined sort on SMP and multicore processors. ------------------------------------------- Version 0.99 (22 March 2006) * Better compiler support: g++ (versions 3.3.x-4.0.x) and Microsoft Visual C++ 7.1 (.NET) * New B+Tree-based implementation of map (compatible with all listed above compilers): I/O-efficient map ------------------------------------------ Version 0.9 (9 August 2005) * STXXL has been ported to Windows. It now can be run under Windows XP and Windows 2000 * STXXL can be compiled now by g++ (versions 3.0.x-3.4.x, 4.0.x) and Microsoft Visual C++ 7.1 (.NET) * New data structure: I/O efficient FIFO queue ------------------------------------------ Version 0.77 (24 March 2005) * An implementation of queue is available ------------------------------------------ Version 0.75 (23 March 2005) * An implementation of map based on B+tree is available ------------------------------------------ Version 0.7 (25 January 2005) * The implementation of pipelining is extended and improved ------------------------------------------ Version 0.6 (5 September 2004) * Tested implementation of the stream package (aka pipelining) is available ------------------------------------------ Version 0.5 (21 November 2003) * The first implementation of the stream package (aka pipelining) is available * Priority queue is available ------------------------------------------ Version 0.2 (Summer 2003) * The first public release * Vectors, stacks, sorting, scanning are available ------------------------------------------ # Local Variables: # mode: text # mode: flyspell # End: