ChangeLog 17 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402
  1. # StarPU --- Runtime system for heterogeneous multicore architectures.
  2. #
  3. # Copyright (C) 2009-2013 Université de Bordeaux 1
  4. # Copyright (C) 2010, 2011, 2012, 2013 Centre National de la Recherche Scientifique
  5. #
  6. # StarPU is free software; you can redistribute it and/or modify
  7. # it under the terms of the GNU Lesser General Public License as published by
  8. # the Free Software Foundation; either version 2.1 of the License, or (at
  9. # your option) any later version.
  10. #
  11. # StarPU is distributed in the hope that it will be useful, but
  12. # WITHOUT ANY WARRANTY; without even the implied warranty of
  13. # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  14. #
  15. # See the GNU Lesser General Public License in COPYING.LGPL for more details.
  16. StarPU 1.1.0 (svn revision xxxx)
  17. ==============================================
  18. New features:
  19. * OpenGL interoperability support.
  20. * Capability to store compiled OpenCL kernels on the file system
  21. * Capability to load compiled OpenCL kernels
  22. * Performance models measurements can now be provided explicitly by
  23. applications.
  24. * Capability to emit communication statistics when running MPI code
  25. * Add starpu_unregister_submit, starpu_data_acquire_on_node and
  26. starpu_data_invalidate_submit
  27. * New functionnality to wrapper starpu_insert_task to pass a array of
  28. data_handles via the parameter STARPU_DATA_ARRAY
  29. * Enable GPU-GPU direct transfers.
  30. * GCC plug-in
  31. - Add `registered' attribute
  32. - A new pass was added that warns about the use of possibly
  33. unregistered memory buffers.
  34. * SOCL
  35. - Manual mapping of commands on specific devices is now
  36. possible
  37. - SOCL does not require StarPU CPU tasks anymore. CPU workers
  38. are automatically disabled to enhance performance of OpenCL
  39. CPU devices
  40. * New interface: COO matrix.
  41. * Data interfaces: The pack operation of user-defined data interface
  42. defines a new parameter count which should be set to the size of
  43. the buffer created by the packing of the data.
  44. * MPI:
  45. - Communication statistics for MPI can only be enabled at
  46. execution time by defining the environment variable
  47. STARPU_COMM_STATS
  48. - Communication cache mechanism is enabled by default, and can
  49. only be disabled at execution time by setting the
  50. environment variable STARPU_MPI_CACHE to 0.
  51. - Initialisation functions starpu_mpi_initialize_extended()
  52. and starpu_mpi_initialize() have been made deprecated. One
  53. should now use starpu_mpi_init(int *, char ***, int). The
  54. last parameter indicates if MPI should be initialised.
  55. - Collective detached operations have new parameters, a
  56. callback function and a argument. This is to be consistent
  57. with the detached point-to-point communications.
  58. - When exchanging user-defined data interfaces, the size of
  59. the data is the size returned by the pack operation, i.e
  60. data with dynamic size can now be exchanged with StarPU-MPI.
  61. - New functionality starpu_mpi_irecv_probe_detached which
  62. first tests if the message is available before calling MPI_Recv.
  63. * Add experimental simgrid support, to simulate execution with various
  64. number of CPUs, GPUs, amount of memory, etc.
  65. * Add support for OpenCL simulators (which provide simulated execution time)
  66. * Add support for Temanejo, a task graph debugger
  67. * Theoretical bound lp output now includes data transfer time.
  68. * Update OpenCL driver to only enable CPU devices (the environment
  69. variable STARPU_OPENCL_ONLY_ON_CPUS must be set to a positive
  70. value when executing an application)
  71. * Add Scheduling contexts to separate computation resources
  72. - Scheduling policies take into account the set of resources corresponding
  73. to the context it belongs to
  74. - Add support to dynamically change scheduling contexts
  75. (Create and Delete a context, Add Workers to a context, Remove workers from a context)
  76. - Add support to indicate to which contexts the tasks are submitted
  77. * Add the Hypervisor to manage the Scheduling Contexts automatically
  78. - The Contexts can be registered to the Hypervisor
  79. - Only the registered contexts are managed by the Hypervisor
  80. - The Hypervisor can detect the initial distribution of resources of
  81. a context and constructs it consequently (the cost of execution is required)
  82. - Several policies can adapt dynamically the distribution of resources
  83. in contexts if the initial one was not appropriate
  84. - Add a platform to implement new policies of redistribution
  85. of resources
  86. * Implement a memory manager which checks the global amount of
  87. memory available on devices, and checks there is enough memory
  88. before doing an allocation on the device.
  89. * Discard environment variable STARPU_LIMIT_GPU_MEM and define
  90. instead STARPU_LIMIT_CUDA_MEM and STARPU_LIMIT_OPENCL_MEM
  91. * Introduce new variables STARPU_LIMIT_CUDA_devid_MEM and
  92. STARPU_LIMIT_OPENCL_devid_MEM to limit memory per specific device
  93. * Introduce new variable STARPU_LIMIT_CPU_MEM to limit memory for
  94. the CPU devices
  95. Small features:
  96. * Add starpu_worker_get_by_type and starpu_worker_get_by_devid
  97. * Add starpu_fxt_stop_profiling/starpu_fxt_start_profiling which permits to
  98. pause trace recording.
  99. * Add trace_buffer_size configuration field to permit to specify the tracing
  100. buffer size.
  101. * Add starpu_codelet_profile and starpu_codelet_histo_profile, tools which draw
  102. the profile of a codelet.
  103. * File STARPU-REVISION --- containing the SVN revision number from which
  104. StarPU was compiled --- is installed in the share/doc/starpu directory
  105. * starpu_perfmodel_plot can now directly draw GFlops curves.
  106. Changes:
  107. * Fix the block filter functions.
  108. * Fix StarPU-MPI on Darwin.
  109. * The FxT code can now be used on systems other than Linux.
  110. * Keep only one hashtable implementation common/uthash.h
  111. * The cache of starpu_mpi_insert_task is fixed and thus now enabled by
  112. default.
  113. * Improve starpu_machine_display output.
  114. * Standardize objects name in the performance model API
  115. * SOCL
  116. - Virtual SOCL device has been removed
  117. - Automatic scheduling still available with command queues not
  118. assigned to any device
  119. - Remove modified OpenCL headers. ICD is now the only supported
  120. way to use SOCL.
  121. - SOCL test suite is only run when environment variable
  122. SOCL_OCL_LIB_OPENCL is defined. It should contain the location
  123. of the libOpenCL.so file of the OCL ICD implementation.
  124. * Fix main memory leak on multiple unregister/re-register.
  125. * Improve hwloc detection by configure
  126. * Cell:
  127. - It is no longer possible to enable the cell support via the
  128. gordon driver
  129. - Data interfaces no longer define functions to copy to and from
  130. SPU devices
  131. - Codelet no longer define pointer for Gordon implementations
  132. - Gordon workers are no longer enabled
  133. - Gordon performance models are no longer enabled
  134. * Fix data transfer arrows in paje traces
  135. * The "heft" scheduler no longer exists. Users should now pick "dmda"
  136. instead.
  137. * StarPU can now use poti to generate paje traces.
  138. * Rename scheduling policy "parallel greedy" to "parallel eager"
  139. * Rename all filter functions to follow the pattern
  140. starpu_DATATYPE_filter_FILTERTYPE. The script
  141. tools/dev/rename_filter.sh is provided to rename your existing
  142. applications using filters.
  143. * Rename function starpu_helper_cublas_init to starpu_cublas_init
  144. * Rename function starpu_helper_cublas_shutdown to starpu_cublas_shutdown
  145. * Rename function starpu_allocate_buffer_on_node to starpu_malloc_on_node
  146. * Rename function starpu_free_buffer_on_node to starpu_free_on_node
  147. Small changes:
  148. * STARPU_NCPU should now be used instead of STARPU_NCPUS. STARPU_NCPUS is
  149. still available for compatibility reasons.
  150. * include/starpu.h includes all include/starpu_*.h files, applications
  151. therefore only need to have #include <starpu.h>
  152. * Active task wait is now included in blocked time.
  153. * Fix GCC plugin linking issues starting with GCC 4.7.
  154. * Fix forcing calibration of never-calibrated archs.
  155. * CUDA applications are no longer compiled with the "-arch sm_13"
  156. option. It is specifically added to applications which need it.
  157. StarPU 1.0.3 (svn revision 7379)
  158. ==============================================
  159. Changes:
  160. * Several bug fixes in the build system
  161. * Bug fixes in source code for non-Linux systems
  162. * Fix generating FXT traces bigger than 64MiB.
  163. * Improve ENODEV error detections in StarPU FFT
  164. StarPU 1.0.2 (svn revision xxx)
  165. ==============================================
  166. Changes:
  167. * Add starpu_block_shadow_filter_func_vector and an example.
  168. * Add tag dependency in trace-generated DAG.
  169. * Fix CPU binding for optimized CPU-GPU transfers.
  170. * Fix parallel tasks CPU binding and combined worker generation.
  171. * Fix generating FXT traces bigger than 64MiB.
  172. StarPU 1.0.1 (svn revision 6659)
  173. ==============================================
  174. Changes:
  175. * hwloc support. Warn users when hwloc is not found on the system and
  176. produce error when not explicitely disabled.
  177. * Several bug fixes
  178. * GCC plug-in
  179. - Add `#pragma starpu release'
  180. - Fix bug when using `acquire' pragma with function parameters
  181. - Slightly improve test suite coverage
  182. - Relax the GCC version check
  183. * Update SOCL to use new API
  184. * Documentation improvement.
  185. StarPU 1.0.0 (svn revision 6306)
  186. ==============================================
  187. The extensions-again release
  188. New features:
  189. * Add SOCL, an OpenCL interface on top of StarPU.
  190. * Add a gcc plugin to extend the C interface with pragmas which allows to
  191. easily define codelets and issue tasks.
  192. * Add reduction mode to starpu_mpi_insert_task.
  193. * A new multi-format interface permits to use different binary formats
  194. on CPUs & GPUs, the conversion functions being provided by the
  195. application and called by StarPU as needed (and as less as
  196. possible).
  197. * Deprecate cost_model, and introduce cost_function, which is provided
  198. with the whole task structure, the target arch and implementation
  199. number.
  200. * Permit the application to provide its own size base for performance
  201. models.
  202. * Applications can provide several implementations of a codelet for the
  203. same architecture.
  204. * Add a StarPU-Top feedback and steering interface.
  205. * Permit to specify MPI tags for more efficient starpu_mpi_insert_task
  206. Changes:
  207. * Fix several memory leaks and race conditions
  208. * Make environment variables take precedence over the configuration
  209. passed to starpu_init()
  210. * Libtool interface versioning has been included in libraries names
  211. (libstarpu-1.0.so, libstarpumpi-1.0.so,
  212. libstarpufft-1.0.so, libsocl-1.0.so)
  213. * Install headers under $includedir/starpu/1.0.
  214. * Make where field for struct starpu_codelet optional. When unset, its
  215. value will be automatically set based on the availability of the
  216. different XXX_funcs fields of the codelet.
  217. * Define access modes for data handles into starpu_codelet and no longer
  218. in starpu_task. Hence mark (struct starpu_task).buffers as
  219. deprecated, and add (struct starpu_task).handles and (struct
  220. starpu_codelet).modes
  221. * Fields xxx_func of struct starpu_codelet are made deprecated. One
  222. should use fields xxx_funcs instead.
  223. * Some types were renamed for consistency. when using pkg-config libstarpu,
  224. starpu_deprecated_api.h is automatically included (after starpu.h) to
  225. keep compatibility with existing software. Other changes are mentioned
  226. below, compatibility is also preserved for them.
  227. To port code to use new names (this is not mandatory), the
  228. tools/dev/rename.sh script can be used, and pkg-config starpu-1.0 should
  229. be used.
  230. * The communication cost in the heft and dmda scheduling strategies now
  231. take into account the contention brought by the number of GPUs. This
  232. changes the meaning of the beta factor, whose default 1.0 value should
  233. now be good enough in most case.
  234. Small features:
  235. * Allow users to disable asynchronous data transfers between CPUs and
  236. GPUs.
  237. * Update OpenCL driver to enable CPU devices (the environment variable
  238. STARPU_OPENCL_ON_CPUS must be set to a positive value when
  239. executing an application)
  240. * struct starpu_data_interface_ops --- operations on a data
  241. interface --- define a new function pointer allocate_new_data
  242. which creates a new data interface of the given type based on
  243. an existing handle
  244. * Add a field named magic to struct starpu_task which is set when
  245. initialising the task. starpu_task_submit will fail if the
  246. field does not have the right value. This will hence avoid
  247. submitting tasks which have not been properly initialised.
  248. * Add a hook function pre_exec_hook in struct starpu_sched_policy.
  249. The function is meant to be called in drivers. Schedulers
  250. can use it to be notified when a task is about being computed.
  251. * Add codelet execution time statistics plot.
  252. * Add bus speed in starpu_machine_display.
  253. * Add a STARPU_DATA_ACQUIRE_CB which permits to inline the code to be
  254. done.
  255. * Add gdb functions.
  256. * Add complex support to LU example.
  257. * Permit to use the same data several times in write mode in the
  258. parameters of the same task.
  259. Small changes:
  260. * Increase default value for STARPU_MAXCPUS -- Maximum number of
  261. CPUs supported -- to 64.
  262. * Add man pages for some of the tools
  263. * Add C++ application example in examples/cpp/
  264. * Add an OpenMP fork-join example.
  265. * Documentation improvement.
  266. StarPU 0.9 (svn revision 3721)
  267. ==============================================
  268. The extensions release
  269. * Provide the STARPU_REDUX data access mode
  270. * Externalize the scheduler API.
  271. * Add theoretical bound computation
  272. * Add the void interface
  273. * Add power consumption optimization
  274. * Add parallel task support
  275. * Add starpu_mpi_insert_task
  276. * Add profiling information interface.
  277. * Add STARPU_LIMIT_GPU_MEM environment variable.
  278. * OpenCL fixes
  279. * MPI fixes
  280. * Improve optimization documentation
  281. * Upgrade to hwloc 1.1 interface
  282. * Add fortran example
  283. * Add mandelbrot OpenCL example
  284. * Add cg example
  285. * Add stencil MPI example
  286. * Initial support for CUDA4
  287. StarPU 0.4 (svn revision 2535)
  288. ==============================================
  289. The API strengthening release
  290. * Major API improvements
  291. - Provide the STARPU_SCRATCH data access mode
  292. - Rework data filter interface
  293. - Rework data interface structure
  294. - A script that automatically renames old functions to accomodate with the new
  295. API is available from https://scm.gforge.inria.fr/svn/starpu/scripts/renaming
  296. (login: anonsvn, password: anonsvn)
  297. * Implement dependencies between task directly (eg. without tags)
  298. * Implicit data-driven task dependencies simplifies the design of
  299. data-parallel algorithms
  300. * Add dynamic profiling capabilities
  301. - Provide per-task feedback
  302. - Provide per-worker feedback
  303. - Provide feedback about memory transfers
  304. * Provide a library to help accelerating MPI applications
  305. * Improve data transfers overhead prediction
  306. - Transparently benchmark buses to generate performance models
  307. - Bind accelerator-controlling threads with respect to NUMA locality
  308. * Improve StarPU's portability
  309. - Add OpenCL support
  310. - Add support for Windows
  311. StarPU 0.2.901 aka 0.3-rc1 (svn revision 1236)
  312. ==============================================
  313. The asynchronous heterogeneous multi-accelerator release
  314. * Many API changes and code cleanups
  315. - Implement starpu_worker_get_id
  316. - Implement starpu_worker_get_name
  317. - Implement starpu_worker_get_type
  318. - Implement starpu_worker_get_count
  319. - Implement starpu_display_codelet_stats
  320. - Implement starpu_data_prefetch_on_node
  321. - Expose the starpu_data_set_wt_mask function
  322. * Support nvidia (heterogeneous) multi-GPU
  323. * Add the data request mechanism
  324. - All data transfers use data requests now
  325. - Implement asynchronous data transfers
  326. - Implement prefetch mechanism
  327. - Chain data requests to support GPU->RAM->GPU transfers
  328. * Make it possible to bypass the scheduler and to assign a task to a specific
  329. worker
  330. * Support restartable tasks to reinstanciate dependencies task graphs
  331. * Improve performance prediction
  332. - Model data transfer overhead
  333. - One model is created for each accelerator
  334. * Support for CUDA's driver API is deprecated
  335. * The STARPU_WORKERS_CUDAID and STARPU_WORKERS_CPUID env. variables make it possible to
  336. specify where to bind the workers
  337. * Use the hwloc library to detect the actual number of cores
  338. StarPU 0.2.0 (svn revision 1013)
  339. ==============================================
  340. The Stabilizing-the-Basics release
  341. * Various API cleanups
  342. * Mac OS X is supported now
  343. * Add dynamic code loading facilities onto Cell's SPUs
  344. * Improve performance analysis/feedback tools
  345. * Application can interact with StarPU tasks
  346. - The application may access/modify data managed by the DSM
  347. - The application may wait for the termination of a (set of) task(s)
  348. * An initial documentation is added
  349. * More examples are supplied
  350. StarPU 0.1.0 (svn revision 794)
  351. ==============================================
  352. First release.
  353. Status:
  354. * Only supports Linux platforms yet
  355. * Supported architectures
  356. - multicore CPUs
  357. - NVIDIA GPUs (with CUDA 2.x)
  358. - experimental Cell/BE support
  359. Changes:
  360. * Scheduling facilities
  361. - run-time selection of the scheduling policy
  362. - basic auto-tuning facilities
  363. * Software-based DSM
  364. - transparent data coherency management
  365. - High-level expressive interface
  366. # Local Variables:
  367. # mode: text
  368. # coding: utf-8
  369. # ispell-local-dictionary: "american"
  370. # End: