Obey the control-flow dependencies of DAX files

[simgrid.git] / ChangeLog

SimGrid (3.4.1-svn) unstable; urgency=low

 SimDag
 * Warn the user about loop dependencies in data flow of DAX files
 * Obey the control-flow dependencies of DAX files

 Java Bindings
 * Fix a bug preventing the tasks from begin garbage collected.

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr>

SimGrid (3.4) stable; urgency=low

 The "Easter in Cargese" release. Also known as (major changes):
 
  * the "se habla Java, Ruby 話せます, fala-se Lua (and deaf-friendly)"
    ~> bindings were greatly improved
    ~> new tracing infrastructure for better visualization introduced
    
  * the "Welcome to configury modernity" release.
    ~> we switched from autotools to cmake, and improved our cdash


 A more detailled list of changes follow (full detail in svn log).

 Java Bindings: Various Cleanups
  * (install java-gcj-compat-dev on debian-like to use them)
  * Remove put/get: no need to export deprecated interface in Java
    Use send/receive instead.
  * Cleanup the examples and add a README per directory
  * Remove example autoDestination (that's the only way to go now)
  * Remove example explicitDestination (was a plain copy of basic)  
  * Make JniException a runtime exception, so that there is no need to
    declare the fact that you may encounter such a beast. I guess that
    nobody will ever want to survive such error.
  * Create specific errors for each MSG case of failure:
    host failure, transfer failure, timeout, task cancelled
  * Cleanup the exceptions that may get thrown by each function
  * Other internal cleanups in Java bindings. Performance still bad :/
 Ruby and Lua Bindings: create them
  * (install ruby1.8-dev/liblua5.1-0-dev on debian-like to use them)
  * That's new and great, you should try them out. 
    Same functionalities than Java bindings, only even less polished
 SimDag:
  * Kill the useless "rate" argument of SD_task_get_execution_time()
    Everyone used to provide -1 as a value, it was not used, and the
    semantic of a possible use wasn't even clear.
  * SD_SCHED_NO_COST: Constant to use as cost in SD_task_schedule()
    either as comm costs or compute costs to mean that there is no
    such thing for that specific task.
  * Add a SD_task_set_name() function
  * Fix SD_task_unschedule() on typed tasks
  * Fix SD_task_get_execution_time() to return seconds, not flop*sec
  * In DAX loader, accept useless 'level' attributes to <job> since
    LIGO DAGs have them (seem to be to ease graphical representation).
 MSG: 
  * Add an example masterslave_mailbox.c using send/receive and not
    the deprecated put/get interface.
  * Kill the MSG_paje_output() function. It's a noop since 2 years.
  * Kill MSG_WARNING and MSG_FATAL return codes: they were not used
    anywere in source.
  * Rename MSG_TIMEOUT_FAILURE into MSG_TIMEOUT for sake of logic
    (declare MSG_USE_DEPRECATED to still have the old name)
  * Add a MSG_task_set_data() function
  * About trace replay (see examples/msg/actions):
    - implement barrier
    - Allow to work with splitted trace files for each process
      Give the specific trace file as argument of each process, 
        and call MSG_action_trace_run(NULL)
      You can still have one merged file for all processes.
    - Fix implementation of collective operations
  * Allow task_execute() on 0-sized tasks (closes #10063)
 SMPI:
  * This is the first release of SimGrid where SMPI is not considered
    beta anymore (even if some corners should still be improved)
  * Port over the new SIMIX_network submodule (internal refactoring)
  * Basic support to log events as with SMPE (use --cfg=SMPE:1)
  * Implement more missing elements of the standard:
    - MPI_COMM_SELF 
    - MPI_MAXLOC MPI_MINLOC + all associated datatype MPI_DOUBLE_INT,
      MPI_FLOAT_INT, etc.
    - MPI_Address() MPI_Get_count() MPI_Type_free() MPI_Type_extent()
      MPI_Scan() MPI_Get_processor_name()
    - Added implementation of missing case for Alltoall (warning: it's
      *not* the bruck variant from OpenMPI; based on Alltoallv instead)
    - SMPI_MPI_Gather() SMPI_MPI_Gatherv() SMPI_MPI_Scatterv()
      SMPI_MPI_Reduce_scatter() SMPI_MPI_Allgather()
      SMPI_MPI_Allgatherv() 
  * Bug fixes include:
    - MPI_Waitsome() was broken
    - Allow relative includes in smpicc
    - Command line cfg argument 'reference_speed' was ignored...
    - Some functions did not properly lead to auto-benching of user code
    - smpicc passes -O2 by default (just like openmpi one)
 SIMIX:
  * add SIMIX_action_suspend() and SIMIX_action_resume() functions
  * Bug fixes about timeouts during communications
  * add SIMIX_message_sizes_output() as a pimple to write to file the
    amount of messages per size. Use gnuplot to get histogram. 
    Pimple because that's the only user-visible function of simix,
     defined directly in xbt.h (irk, sorry)
  * About semaphores:
     - Add a SIMIX_sem_get_capacity() function
     - Fix interactions with processe resume/suspende
     - release_forever() was stupidly broken
     - Fix SIMIX_display_process_status() for processes in a semaphore
     - Make SIMIX_sem_block_onto() user-visible
  * Refactoring context stuff:
    - Use pseudo-OOP for better modularity
    - reimplement SIMIX_process_kill() without process_schedule() so
      that the latter can take as invariant that it is called from
      maestro.
    - Merge context_start into context_new for sake of simplicity
 SURF:
  * Add a Vivaldi network model, coded live during SUD'10 ;)
  * Rename configuration variables to start a hierarchy:
    o cpu_model -> cpu/model
    o network_model -> network/model
    o workstation_model -> workstation/model
  * New configuration variables:
    o network/bandwidth_factor: correction to bandwith
    o network/latency_factor: correction to latency
    o netwotk/weight_S: correction to the weight of competing streams
  * Add a long description to the models, that users can see with such
    argument on the command line: --cfg=cpu/model:help
  * --help-models display the long description of all known models
 XBT:
  * config: add the ability to set a default value after registration
    Does not override any previously set value (e.g. from cmd line)
  * dict: allow to have integer key and data.
    When so, you need to use the following functions
     void xbt_dicti_set(xbt_dict_t dict, uintptr_t key, uintptr_t data);
     uintptr_t xbt_dicti_get(xbt_dict_t dict, uintptr_t key);
     void xbt_dicti_remove(xbt_dict_t dict, uintptr_t key);
    In contrary to regular dicts, the key is not malloced before copy.
    Mixing scalar and regular elements in the same dict is not tested 
      (but may work).
  * Allow to use xbt_dynar_shrink() to expend the dynar instead
 Tracing for Visualization:
  * SimGrid is now instrumented in order to generate a trace file for
    visualization analysis: to use it, need to compile SimGrid with the
    "tracing" option enabled, and instrument the program using SimGrid with
    TRACE_start, TRACE_category, TRACE_msg_set_task_category and TRACE_end
    (among other functions).
  * The instrumentation only traces the platform utilization for now
  * Documentation to use the tracing functions and how to analyze the
    traces with the Triva tool is written.
  * More information about: SimGrid FAQ (in the section Tracing Simulations
    for Visualization)
 Build system:
  * We moved to cmake as default build system. Autotools support will
    be dropped soon. Check the FAQ for more info about how to use it.
  * Greatly improved our cdash/ctest interactions
    Check http://cdash.inria.fr/CDash/index.php?project=Simgrid
  * Added memory checking tests with valgrind; lot of memleak fixing.
    This may be the first release of simgrid with so few memory issues
  * Added code coverage tests. 
    Our coverage is still improvable, but at least we see it on cdash.
   
 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr> Wed, 28 Apr 2010 28 17:11:16 +0100

SimGrid (3.3.4) stable; urgency=low

 The "Desktop Grid needs love too" release (also called Xmas release).

 Models improvements:
 * Major speedup in the maxmin system solving by using lazy evaluation
   Instead of solving completely the maxmin system at each iteration, 
     only invalidate (and recompute) the modified parts. 
   This new feature is enabled in default models but you can try to
     turn it on with "--cfg:maxmin-selective-update=1" for other models.
 * Cas01 IMproved as default CPU model
   This CPU model is the same Cas01 model, but it uses the
     maxmin-selective-update flag and a heap structure to manage
     actions on SURF kernel. 
   It reduces the complexity to find the next action to finish and,
     consequently, it's faster than the old Cas01.
   This is the new default CPU model (Cas01).   
 * Rename the old Cas01 model to Cas01_fullupdate
   Keep the old cpu model Cas01 with the new name of Cas01_fullupdate.
   Use "--cfg=cpu_model:Cas01_fullupdate" to use the old default CPU model.
 * CpuTI (CPU Trace Integration)
   A new CPU model whose objective is simulate faster when using
     availability trace files. 
   Instead of using a full featured, over engineered maxmin system for
     CPU modeling, this model does the pre-integration of traces files
     to calculate the amount of CPU power available, and so, executes
     faster than the old CPU models. 
   Use "--cfg=cpu_model:CpuTI" to change to this CPU model.
 * Use LV08 as default network model since it gives better accuracy
    for small messages and shouldn't change things for big ones.
   Use --cfg=network_model:CM02 to get the previous behavior.
   
   
         ******************************************
         *DO NOT MIX 3.3.4 RESULTS WITH OLDER ONES* 
         ******************************************
   * The new CPU model may changes simulations!
     The point is that events occurring at the exact same timestamp
        are not scheduled in the same order with the old and new 
        version. This may be enough to completely change the execution
        of simulations in some cases. 
   * The new network model will change simulations!
     This new model is more realistic than the previous one, so you
       should consider redoing your old experiments with this model.
     Sorry for the inconvenience.
     
 Build System:
 * Introduce the supernovae compilation mode
   When compiled that way, the whole SimGrid (or almost) is put in a
     single compilation unit and compiled in one shoot. 
  This is to help gcc which has difficulties to inline stuff from one
     file into another.
  The speedup seem to be above 15%, althrough more tests are needed on
     amd64 to confirm that gain.

 MSG:
 * Port of MSG's mailbox on top of SIMIX network
   The put/get mechanism was greatly simplified on the way.

 SIMIX:
 * New SIMIX network module. Provides:
   - Mailbox: rendez-vous mecanism to find with who you want to speak
   - Synchronous send/recv: easier and hopefully faster since the
     logic is handled in the maestro process directly now
   - Asynchronous send/recv: you dreamt of it? It's here now
     Too bad that nobody cared enough to propagate the change to MSG.
 * Add semaphores as SIMIX synchronization mechanism.
   
 SimDag:
 * new function SD_daxload(char*) to load a DAX file 
   (see http://vtcpc.isi.edu/pegasus/index.php/WorkflowGenerator)
 * Introduce typed tasks. Specify its kind and cost at creation. 
   At scheduling, just give where it should be placed, and the cost
   for each involved resource is automatically computed.
   Existing constructors so far (more to come of course):
    - SD_task_create_comm_e2e() for end-to-end communication
    - SD_task_create_comp_seq() for sequential computation
   Use SD_task_schedulev() / SD_task_schedulel() to schedule them.
 * new function SD_task_dump() for debuging display
 * new function SD_task_dotty(task,FILE*) writing to file the info
   about the task in dotty format
 * SD_task_dependency_exists() can now cope with having one of its
   arguments NULL. If so, it tests whether the other argument has any 
   dependency.
 * Add getters on list of preceding/following tasks:
    SD_task_get_parents(task) and SD_task_get_children(task)
 * Add getters on amount of workstations and list:
    SD_task_get_workstation_count(t) and SD_task_get_workstation_list(t)
 * Add getter on task kind: SD_task_get_kind(task)
 * Update the start_time and finish_time of tasks on completion/failure
 * Bugfix: Remove task from state swags when destroyed
 
 GRAS:
 * New function: void gras_cpu_burn(double flops) -- a simple CPU burner

 XBT:
 * New function: xbt_dynar_dopar(dynar,fun) to map a function over the
   dynar with one separate thread per value of the dynar.
 * Change the prototype of xbt_thread_create(), sorry. 
   Added a boolean parameter indicating whether we want to join this
   thread (used in SG only for now)
 * Implement xbt_thread_join and xbt_thread_yield in SG also.
   
 Bug fixes:
 * GTNetS wrappers should now be usable again (and betterly tested too)
 * Fix a major regression from 3.2 where the timeout provided to
   MSG_task_put_with_timeout() was used as absolute time before which
   the comm should be done.
 * Start to fix the <cluster> tag. 
   - Internal links should be good now (beside of the loopback, which
     use the private link instead)
   - paths to the external world is still rather broken
   - the <route:multi> tag is just broken. Actually that's brain-dead.
     We need sth like <route:multi src="myCluster" dst="$*-${myCluster}">
     to make it less stupid
   ** Check your platform with teshsuite/simdag/platforms/flatifier **
 * Fix a source-level compatibility glitch from 3.2: after defining
   MSG_USE_DEPRECATED, you can use the old name
   MSG_task_put_with_time_out() for MSG_task_put_with_timeout()
 * Allow to compile from the SVN with automake 1.11
 * Fix some problems when using the "start_time" tag in deployment XMLs.
 * Fix #8569: XBT/synchro.h has redundant declarations
 * Fix #8563: MSG return values and exceptions
   Introduce a MSG_TIMEOUT_FAILURE return code and use it consistently.
 * Integrate patch #8636: Obey DESTDIR when installing documentation.
   Thanks to Robson Peixoto.
 * Fix a vicious bug in dictionaries inducing that some elements were
   not freed on xbt_dict_free()

 Portability report of this version:
  * Main portability targets:
    - linux (ubuntu (804/810/910) /debian (4/5/testing) /fedora (core11)) 
      on (amd64/i386/ia64)
    - mac leopard on i386
    Known problems: http://cdash.inria.fr/CDash/index.php?project=Simgrid
     but nothing critical.
  * Other platforms: windows, AIX and others were not tested for this release
  
 Timing report of this version:
  * Lazy evaluation brings arbitrary speedup (ie, speedup depending on
    scenario parameters). From 8h to a few seconds in desktop grid settings.
  * Supernovae brings about 25% speedup on i386.

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr> Thu, 24 Dec 2009 19:07:39 +0100

SimGrid (3.3.3) stable; urgency=low

 The "Need for Speed" release.
 
 The timings done to validate the 3.3.2 were faulty. 
 Instead of being 5% faster, it was 15% slower (compared to 3.3.1).
   
 The problem was a conversion from a manually handled vector to
   xbt_dynar_t on the critical path. 
 xbt_dynar_foreach calls functions, inducing stack management crap.

 We inlined these functions and xbt_dynar_foreach is now breath taking.
 We also inlined xbt_swag_belong on the way.

 Here are some approximate speedup measurements (on master/slaves
  simulations lasting between 10s and 20s each):
   3.3.1                   -> 3.3.2: about same performance
   3.3.2                   -> 3.3.3: 40% speedup
   3.3.1                   -> 3.3.3: 40% speedup
   3.3.1 with inline patch -> 3.3.3: 30% speedup
   
 Our reading is that the refactoring which occurred in 3.3.2 made us
  suffer much more from the xbt_dynar_foreach low performance, but
  once we solved this, this refactoring proved to be very performance
  effective. From the 40% speedup, somehow, 10% are due to the
  inlining and 30% to the refactoring.

 That's a pitty that gcc cannot inline functions placed in other files
  alone. We have to choose between:
  - break the encapsulation (by putting private data structures and
    accessors in headers files to help gcc)
  - live with low performance 
  - switch to a decent compiler such as icc (not quite possible).

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr> Thu, 20 Aug 2009 21:21:33 +0200

SimGrid (3.3.2) stable; urgency=low

 The "Simplicity does not preceed complexity, but follows it" release.

 The main contributors of this release were (lexical order):
   Silas De Munck, Stéphane Genaud, Martin Quinson, Cristian Rosa.
       
 SURF: 
  * Extract the routing logic into its own object.
    (was dupplicated in network.c and workstation_LV07.c;
     Allows to implement other ways of storing that info)
    => kill now useless network_card concept
    - Use dynar to represent routes (instead of void** + int*)
    - kill link_set (use surf_network_model->resource_set instead)
    - Add a command-line option to choose the routing schema to use
    - Add three new models: 
      * Floyd (shortest path computed at initialization)
      * Dijikstra (shortest path recomputed all the time)
      * Cached Dijikstra (shortest path computed on need)
      All these models where contributed by Silas De Munck, and are
      described in his ICCS09 paper.

  * Simplify model declaration
    (less redirections, less function to write when defining a model)
    - Factorize stuff between models:
      - model_init/exit
      - Set of resources:
        surf_model_resource_set(model)
        surf_model_resource_by_name(model, name)
    - Unify the types of models in s_surf_model_t (using an union)
    - Embeed fields of common_public directly into s_surf_model_t
    - Rename model methods:
      action_free ~> action_unref
      action_change_state ~> action_state_set
      action_get_state    ~> action_state_get
    - Change model methods into functions :
      (model)->common_public->action_use  ~> surf_action_ref
      
  * Implement a generic resource; use it as ancestor to specific ones
    (allows to kill duplicated code in models)
    Drawback: timer command don't need no name nor properties;
              workstation_CLM03 don't need no properties
    (but I guess we can live with those few bytes wasted)
    
  * Improve the action object model
    - implement a constructor avoiding dupplicated code about field
      initialization in generic_action part.
      
  * Kill the SDP model: it has an external dependency, is deprecated
    in flavor of modern lmm models, and didn't compile since a while
 
 SIMIX:
  * Relocation of the context module from XBT to SIMIX.
    (the context were decoupled from the simix processes, duplicating a lot of code)
    => a lot of code was factorized
    - less overhead is introduced during scheduling
    - simpler API for the context factory
    - the logic for process creation,destruction and manipulation was simplified
  * Simplification of the s_smx_process_t data structure.
    => accesing the simix level data associated to a process is faster now, 
       and the code is a lot more readable.
       
 SMPI:
  * Implement some more MPI primitives: 
    MPI_Bcast, MPI_Waitany, MPI_Waitall, MPI_Reduce, MPI_Allreduce, MPI_Scatter, MPI_Sendrecv, MPI_Alltoall
    -implementation: Bcast: flat or 2-ary tree (default), 
                     Barrier: 4-ary tree,
                     Reduce: flat tree
                     Allreduce: Reduce then Bcast
                     Alltoall: "basic_linear" if data per proc < 3Kb, "otherwise pairwise". 
                               Not yet implemented: "Bruck" for data per proc < 200b and comm size > 12
                     Alltoallv: flat tree, like ompi
                     Scatter: flat tree
  * Add support for optimized collectives (Bcast is now binomial by default)
  * Port smpirun and smpicc to OS X

 SimDag:
  * Kill SD_link_get_properties: hard to maintain and makes very little sense
    Shout out if you used it.
    
 GRAS:
  * Display the list of still queued messages in SG mode when existing
    the process.

 XBT:
  * Add xbt_set_get_by_name_or_null() [Silas De Munck]
  * Add xbt_graph_node_get_outedges() [Silas De Munck]
  * Add xbt_str_from_file(FILE*)
  * Add xbt_dict_get_key achieving a linear reverse search
  * Remove the context module 

 Portability report of this version:
  * Main portability targets:
    - Linux(debian)/x86/context   
    - Linux(debian)/x86/pthreads 
    - Linux(debian)/amd64/context 
    - Linux(debian)/amd64/pthreads
    On these, we still have the eratic breakages of gras/pmm and 
      amok/saturate_sg reported in previous version. We still think
      that the tests are the cause of the fault, not the tested code.
      
    - Mac OSX Leopard/x86/context
    Still false negative in tesh autotesting.
    Smpi still fails, but this time because readlink does not accept -f
    Everything seems to work properly beside of that.
    
  * Exotic platforms:
    - AIX version 5.3 (only tested contexts this time)
      Smpi still fails there because mktemp is not installed. 
      Everything seems to work properly beside of that.
    - OpenSolaris 11
      I managed to compile it for the first time, but several breakages.
      Won't delay the release for this exotic platform.
    
  * Windows: it's still lagging behind. If you want to help, please
    stand up.

 Timing report of this version:
  This version seem to be more than 5% faster than 3.3.1 (on linux
    64bits with contextes). The gain is less than expected, we are
    investigating this for next release.

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr> Wed, 19 Aug 2009 17:07:12 +0200

SimGrid (3.3.1) stable; urgency=low

 OVERALL CHANGES:
  * Implement a --cfg-help to show existing configuration variables
  * Build chain do not require doxygen in maintainer mode

 GRAS:
  * fix a bug on struct sizeof computation, which prevented the
    exchange of arrays of structs in some conditions
    - added a regression test about this in datadesc_usage
  * Allow the exchange of 0-long dynamic vectors.
    - for that, use -1 as indicator of dynamic size instead of 0
    - This implied to change any size from unsigned long to long,
      reducing a bit communication abilities, but I guess that with
      64bits being quite common, this is more than enough.
    - This also induce a protocol change, thus bumping network protocol
      version from 0 to 1 (if we have external users, we have to get
      clean on that point too ;)
    - added two regression tests about this in datadesc_usage
  * Be more verbose when propagating local exceptions
    This helps debugging.
  * Display the status of simulated processes when receiving SIGINT in
    simulation mode

 MSG:
  * Allow to control the simulation from a trace file.
    New functions MSG_action_register() and MSG_action_trace_run()
    The first one allows to associate a function execution to each
     kind of action while the second one parses a trace file and
     triggers the corresponding actions within the system.
    For now, only a toy example is provided in examples/msg/actions
  * Add an exemple of process migration in examples/msg/migration
  * Fix a bug in task exchange which broke MSG_task_get_sender()
    Add a teshsuite regression test for that.
    [Bug: if MSG_task_get_sender() is called after sender exit,
     bad things happen]
  * Fix a bug which prevented suspend/resume to work properly
  * Display the status of simulated processes when receiving SIGINT
    This fixes a regression of v3.3. due to the introduction of SIMIX
  * Bug fixing in failure management:
    - trace could not start by a failure at time 0
    - failure during communications were not working
        
 SIMIX:
  * Add SIMIX_process_set_name() to change the name of the current
    process in the log messages.
  * Store smx_hosts in a dict since we only retrieve them by name
  * Move the configuration infrastructure to surf

 SIMDAG:
  * Move the configuration infrastructure to surf

 SMPI: 
  * Massive internal cleanups:
    - Store internal structures on processes instead of hosts (allows
      to have more than one process per host, in addition of being more 
      logical)
    - Cleanup the initialization/finalization process
    - Kill a whole bunch of unneeded synchronization: 
      processes run in exclusive manner within the simulator
    - Move queues from global tables to process data fields
  * Improve smpirun:
    - now accept -platform and -hostfile arguments
    - Pass the right rank value to processes according to the hostfile
  * Compile the examples by default, and use them as regression tests
  * Implement MPI_Wtime()
  * Change the reference speed to a command line option
  
 SURF:
  * TCP_gamma can now be specified as command line option using
    --cfg=TCP_gamma:10000000.0
  * Change the --surf-path cmd line option into --cfg=path:
  
 XBT:
  * Also include strbuff from xbt.h public header
  * xbt_ex_display(): do not free the exception after displaying 
    This allows to do more with the given exception afterward.
    Users should call xbt_ex_free() themselves.
    
    

 Portability report of this version:
  * Main portability targets:
    - Linux(debian)/x86/context   
    - Linux(debian)/x86/pthreads 
    - Linux(debian)/amd64/context 
    - Linux(debian)/amd64/pthreads
    These targets fail about 1/10 of times on gras/pmm, but we believe
      that this is because of the test, not because of simgrid.
    amok/saturate_sg fails even more rarely, and the test may not be
      the problem.
      
    - Mac OSX Leopard/x86/context
    The test suite still spits tons of errors because some obscure
      force prevents us from removing the temporary directories
      arguing that they still contain some metadata I've never heard of.
    Smpi fails because seq is not installed.
    Everything seems to work properly beside of that.
    
  * Exotic platforms:
    - AIX version 5.3 (both contexts and pthread)
      Smpi still fails there because mktemp is not installed. 
      XML inclusions seems rosty on AIX.
    
  * Windows: it's still lagging behind. If you want to help, please
    stand up.

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr>  Sat, 27 Jun 2009 00:14:30 +0200

SimGrid (3.3) stable; urgency=high

 OVERALL CHANGES:

  * JAVA BINDINGS for MSG (you dreamt of them? We made them)
    [Malek Cherier & Mt]

  * Introduce the SIMIX module: factorize code between MSG and GRAS.
    [Bruno Donassolo]
  
    Until now, GRAS were using MSG as an interface to SURF. It was
    quite difficult because both interface have several differences
    (MSG channels vs GRAS sockets were the most notable point).
   
    This also opens the gate to SMPI (which should occur soon) and speed
    up simulations by to 40% (even if it were not the main goal).
  
    ************************************** 
    *DO NOT MIX 3.2 RESULTS WITH 3.3 ONES* Simix may changes simulations!
    **************************************
    The point is that events occuring at the exact same timestamp are
    not scheduled in the same order with the old and new version. This
    may be enough to completely change the execution of simulations in
    some cases. Sorry for the inconvenience.

  * Cleanup and upgrade the XML format to push further scalability
    issues (check http://hal.inria.fr/inria-00256883/ for more info)

  * Improve the testing infrastructure with tesh. Now a very large part of
    the code is tested not only by being run but also by checking that the
    output match an expected output [Mt].

  * Move on to FleXML v1.7 for the embeeded XML parsers. This version
    is really less memory-demanding, which should allow you to use
    larger files in SimGrid [AL].
    
  * Inform valgrind about our contextes, so that it becomes usable
    with the default (and more effecient) version of SimGrid
    [contributed by Sékou Diakite, many thanks]

 GRAS:
  * Introduce a listener thread in charge of receiving incomming
    messages from the network. It allows to overlap communication and
    computation but most notably, it removes some stupid deadlocks due
    to the fact that so far, a process could not send and receive at
    the same time. This made most non trivial communication schema
    impossible.
  * Convert the PIDs from long int to int to match the MSG ones (and
    linux ones too) [Mt]
  * New function: gras_agent_spawn() to launch a new process on
    current host. Only working in simulation for now. [Mt]
  * New function: gras_os_hostport() returning a constant form (ie,
    not needing to be freed) of "gras_os_hostname():gras_os_myport()"

 XBT:
  * Make the backtrace of exceptions more human readable [Mt]
  * New module: xbt/str [Mt]
    a ton of string utility functions (split, join, printf to a newly
    allocated buffer, trim, etc)
  * New module: xbt/hash [Mt]
    SHA1 hashing algorithm (more to come if needed)
  * New module: xbt/synchro [Mt]
    synchronization tools (mutex and conditions) working the same way
    in simulation and in real life (mainly useful for GRAS, but not
    only).
  * New module: xbt/queue [Mt]
    classical producer/consumer synchronization scheme
  * xbt_dynar_new_sync() creates a synchronized dynar. All access
    (using the classical functions will get serialized) [Mt]
  * Make dictionary internal table dynamic. No need to specify its size
    anymore; functions xbt_dict_new_ext() and xbt_dict_hashsize_set()
    thus dropped. [Mt].
  * Make sure the log channels are organized as a tree under windows
    (because of ANSI C compatibility issue, any channel were child of
     root directly) [Mt].

 SURF:
  * Cleaned many thing in surf and fixed a few bugs [AL].
  * Add a nice command line configuration mechanism to compose models [AL].
  * Add a new model for parallel tasks (ptask_L07) that is less buggy than
    the previous one (KCCFLN05). It relies on something that looks like
    a max-min sharing mechanism but cannot be written as such. A new solver
    was thus designed [AL].
  * Add a new solver to lmm. Based on Lagrange optimization and
    gradient-based descent, it enables to efficiently maximise systems s.a
  
     sum f_i(x_i) s.t Ax<= b  with A_{i,j}>=0 and f_i a concave function.

    This solver enables to propose two new network models for TCP Reno and
    TCP Vegas based on Low's work. These models still need to be fully
    tested though [Pedro Velho].

 SIMDAG [AL]:
  * Bug fix in SD_simulate. Now the time bound given as argument is
    used.
  * Use the new parallel task model (ptask_L07) as default.
  * Use the SURF command line configuration mechanism.
  * 0-size tasks (for synchronization) should now work.

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr> Sun Apr 12 05:20:36 CEST 2009

SimGrid (3.2) stable; urgency=high

  OVERALL CHANGES:
   * Port to windows.
     We still experience issues on this platform, but we believe that at
     least MSG is usable.

  GRAS API BREAKAGE (for simplification purpose, sorry):
   * the gras_msgtype_by_name is not used anymore. Instead of 
       gras_msg_send(toserver, gras_msgtype_by_name("request"), &request);
     you can write (and must)
       gras_msg_send(toserver, "request", &request);
   - If you still want to pass a gras_msgtype_t to the function (to cache
     the type and avoid the lookup time), use the gras_msg_send_() variant.
   - Impacted functions:
     gras_cb_register, gras_cb_unregister, gras_msg_send, gras_msg_wait,
     gras_msg_rpccall, gras_msg_rpc_async_call, gras_msg_wait_ext
   * The callbacks are now expected to return 0 when everything went well
     (just like the main() function)

  GRAS new features and improvements:
  * New module mecanism where user code can use per process globals [Mt]
    This is similar to gras_userdata_*() functions, but for libraries. It
      factorize some code developped over and over in the examples and AMOK.
    It has still to be documented and used (only amok/peermanagement is
      converted for now).
  * Fix a vicious bug in the TCP buffering mecanism which leaded to message
    loss when they were small enough to fit into the buffer and sent quickly
    enough so that they can all get received in one shoot.   
  * gras_datadesc_by_name and gras_msgtype_by_name: now raise an exception
    if not found. Use the *_or_null() variant for the old semantic.
  * In gras_msg_handle, do not discard messages without callback.
    They are probably messages to be explicitly awaited later (ie, proofs of
    mis-synchronization in userland since they are sent before being awaited)
    No big deal usually.
  * gras_socket_meas_send/recv: semantic changed!
    The numerical arguments used to be (1) the total amount of data to send
    and (2) msg_size. This was changed to (1) msg_size and (2) amount of
    messages. This was need for the fool willing to send more than MAXINT
    bytes on quite fat pipes.       
        
  AMOK:
  * Do really rename the hostmanagement module to peermanagement. [Mt]
    Ie, rename functions from amok_hm_* to amok_pm_*. This breaks the API,
    but this is rather new and this was documented in the module
    documentation (poor excuses, I admit)
  * Bandwidth measurement semantic changed! This follows the changes to
    gras_socket_meas_send/recv explained above.
    
  SIMDAG:
  * A sequential mode has been added to the workstations. When a workstation
    is in sequential mode, it can execute only one task, and the other tasks
    are waiting in a FIFO. [Christophe Thiery]

  SURF:
  * The KCCFLN05 workstation model now handles parallel tasks. It is the
    model for SIMDAG. [Christophe Thiery]
  * Bug fix in the maxmin solver: Some values were close to 0 instead of
    equal to 0, which caused some bad behaviors in
    saturated_constraint_set_update. I now use a threshold mechanism like in
    surf. [AL]

  XBT:
  * When running manually src/testall, you select specific units [Mt]
    testall is the result of our cunit mecanism, and should replace all
    the scripty thingy around since bash don't run easily on billware.

  * A mallocator system has been added. [Christophe Thiery]
    Mallocators allow you to recycle your unused objects instead of freeing them
    and allocating new ones.

  Documentation update:
  * FAQ reworking + New FAQs:
    - "Valgrind spits tons of errors!" [Mt]
    - "How to repport bugs" [Mt]
    - "Cross-compiling a Windows DLL of SimGrid from Linux" [Mt]
    - "What is the difference between MSG, SimDag, and GRAS?" [Mt]
    - Communication time measurement within MSG [AL]
    - I experience weird communication times when I change the latency [AL]
  * GRAS tutorial [Mt]
    It contains:
     - an introduction to the framework and to the used communication model
     - an initiatic tour introducing the most proheminent features:
       o Part 1: Bases
         . Lesson 0: Installing GRAS
         . Lesson 1: Setting up your own project
       o Part 2: Message passing
         . Lesson 2: Exchanging simple messages
         . Lesson 3: Passing arguments to the processes (in SG)
         . Lesson 4: Attaching callbacks to messages
         . Lesson 5: Using globals in processes
         . Lesson 6: Logging informations properly
         . Lesson 7: Using internal timers
         . Lesson 8: Handling errors through exceptions
         . Lesson 9: Exchanging simple data
         . Lesson 10: Remote Procedure Calling (RPC)
         . Lesson 11: Explicitely waiting for messages
         . Recapping of message passing features in GRAS
     - A HOWTO section containing:
       o HOWTO design a GRAS application
       More are due, of course. They will come latter. In the meanwhile, you can 
       check the examples which are still here.

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr> Fri Mar 16 21:11:46 CET 2007

SimGrid (3.1) stable; urgency=high

  General:
  * Port to gcc 4.x   
    There was a stack corruption somewhere, visible only when optimizing
    with these versions. [Vince]

  SIMDAG:
  * This is a NEW module! SimDAG (SD for short) is a revival of the old SG
    module that enabled to play with Directed Acyclic Graphs. It is built
    directly on top of SURF and provides an API rather close to the old
    SG. Some old codes using SG are currently under rewrite to check that
    all needful functions are provided. [Christophe Thiery]
        
  SURF:
  * Complete rewrite of the KCCFLN05 workstation model. It is now an
    extension of the classical CLM03 model that gracefully handles
    failures. This is now the default model for MSG and GRAS. It doesn't
    handle parallel tasks yet though. [AL]
  * Bug fix: Weights were not correctly set in the network part. 
    WARNING: This may have resulted in incorrect results with simulations
    where there are more than one flow on a given link. [AL]

  SURF, MSG, GRAS:
  * After a (long ?) discussion on simgrid-devel, we have decided that the
    convention we had on units was stupid. That is why it has been decided
    to move from (MBits, MFlops, seconds) to (Bits, Flops, seconds). 
    WARNING : This means that all previous platform files will not work as
    such with this version! A warning is issued to ask users to update
    their files. [AL]
    A conversion script can be found in the contrib module of the CVS, under
    the name contrib/platform_generation/surfxml_update.pl [MQ]

  MSG,GRAS:
  * Bug fix: Processes were started in reverse order, wrt deployment file.
    WARNING: if your code relies on this bug, please fix it.    [AL]
  * Bug fix: Add a test in MSG_task_execute to stop whenever a task is
    being executed on two different locations.                  [AL]
  * Bug fix: Failures are now better supported thanks to Derrick's tests
    (there was many failure situations I hadn't thought of and that weren't
    correctly handled). [AL]
  * New function: MSG_host_is_avail indicates you whether a given m_host_t
    is up or down. [AL]

  GRAS:
  * New! a real RPC mecanism, as it ought to be since too long. [MQ]
      Exception occurring on server-side are propagated back to client (!).
      
    API CHANGE: the callback changed their prototype. Change:
        int my_handler(gras_socket_t expeditor, void *payload_data) {
      to:
        int my_handler(gras_msg_cb_ctx_t ctx  , void *payload_data) {
          gras_socket_t expeditor=gras_msg_cb_ctx_from(ctx);
      and you're set.
  * New! function: gras_msg_handleall to deal with all messages arriving
      within a given period.
  * New! function: gras_socket_server_range to get a server socket in a
    range of port numbers (ease to avoid port number conflicts) [MQ]
  * New! gras processes display their backtrace when they get a SIGUSR1
      or when Ctrl-C is pressed. Use Ctrl-C Ctrl-C to exit.
      Sweet to debug RL processes [MQ]

  AMOK:
  * Bandwidth module: 
    - Do not force experiment sizes to be expressed in kb, or it becomes
      impossible to measure the latency this way (needs one byte-long tests)
    WARNING: this changes the amok_bw_* function semantic. [MQ]
    - Implements the link saturation stuff. [MQ]
  * Peer management module: 
    New! module factorizing code that we wrote over and over [MQ].
      
  XBT:
  * New module: cunit (my jUnit implementation in ansi C) [MQ]
    - Test units are placed directly into the library code, they get extracted
      automatically and placed into the src/testall binary.
    - Convert most of the XBT tests to this system.
  * New functions: xbt_dynar_getfirst_as() and xbt_dynar_getlast_as() [MQ]
  * XML parsing: rewrote parts of flexml to enable multiple xml parsers to
    live in the same C code. This required to change a little bit the API
    of surfxml parsing but shouldn't be an issue for end-users. [AL]
  * New module: sparse graph structure with basic algorithms (this is work
    in progress and the API is not considered to be frozen yet). [AL]
  * Display more information on backtraces: source line & function names are
    now displayed just like valgrind does (rely on addr2line tool) [MQ]
  * New function: xbt_backtrace_display(). Sweet while debuging [MQ]
  * Reworked a little bit some #include statements to load only required
    headers. Some user code that relied on SimGrid to include stdlib or
    stdio may need to include it by themselves. [AL]
  * Fixed xbt/log.h. A missing SG_BEGIN_DECL prevented compilation with
    g++. [AL]
  * Renamed xbt_host_t into xbt_peer_t since it betterly describes what I
    meant. This breaks the API of AMOK and of xbt/config. Sorry about this,
    but I guess that almost nobody used those parts. [MQ]

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr> Fri, 14 Jul 2006 01:32:27 +0200

SimGrid (3.0.1) stable; urgency=low

  XBT:
  * Unfortunately, I had missed 5 misnamed functions:
      xbt_fifo_item_t xbt_fifo_newitem(void);
      void xbt_fifo_freeitem(xbt_fifo_item_t);
      xbt_fifo_item_t xbt_fifo_getFirstItem(xbt_fifo_t l);
      xbt_fifo_item_t xbt_fifo_getNextItem(xbt_fifo_item_t i);
      xbt_fifo_item_t xbt_fifo_getPrevItem(xbt_fifo_item_t i);
    They're now deprecated. Please use their new versions:
      xbt_fifo_item_t xbt_fifo_new_item(void);
      void xbt_fifo_free_item(xbt_fifo_item_t);
      xbt_fifo_item_t xbt_fifo_get_first_item(xbt_fifo_t l);
      xbt_fifo_item_t xbt_fifo_get_next_item(xbt_fifo_item_t i);
      xbt_fifo_item_t xbt_fifo_get_prev_item(xbt_fifo_item_t i);
    [AL]
  * Bugfix: really disconnect fifo items which are remove_item()ed [AL]
  * Documentation: xbt_log module unmercifully reworked [MQ]
  * Bugfix: there was a problem with the ending of contexts with 
    the pthread backend. It caused some weird deadlock or behavior
    depending on the pthread implementation. [AL]
  * Bugfix: get the exceptions raised in the simulator repport where
    and why they come from when they are not catched in time [AL, MQ]

  SURF:
  * Bugfix: Do repport the error when two non-connected hosts try to
    exchange data (Thanks to Flavien for stumbling into this one) [AL]
  
  SURF:
  * Add additionnal checkings on communications. Assert that two
    communicating hosts are connected by a set of links... [AL]
        
  MSG:
  * Add additionnal checkings on channel values in communication [AL]
  * New: MSG_task_get_source to see on which host a task was generated [HC]
  * New: int MSG_task_probe_from_host(int channel, m_host_t host): returns
    the number of tasks waiting to be received on channel and sent
    by host. [AL]
  * New: MSG_error_t MSG_task_get_from_host(m_task_t * task, int channel, m_host_t host); 
    waits for the first task coming from a given host.. [AL]
        
  GRAS new functionnalities: [MQ]
  * Enhance the parsing macro to allow the size of multidimentional objects
    to be given thru annotations.
  * New example (and documentation): Matrix Multiplication a la RPC 
    (as when I was young!) and fix a bunch of bugs found on the way.

  GRAS performance improvements: [MQ]
  [DataDesc]
  * Reduce the amount of cbps creation/destruction by making it static to 
    datadesc_send/recv() and using a (newly created) cbps_reset (based on 
    dynar_reset ())
  [Virtu]
  * Change libdata to a set so that we can search for stuff by ID (and thus 
    reduce the insane amount of dict lookups)     
  
  [Transport]
  * Actually implement gras_datadesc_copy() so that we don't have to mimick
    RL communication on top of SG since it's so uneffective. 
    It may also be used for inter-thread communication in RL, one day. 
  * Use gras_datadesc_copy() to exchange messages on top of SG 
    Allows to:
    - improve message exchange performance on top of SG
    - deprecate transport_plugin_sg.c:gras_trp_sg_chunk_send() & recv()
  * Don't exchange on the network the size of the used part of buffer,
    instead, specify the possible buffer size to read(). 
    Advantages:
     - reduces the amount of read/write calls (one pair per exchange)
     - reduces the amount of exchanged data (the size)
     - allows to retrieve all arrived data on receiver side, if we don't need
       it right now (subsequent read will peek the buffer)
     - allows the receiver to proceed with the begining of the stream before
       everything is arrived
     - make it possible to build an iov transport (using readv/writev)
    Extra difficulty: 
     - take care of the data with non-stable storage (like stacked data),
       and bufferize them.
  * If possible, TCP send uses vector I/O (when writev() is here) 
     - Don't use it for receive since we send data sizes and data on the
       same stream, so we wouldn't be able to chain large amount of chunks
       before having to flush the stuff to read the size.
  * Rework the transport plugin mecanism to simplify it and reduce the
    amount of pointer dereferencement when searching for the right function 
    to use. 

  * I guess that now, we do almost as few system calls as possible while
    doing as few data copy as possible.

    To improve it further, we could try to send all the sizes first and then
    all the data (to use iov on receiving size), but it's only a partial
    solution: when you have 2 dimensional data, the sizes of the second
    dimension is data of the first dimension, so you need 3 streams.

    I'm not sure the potential performance gains justify the coding burden.

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr>  Fri, 21 Oct 2005 14:42:20 +0200

SimGrid (3.00) stable; urgency=high
  
 SURF:
  * New! Give the possibility to hijack the surf parser and thus bypass 
    MSG_create_environment and MSG_launch_application. Have a look at
    examples/msg/msg_test_surfxml_bypassed.c to see how it can be done.
        
 -- Arnaud Legrand <simgrid-devel@lists.gforge.inria.fr>  Sat, 20 Aug 2005 23:25:25 -0700

SimGrid (2.96) unstable; urgency=low

  AKA SimGrid 3 rc 2.
  
  XBT:
  * New! Exception handling with setjmp or such (code from OSSP ex) [MQ]
    This deprecates the xbt_error_t mecanisms. 
    It modifies (simplifies) all XBT and GRAS API.
    MSG API keeps unchanged (exceptions raised by XBT are catched from 
     within MSG and masked with existing error handling facilities)

  SURF:
  * New! Add a FATPIPE model. [AL]
  * New! Add a parallel task model. [AL]
  * New! Add automatically a loopback interface (in the default
    network model) if none was precised.

  MSG
  * Bugfix: MSG_process_resume now works with the current running process. 
    [AL]
  * New! Add MSG_parallel_task_create and MSG_parallel_task_execute. [AL]
  * Modification of MSG_task_get_compute_duration. Once a task has been
    processed, the value returned by this function is now equal to 0. [AL]
  * New! Add double MSG_task_get_remaining_computation(m_task_t task) and
    MSG_error_t MSG_task_cancel(m_task_t task). Add a state
    (MSG_TASK_CANCELLED) to MSG_error_t corresponding to the cancelation
    of a m_task. For now, MSG_task_cancel only works with computation
    tasks. [AL]
  * New! Add double MSG_get_host_speed(m_host_t h) that returns the speed
    of the processor (in Mflop/s) regardless of the current load on the
    machine. [AL]
  * API Change: use proper naming convention for MSG_getClock and 
    MSG_process_isSuspended: MSG_get_clock and MSG_process_is_suspended.
    [AL]
  * New! Add void MSG_task_set_priority(m_task_t task, double priority). 
    This function changes the priority of a computation task. This priority
    doesn't affect the transfer rate. A priority of 2 will make a task 
    receive two times more cpu power than the other ones. This function 
    has been added to suit the needs of Nguyen The Loc and hasn't been that
    much tested yet. So if it fails, please report it and send me your code. 
    [AL]
  * API Change: removed all functions and types that were marked "deprecated" 
    since many months. Renamed MSG_global_init_args to MSG_global_init.

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr>  Mon,  8 Aug 2005 17:58:47 -0700

SimGrid (2.95) unstable; urgency=low

  XBT
  * Steal some nice code to GNU pth to fix context detection and usage [AL]
  * Cleanup in the xbt_config API; add configuration callbacks. [MQ]
  * Cleanup in the initialization API: the unused "defaultlog" is dead. [MQ]

  SURF
  * Bugfix: Allow absolute paths for platform description files [MQ]
  * Bugfix: do free the variables after use. Leads to drastic performance 
    improvement [AL] 
  * Implement max_duration (ie, timeouts) on resources [AL]

  MSG
  * Implement MSG_config to configure MSG at runtime. xbt_cfg test on a real
    case ;) [MQ]
  * Implement MSG_channel_select_from() to help GRAS now that SURF provide
    the needed support (timeouts) [AL]

  GRAS (new features)
  * Implement measurement sockets. You can now get the bandwidth between two
    hosts thanks to AMOK (see below). [MQ]
  * gras_datadesc_dynar() builds a dynar type descriptor, allowing to send
    dynar over the network (yeah) [MQ]
  * Real (even if simplistic) implementation of gras_os_myname() on RL [MQ]
  * simple/static token-ring example. [Alexandre Colucci and MQ]
  * Use MSG_channel_select_from() and remove the *slow* hack we had to put
    in place before [MQ]
  
  GRAS (bug fixes)
  * Differentiate the types "char[22]" and "unsigned char[22]" in automatic
    type parsing. "short" and "long" modifiers were also ignored; other
    modifier (such as reference level) are still ignored. [MQ] 
  * Embeed the buffer size within the buffer itself on SG. [MQ]
    That way, send() are atomic and cannot get intermixed anymore (at least
    the ones which are less than 100k; bigger messages still have the issue)
  * Array size pushed by the field, not by the field type (or each
    and every long int will push stuff to the cbps) [MQ]
  * use select() to sleep since it allows to portably sleep less than one
    second. [MQ]

  GRAS (minor cleanups)
  * <project>.Makefile.local (generated from gras_stub_generator) |MQ]:
    - Do clean .o files
    - Compile with -g
  * Type Callbacks now receive the gras_datadesc_type_t they work on as argument.
  * type category 'ignored' killed as it was never used and were difficult
    to transmit.
  * whether a type can cycle or not is now a flag, leaving room for more
    flags (as "ignored", if we feel the need one day ;)
  * Rename raw sockets to measurement sockets since "raw" has another
    meaning in networking community. 
  
  AMOK 
  * Advanced Metacomputing Overlay Kit introduction. It is based over GRAS
    and offers features not belonging to GRAS but that most applications
    need. One day, it may be a set of plugins loadable at runtime.
  * New module: bandwidth 
    bandwidth measurement between arbitrary nodes running this module. [MQ]

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr>  Thu, 30 Jun 2005 16:29:20 -0700

SimGrid (2.94) unstable; urgency=low

  The first beta release of SimGrid 3 !

  >>>Arnaud<<<
  (documentation)
  * Update the main page and the FAQ. Adding references to gforge.

  (gras)
  * Add a gras_os_getpid function.

  (msg)
  * Add MSG_task_get_compute_duration() and MSG_task_get_data_size()
  * Extend the logs so that they also print PID, hostname, date, ... if
    available.
  * Convert the MSG example to the use of xbt_logs instead of PRINT_MESSAGE,
    and kill the old version which were in testsuite/
  * Rewrite tools/MSG_visualization/colorize.pl for using with logs instead
    of PRINT_MESSAGE

  (xbt)
  * Add xbt_os_time(). As the rest of xbt/portability, this is not public
    for users. Instead, each programming environment (GRAS, MSG,...) use it
    when needed to provide such a feature to users.
    Don't shortcut the mecanism or you will also shortcut the virtualization
    you need on the simulator.

  >>>Martin<<<
  (infrastructure)
  * Cleanups in configury with regard to compile optimization/warning flags.
    Also add -fno-loop-optimize to any powerpc since it's the optimization
    killing gcc (< 3.4.0).
  * Doxygen cleanups: move MSG examples, kill the second Doxygen phase
    needed by MSG examples complications
  * Borrow configury beautifications from PHP

  (xbt)
  * Bugfix: XBT_LOG_NEW_DEFAULT_CATEGORY now compiles without compiler
    warning (thanks loris for stumbling into this one).
  * Bugfix: stop loading private headers (gras_config.h) from the public
    ones (xbt/swag.h).

  (gras)
  * Change SIMGRID_INSTALL_PATH to GRAS_ROOT in Makefiles generated for user.
  * Rename gras_get_my_fqdn to gras_os_myname and implement it in the simulator
    RL would imply a DNS resolver, which is *hard* to do in a portable way
    (and therefore delayed).
  * Implement a real timer mecanism and use it in timing macros. This allows
    to avoid rounding errors and get a 0.000005 sec precision in timing
    macros. While I was at it, various cleanups:
     - allow to declare more than one timed section per file (fix a stupid bug)
     - move some private declaration to the right place
     - merge conditional execution and timing macros into emulation module
     - document the module
     - make sure the module cleanups its mess on gras_exit
  * Documentation improvements:
     - (new) how to compile applications using GRAS
     - (new) emulation support (timing macros)

 -- Da SimGrid team <simgrid-devel@lists.gforge.inria.fr>  Fri, 13 May 2005 10:49:31 +0200

SimGrid (2.93) unstable; urgency=low

  Alpha 4 on the path to SimGrid 3 (aka the "neuf-trois" version)

  [Arnaud]
   - Use Paje properly where used. Still to be sanitized properly.
   - Portability fix: Add an implementation of the contexts using pthread

  [Martin]
  (misc)
   - Add xbt_procname(): returns the name of the current process.
     Use it to show the current process's name in all logging.
  (infrastructure)
   - fix detection of older flex version and the reaction, since we do
     depend on modern ones (we use lex_destroy)
   - Better separation of SG and RL in the libs: remove all simulation code
     from libgras. As a result, this lib is now only 200k when stripped.
     Some of the xbt modules may also be duplicated (two sets and such) and
     should be cleaned/killed before SG3.
   - Insist on using xlC on AIX because of weird problems involving gcc there.
   - Cleanup the make remote stuff. This is now done by scripts
     tools/graspe-{master,slave} (GRAS Platform Expender). This is still
     mainly for our private use, but we're working on changing them to user
     tools, too.
  (gras)
   - Bugfix: flush the socket on close only if there is some *output*.
   - Bugfix: flush idempotent when there's nothing to send (don't send size=0)
  (msg)
   - Add MSG_task_get_name. The task names are mainly for debugging purpose,
     but anyway.

 -- SimGrid team <simgrid2-users@listes.ens-lyon.fr>  Fri,  4 Mar 2005 14:32:37 -0800

SimGrid (2.92) unstable; urgency=low

  Alpha 3 on the path to SimGrid 3
  
  [Arnaud]
  (gras)
   - New! First try of benchmarking macros.
   - New! First try so that gras_stub_generator generate deployment and
     remote compilation helpers.
  (msg)
   - Bugfix: Initialization fix in msg_test.

  [Martin]
  (surf)
   - Bugfix: applied patch to lexer so that it doesn't need a huge heap.
  (xbt)
   - Bugfix: let dicts work with NULL content (_foreach didn't) and cleanups
  (gras)
   - API Change: gras_os_sleep to take the amount of seconds as a double.
     Accepting an int was error prone since it was the only location where
     seconds were coded as such. It leaded to damn rounding errors.
   - Bugfix: Hard to belive that timers ever worked before this.

 -- SimGrid team <simgrid2-users@listes.ens-lyon.fr>  Wed, 23 Feb 2005 22:09:21 +0100

SimGrid (2.91) unstable; urgency=low

  Alpha 2 on the path to SimGrid 3
  
  [Arnaud]
  (surf)
   - Bug fix in the lmm_solver.
  (msg)
   - New! Interface to Paje (see http://www-id.imag.fr/Logiciels/paje/) 
     through the function MSG_paje_output.
   - New! Introducing two new functions MSG_process_kill() and MSG_process_killall().
   - It is possible to bound the rate of a communication in MSG with 
     MSG_task_put_bounded() (was already in the previous version but I had forgotten 
     to write it in the changelog).
   - Bug fix to let GRAS run on top of MSG until we move it directly on top
     of the SURF.
    
  [Martin]
  (infrastructure)
   - Various cleanups to the autotools stuff
   - Begin to move Gras examples to examples/gras/
   - Let make distcheck work again (yeah!)
  (documentation)
   - documentation overhauled using doxygen. 
     gtk-doc-tools is dead in SimGrid now.
   - Automatically extract all existing logging categories, and add the list
     to the documentation (long standing one, to say the less)
  (gras)
   - Cleanup the known architecture table. Reorder the entries to group what
     should be, and use a more consistent naming scheme.
     (some of the test dataset are still to be regenerated)
   - New! Allow library to register globals on each process just as userdata
     does. 
      This is implemented using a xbt_dict and not a xbt_set, so we loose the
       lookup time (for now).
      Use it in msg and trp.
      This cleans a lot the internals and helps enforcing privacy of the
       headers between the gras components.
   - New! Add a timer mechanism, not unlike cron(8) and at(1). 
   - Bugfix: gras_os_time was delirious in RL.
   - Bugfix: gras_trp_select/RL don't run into the wall when asked to select
     onto 0 sockets.
   - Reenable GRAS now that it works.

 -- Arnaud Legrand <Arnaud.Legrand@imag.fr>  Mon, 14 Feb 2005 14:02:13 -0800

SimGrid (2.90) unstable; urgency=low

  Alpha 1 on the path to SimGrid 3

  * It is a long time since the last release of SimGrid. I'm sorry about
    that but as I had told you, I was rewriting a lot of things. I apologize 
    to those who had been reporting bugs to me and that I had not answered. 
    If your bug is still in the new version, please tell me. Here is a 
    summary of the main changes.

  * REVOLUTION 1: The SimGrid project has merged with the GRAS project
    lead by Martin Quinson. As a consequence SimGrid gains a lot in
    portability, speed, and a lot more but you'll figure it out later. 
    SimGrid now comprises 3 different projects : MSG, GRAS and SMPI. 
    I wanted to release the new MSG as soon as possible and I have 
    broken GRAS, which is the reason why, for now, only MSG is fully 
    functional. A laconic description of these projects is available 
    in the documentation.
  
  * REVOLUTION 2: I have removed SG and I am now using a new simulation
    kernel optimized for our needs (called SURF but only the developers
    should use it). Hence, MSG is now roughly 30 times faster and I think
    that by rewriting a little bit MSG, I could event speed it up a little
    bit more. Beside the gain in speed, it is also much easier to encode a
    new platform model with SURF than it was with SG. More to come...
  
  * REVOLUTION 3: I have tried to change a little as possible the API of
    MSG but a few things really had to disappear. The main differences
    with the previous version are :
       1) no more m_links_t and the corresponding functions. Platforms are
         directly read from a XML description and cannot be hard-coded
         anymore. The same format is used for application deployment
         description. The new format is described in the documentation. 
         Have a look in tools/platform_generation. There is a tiny script 
         that converts from the old platform format to the new one. Concerning
         the application deployment format, parsing the old one is tricky.  
         I think most of you should however be able to convert your files.  If 
         it is really an issue, I can write a C code that does the conversion. 
         Let me know.
       2) the toolbox tbx does not exist anymore. We now have a library
          with much more data-structures but without the hash-tables (we have 
          dictionaries that are much faster).

 -- Arnaud Legrand <Arnaud.Legrand@imag.fr>  Mon, 31 Jan 2005 10:45:53 -0800

*****************************************************************************
* Follows the old GRAS changelog. It does not follow the same syntax, but I *
* don't feel like converting the oldies. (Mt)                                *
*****************************************************************************

2005-01-31 Arnaud
  Version 2.90: "the long awaited one"
  - Finished rewriting and debugging MSG. Rewrote the documentation.
  - disable GRAS for now since it needs to be ported to the newest SG

2004-12-16 Martin
  - Finish the port to windows (using mingw32 for cross-compile)

2004-11-28 Arnaud
  - Main loop and datastructures of SURF. A cpu resource object is
    functional. Surf can thus be used to create cpu's with variable
    performance on which you can execute some actions.
        
2004-11-15 Martin Quinson
  - Port to ARM. Simply added the alignment and size descriptions. Should
    work, but the ARM machines are so slow that I didn't had the opportunity
    to 'make check' over there yet.

2004-11-15 Arnaud Legrand
  - Trace manager now written. It uses a heap structure and is therefore
    expected to be efficient. It may however be speeded up (particularly
    when many events occur at the same date) by using red and black
    trees. One day maybe... 
  - Max-min linear system solver written. It uses a sparse matrix
    structure taking advantage of its expected use. Most operations are
    O(1) and free/calloc are called as few as possible. The computation of
    the minimum could however be improved by using a red and black tree
    (again ! ;).

2004-11-03 Arnaud Legrand
  - Rename every gras_* function that was in xbt/ to its xbt_
    counterpart.
  - Add a heap and a doubly-linked list to xbt
  - Added a dichotomy to the dictionaries. make check works as well before
    so I assume that the patch is correct. I do not know however if things
    run effectively faster than before now. :)

  Inclusion of the SimGrid tree in the GRAS one. The archive is renamed to
  SimGrid, and the version number is bumped to 2.x

2004-10-29 Martin Quinson
  - Introduction of the remote errors. 
    They are the result of a RMI/RPC on the remote machine.
    ErrCodes being scalar values, you can't get the host on which those
    errors did happen. Extending the error mechanism as in Gnome is possible.
    No idea yet whether it is a good idea.
    
2004-10-28 Martin Quinson
  - Interface revolution: the Starred Structure Eradication.
    I used to do typedef struct {} toto_t; and then handle *toto_t.
    Arnaud (and Oli) didn't like it, and I surrendered. Now, you have:
      - ???_t is a valid type (builded with typedef)
      - s_toto_t is a structure (access to fields with .)
      - s_toto   is a structure needing 'struct' keyword to be used
      - e_toto_t is an enum
      -   toto_t is an 'object' (struct*)
    Exemple:
      typedef struct s_toto {} s_toto_t, *toto_t;
      typedef enum {} e_toto_t;
    Moreover, only toto_t (and e_toto_t) are public. The rest (mainly
     s_toto_t) is private.
    
  - While I was at it, all gras_<obj>_free() functions want a gras_<obj>_t*
    so that it can set the variable to NULL. It was so for dicts and sets,
    it changed for dynars.
    
  - Fix a bunch of memleaks in dict_remove
  - Fix a bug in sg/server_socket opening: it failed all the time.

2004-10-07 Martin Quinson
  - Speed up dynar lookup operation a bit.
  
    gras_dynar_get is dead. 
    
    Now, you can choose between gras_dynar_get_cpy (the old gras_dynar_get
    but should be avoided for efficiency reasons) and gras_dynar_get_ptr
    (which gives you the address of the stored data).
    
    gras_dynar_get_as is an helpful macro which allows you to retrieve a
    copy of the data using an affectation to do the job and not a memcpy.
    
    int toto = gras_dynar_get_as(dyn,0,int); rewrites itself to
    int toto = *(int*)gras_dynar_get_ptr(dyn,0);
    
    It does not really speedup the dynar test because they are
    setting elements all the time (and look them seldom). But the dict does
    far more lookup than setting.

    So, this brings the dict_crash test from ~33s to ~25s (200000 elms).

2004-10-05 Martin Quinson
  - Allow to (en/dis)able the cycle detection at run time.
  
    Whether we should check for cycle or not is now a property of each
    datatype. When you think there may be some cycle, use datadesc_cycle_set.
    datadesc_cycle_unset allow to remove this property when previously set.
    
    Note that the cycle detection is off by default since it impacts the 
    performance. Watch the data you feed GRAS with ;)
    
    This property is hereditary. Any element embedded in a structure having it
    set have it set for the time of this data exchange.
    
    You should set it both on sender and receiver side. If you don't set it on
    sender side, it will enter an endless loop. If you forget on receiver
    side, the cycles won't be recreated after communication.
    
  - Header reorganization.
    Kill gras_private.h, each submodule must load the headers it needs.

2004-10-04 Martin Quinson
  - Interface revolution: do not try to survive to malloc failure.
  
    Now, gras_malloc and friends call gras_abort() on failure.
    As a conclusion, malloc_error is not a valid error anymore, and all
      functions for which it was the only gras_error_t return value are
      changed. They now return void, or there result directly. 
    This simplify the API a lot.

2004-09-29 Martin Quinson
  - Re-enable raw sockets.
    Created by gras_socket_{client,server}_ext;
    Used with gras_raw_{send,recv}
    No select possible.
    
    It should allow to kill the last bits of gras first version soon.
  
    This is not completely satisfactory yet (duplicate code with
     chunk_{send,recv}; a bit out of the plugin mechanism), but it should
     work. 

  - Simplify transport plugin (internal) interface by not passing any
    argument to _server and _client, but embedding them in the socket
    struct directly. 

2004-09-28 Martin Quinson
  - Finish the port to AIX.
    autoconf was my problem (segfault within the malloc replacement
    function. No idea why)
        
2004-09-16 Martin Quinson
  - Fix some size_t madness on 64bit architectures.
  
2004-09-08 Martin Quinson
  - Reduce the number of system headers loaded, overload some more system
    calls (such as malloc to cast the result of the system one, and work
    properly on AIX)
  - Fix and reintroduce the config support

2004-09-07 Martin Quinson
  - Source code reorganization to allow Arnaud to surf all over there.
  - Allow to document the logging categories.
  - Remove all uppercase from logging categories and useless cleanup in names.

2004-08-18 Martin Quinson
  Version 0.6.2 (protocol not changed; API changed)
  - Interface cleanup: gras_msgtype_by_name returns the type (instead of a
     gras_error_t), and NULL when not found. Functions expecting a msgtype
     as argument (msg_wait; msg_send) deal with NULL argument by providing a
     hopefully usefull message.
  - Portability to prehistoric sparcs again

2004-08-17 Martin Quinson
  Version 0.6.1 (protocol not changed; ABI not changed)
  - prealloc some buffers to speed things up

2004-08-11 Martin Quinson
  Version 0.6 (protocol not changed; ABI expended)
  - The parsing macro can deal with the references, provided that you add
    the relevant annotations (using GRAS_ANNOTE(size,field_name))

2004-08-09 Martin Quinson
  Version 0.5 (protocol not changed; ABI changed)
  - Allow to off turn the cycle detection code in data exchange at
    compilation time. It should be at run time, but I'm short of time (and
    the config stuff is still broken). That way, we keep dict out of the
    critical path, which is good because the performance is poor:
     - search not dichotomial yet
     - dynar give no way to access their content and memcpy everytime
  - In composed data description (struct, ref and so on), stop foolness of
    keeping the subtype's ID, but store the type itself. This keeps sets out
    of the critical path, which is good since they rely on dynar and
    dictionnaries. The only loose of that is that we cannot detect the
    redeclaration of a structure/union with another content (but I'm not sure 
    the code detected well this error before anyway). We still can detect
    the redefinition discrepancy for the other types.
  - Use a whole bunch of optimisation flags (plus -fno-strict-aliasing since
    it breaks the code because of type-punning used all over the place).
    This breaks on all non-gcc architectures (for now).
    
  All those changes (plus the buffer of last time) allow me to gain 2 order
  of magnitude on cruel tests consisting of 800000 array of integers on two
  level of a hierarchical structure (200 secondes -> 4 secondes)
  
  API change:
    - the selector of reference must now return the type it points to, not
      the ID of this type.

2004-08-06 Martin Quinson
  Version 0.4 (protocol changed; ABI not changed)
  - Allow to pass --gras-log argument to processes in simulation mode. Really.
  - New debugging level: trace (under debug) to see effect of GRAS_IN/OUT
  - Implement a buffer transport, and use it by default (it relies on tcp in
     real life and on sg in simulation).
    That's a bit hackish since I had a new field to the structure to store
     its data without interfering with the subtype ones. Inheritance
     is tricky in C. And that's a kind of reverse inheritance with one class
     derivating two classes. Or maybe a game with java interfaces. Anyway,
     that's damn hard in C (at least).
    Moreover, I got tired while trying to ensure plugin separation and
     genericity in SG mode. MSG wants me to do weird things, so let's go for
     cruel hacks (temporarily of course ;).
     See comment in transport_private.h:71
  - do not include all the _interface headers in private but in the files
    which really need them (to cut the compilation time when they are
    modified) 

2004-07-26 Martin Quinson
  Version 0.3 (protocol not changed; ABI changed)
  - Major overhault of the datadesc interface to simplify it:
    - shorted the function names:
      s/gras_datadesc_declare_struct/gras_datadesc_struct/ and so on
    - add a trivial way to push/pop integers into the cbps without malloc.
      This allows to make really generic sub_type description, which simply
        pop their size of the stack.
    - add a function gras_datadesc_ref_pop_arr() which does what users want
      most of the time: Declare a dynamic array (which pops its size of the
      stack) and declare a reference to it. Poor name, but anyway.
    - kill the post-send callback, add a post-receive one
    
2004-07-23 Martin Quinson
  Version 0.2 (protocol changed; ABI changed)
  - add some testing for cpbs in the test cases, and fix some more bugs.
    This invalidate again the little64 data file, since I cannot regenerate
    it myself.
  - remove an awfull optimization in the logging stuff, allowing me to:
    - understand it again
    - learn gcc how to check that the argument match the provided format
    - fix all errors revealed by gcc after that
  - internal keys of dict are not \0 terminated. Deal with it properly in
    loggings instead of segfaulting when the user want to see the logs :-/

2004-07-22 Martin Quinson
  - Fix some stupid bug preventing cbps (callback postit) from working

2004-07-21 Martin Quinson
  - Some documentation cleanups
  - remove the useless last argument of msgtype_declare
  - rename the Virtu functions to fit into the 'os' namespace
  - move headers src/include -> src/include/gras/ and stop fooling with 
    gras -> . symbolic link
  - make distcheck is now successful

2004-07-19 Martin Quinson
  Version 0.1.1
  - Build shared library also
  - Install html doc to the right location
  - stop removing maintainer files in make clean
  - build tests only on make check
  
2004-07-13 Martin Quinson
  version 0.1
  - No major issue in previous version => change versionning schema
  - Re-enable little64 convertion test now that Abdou kindly regenerated the
    corresponding dataset.
  
2004-07-11 Martin Quinson
  version 0.0.4
  - Get it working with any kind of structure (we can compute the padding
    bytes remotely for all the architectures I have access to)
  - Implement the structure parsing macro (still not quite robust/complete)
  - Improvement to the remote testing toysuite
  
2004-07-10 Martin Quinson
 [autoconf mechanism]
  - get ride of a bunch of deprecated macros
  - actually run the test for two-compliment, not only compile it :-/
  - test whether the structures get packed (and bail out if yes. Damn.
    Alignment is a serious matter)
  - test whether the structures get compacted (but respecting the alignment
    constraints of each types)
  - test whether the array fields of structures can straddle alignment boundaries
 [base]
  - Damnit, double are bigger than float (typo in creation of 'double' datadesc)
    (took me 2 hours to find that bug, looking at the wrong place)
  - Add gras_datadesc_declare_{union,struct}_close(). They must be used
    before sending/receiving and are used to compute the offsets of fields
  - Given that padding size depend even on compiler options, keep track of
    alignment and aligned_size only for the current architecture. Not a big
    deal since we send structure fields one after the other (seems
    reasonable).    
  - Add the datastructure used for IEEE paper by the PBIO guys to the test
    program, let it work on linux/gcc/little32. portability todo.

2004-07-08 Martin Quinson
  - import and improve remote compilation support from FAST
  - make sure make check works on half a dozen of machines out there

2004-07-07 Martin Quinson
 Let's say it's version 0.0.3 ;)
  - Implement conversions (yuhu!)
  - Let it work on solaris (beside conversion, of course)
  - Stupid me, using rand() to generate the conversion datatests in not wise.

2004-07-06 Martin Quinson
  - Let make dist work, since I'm gonna need it to compile on remote hosts
  - Let Tests/datadesc_usage write the architecture on which the file was
    generated as first byte.
  - Add PowerPC (being also IRIX64), SPARC (also power4) and ALPHA
    architecture descriptions. 
  - Add datadesc_usage.{i386,ppc,sparc} files being the result of execution
    on those architectures.
  - Optimization: send/recv array of scalar in one shoot

2004-07-05 Martin Quinson
  - YEAH! GRAS/SG and GRAS/RL are both able to run the ping example !
  
  - Plug a whole bunch of memleaks
  - each process now have to call gras_{init,exit}. One day, their log
    settings will be separated
  - Continue the code factorisation between SG, RL and common in Transport.

2004-07-04 Martin Quinson
 [Transport]
  - Redistribution between SG and RL. 
    We wanna have to accept in SG, so move accepted related parts of RL in
    the common part. (more precisely, the dynar of all known sockets is no
    more a static in transport.c, but part of the process_data)
 [Core/module.c] 
 [gras_stub_generator]
  - Bug fix: Do call gras_process_init from gras_init (wasnt called in RL).

2004-07-03 Martin Quinson
  - Create a new log channel tbx containing dict, set, log, dynar (to shut
    them all up in one shot)
 [DataDesc]
  - Fix the ugly case of reference to dynamic array.
  - New (semi-public) function gras_datadesc_size to allow the messaging
    layer to malloc the needed space for the buffer.
 [Transport]
  - gras_socket_close now expect the socket to close (and not its address to
    put NULL in it after it). This is because the socket passed to handlers
    is one of their argument (=> not writable).
 [Messaging]
  - propagate the interface cleanup from last week in datadesc, ie remove a
    superfluous level of indirection. User pass adress of variable
    containing data (both when sending and receiving), and not of a variable
    being a pointer to the data. Let's say that I like it better ;)
      The price for that is constructs like "int msg=*(int*)payload" in
    handlers, but it's a fine price, IMHO.
 [examples/ping]
  - Let it work in RL (yuhu)

2004-06-21 Martin Quinson
 [Transport]
   - porting SG plugin and SG select to new standards (works almost).
   - plug memleaks and fix bugs around.
   
 [DataDesc] 
   - cleanup the prototype of data recv and force users to specify when they 
     want to handle references to objects. Test case working even for cycles.
   - plug memleaks. Valgrind is perfectly ok with this.

2004-06-12 Martin Quinson
 [Transport] 
   - cleanup the separation between plugin and main code in plugin creation 

2004-06-11 Martin Quinson
 [Transport]
   - Reput hook for raw sockets, needed for BW experiments
   - kill a few lines of dead code
 [Data description] Interface cleanup
   - gras_datadesc_by_name returns the searched type or NULL.
     That way, no variable is needed to use a type desc once, which makes
      the code clearer.
   - gras_datadesc_declare_[struct|union]_append_name is removed. The last
      two parameters were strings (field name, type name), leading to
      common errors.
 [Dicos] Interface cleanup
   - gras_dico_retrieve -> gras_dico_get ; gras_dico_insert -> gras_dico_set 
     This is consistant with the dynar API.

2004-04-21 Martin Quinson
 [Messaging]
   - Porting to new standards.
 [Data description]
   - interface cleanup. 
     There is no bag anymore, no need to take extra provision to mask the
       pointers behind "ID". 
     Better splitup of functions between files create/exchange/convert.
       This is still a bit artificial since convert and receive are so
       interleaved, but anyway.
 [Virtu(process)]
   - add a queued message list to procdata (the ones not matching criteria
     in msg_wait)
   - factorize some more code between SG and RL wrt procdata
 [Tests]
   - use gras_exit in example to track memleaks
   - get rid of gs_example now that GS is properly integrated into gras
   - update run_test to integrate the lastest tests (datadesc)
 [Logging]
   - rename WARNINGn macros to WARNn since it prooved error-prone
     
2004-04-19 Martin Quinson
 [Data description]
   - register init/exit functions within gras module mechanism   
   - send/receive function. 
   Convertion is not implemented, but short-cutted if not needed.
   struct/array elements are sent one by one (instead of block-wise), but
     nobody really cares (yet). Get a prototype before optimizing.
   - tests (using a file socket) for DD send/receive on:
     - base types: int, float
     - array: fixed size, string (ie ref to dynamic string)
     - structure: homogeneous, heterogeneous
     - chained list, graph with cycle
   Believe it or not, valgrind is not too unhappy with the results. The
    cycle happily segfaults, but the others are ok. And I'm sick of pointers
    for now.
 [Transport]
   [File plugin] 
     - Bugfix when using a filename explicitely (instead of '-')

2004-04-09 Martin Quinson
 [Transport plugins]
   - factorize more code between RL and SG in socket creation
   - Complete the implementation and tests of:
     o TCP
     o file (only in RL, and mainly for debugging)
     
     I lost 3 days to design a portable address resolver, and then decided
       that the prototype mainly have to run on my box.
     Addressing portability too early may be like optimizing too early :-/
 [Tests]
   - use gras_init in the Tests instead of the crappy parse_log_opt 
     (the latter function is removed)
 [Conditional execution]
   - New functions: gras_if_RL/gras_if_SG (basic support for this)
 [Code reorganisation]
  - Get rid of libgrasutils.a since it makes more trouble than it solves.
    Build examples against the RL library, since there is no way to disable
    its creation for now.

For information, the beginning of coding on GRAS was back in june
2003. I guess that every line has been rewritten at least twice since
then.