SMPI:
- SMPI is now tested with ~45 proxy apps from various sources, with none or
- only minor patching needed : check github.com/simgrid/SMPI-proxy-apps
+ only minor patching needed: check github.com/simgrid/SMPI-proxy-apps
- Replay: The replay file has been re-written in C++.
- Replay: Tags used for messages sent via MPI_Send / MPI_Recv are now
supported. They are stored in the trace and used when replayed.
- Basic support of MPI_Cancel. Robustness not guaranteed.
- Support of MPI_Win_allocate_shared, MPI_Win_shared_query, MPI_Comm_split_type
(only for MPI_COMM_TYPE_SHARED).
- - New option : smpi/privatize-libs, to add external shared libs to be privatized
+ - New option: smpi/privatize-libs, to add external shared libs to be privatized
by SMPI. They will be copied locally and loaded separately by each process.
Example --cfg=smpi/privatize-libs:"libgfortran.so.3;libscalapack.so".
- - Tracing : add tracing for MPI_Start, Startall, Testall, Testany
+ - Tracing: add tracing for MPI_Start, Startall, Testall, Testany
- Interception of getopt, getopt_long and getopt_long_only calls to avoid issues
with internal index optind with multiple processes. Only works if MPI_Init has
already been called.
- - Fortran : SMPI builds a mpi.mod file which should allow use of "use mpi"
+ - Fortran: SMPI builds a mpi.mod file which should allow use of "use mpi"
syntax without preprocessing tricks.
XBT:
- #150: Inconsistent event names in SMPI replay
- #258: daemonized actors hang after all non-daemonized actors have completed
- #267: Linker error on unit_tmgr
- - #269: SMPI : tracing of MPI_Wait/all/any broken
- - SMPI : Fix various crashes with combined use of MPI_PROC_NULL and MPI_IGNORE_STATUS
+ - #269: SMPI: tracing of MPI_Wait/all/any broken
+ - SMPI: Fix various crashes with combined use of MPI_PROC_NULL and MPI_IGNORE_STATUS
----------------------------------------------------------------------------
Virtual Machines
- Live migration is getting moved to a plugin. Dirty page tracking is
the first part of this plugin. This imply that VM migration is now
- only possible if one this function is called :
+ only possible if one this function is called:
- C/MSG: MSG_vm_live_migration_plugin_init()
- C/C++: sg_vm_live_migration_plugin_init()
- Java: Msg.liveMigrationInit()
SMPI
- New algorithm to privatize globals: dlopen, with dynamic loading tricks
- New option: smpi/keep-temps to not cleanup temp files
- - New option : smpi/shared-malloc-blocksize . Relevant only when global shared
+ - New option: smpi/shared-malloc-blocksize . Relevant only when global shared
mallocs mode is used, allows to change the size of the fake file used
(default 1MB), to potentially limit the number of mappings for large runs.
- Support for sparse privatized malloc with SMPI_PARTIAL_SHARED_MALLOC()
- Fortran ifort and flang compilers support
- - New RMA calls supported (experimental) :
+ - New RMA calls supported (experimental):
- MPI_Win_allocate, MPI_Win_create_dynamic, MPI_Win_attach
- MPI_Win_detach, MPI_Win_set_info, MPI_Win_get_info
- MPI_Win_lock_all, MPI_Win_unlock_all, MPI_Win_flush
* smpirun script should be (much) faster for large deployments.
- * SMPI tracing : fixed issue with poor matching of send/receives.
+ * SMPI tracing: fixed issue with poor matching of send/receives.
- * Replay : Fix broken waitall
+ * Replay: Fix broken waitall
New functions and features
* MSG_parallel_task_execute_with_timeout, to timeout computations.
SMPI:
* New functions
- - Onesided early support for : MPI_Win_(create, free, fence, get_name, set_name, get_group), MPI_Get, MPI_Put, MPI_Accumulate, MPI_Alloc_mem, MPI_Free_mem.
+ - Onesided early support for: MPI_Win_(create, free, fence, get_name, set_name, get_group), MPI_Get, MPI_Put, MPI_Accumulate, MPI_Alloc_mem, MPI_Free_mem.
- MPI_Keyval*, MPI_Attr* functions, as well as MPI_Comm_attr*, MPI_Type_attr* variants (C only, no Fortran support yet)
- MPI_Type_set_name, MPI_Type_get_name
- MPI_*_c2f and MPI_*_f2c functions
- Activate a lot of new tests from the mpich 3 testsuite
* Features
- Constant times can be injected inside MPI_Wtime and MPI_Test through options smpi/wtime and smpi/test
- - InfiniBand network model added : Based on the works of Jerome Vienne
+ - InfiniBand network model added: Based on the works of Jerome Vienne
http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf
- When smpi/display_timing is set, also display global simulation time and application times
- Have smpirun, smpicc and friends display the simgrid git hash version on --git-version
* Collective communications
- SMP-aware algorithms are now dynamically handled. An internal communicator is created for each node, and an external one to handle communications between "leaders" of each node
- - MVAPICH2 (1.9) collective algorithms selector : normal and SMP algorithms are handled, and selection logic is based on the one used on TACC's Stampede cluster (https://www.tacc.utexas.edu/stampede/).
+ - MVAPICH2 (1.9) collective algorithms selector: normal and SMP algorithms are handled, and selection logic is based on the one used on TACC's Stampede cluster (https://www.tacc.utexas.edu/stampede/).
- Support for Rabenseifner Reduce/Allreduce algorithms (https://fs.hlrs.de/projects/par/mpi//myreduce.html)
* Replay
- Replay now uses algorithms from wanted collective selector
- Memory occupation of replay should now be contained (temporary buffers allocated in collective algorithms should be shared between processes)
- Replay can now replay several traces at the same time (check examples/smpi/replay_multiple example), to simulate interactions between several applications on a given platform. User can specify the start time of each instance. This should also allow replay + actual applications to run.
* Bug fixes
- - [#17799] : have mpi_group_range_incl and mpi_group_range_excl better test some corner cases
+ - [#17799]: have mpi_group_range_incl and mpi_group_range_excl better test some corner cases
- Correctly use loopback on fat-tree clusters
- Asynchronous small messages shouldn't trigger deadlocks anymore
* Energy/DVFS cleanup and improvement
* New functions
- Add a xbt_heap_update function, to avoid costly xbt_heap_remove+xbt_heap_insert use
- Add a xbt wrapper for simcall_mutex_trylock (asked in [#17878])
- - Add two new log appenders : rollfile and splitfile. Patch by Fabien Chaix.
+ - Add two new log appenders: rollfile and splitfile. Patch by Fabien Chaix.
- xbt_dirname and xbt_basename for non-POSIX systems
MC
* The model checker now runs as a separate process.
one node.
* Collective communication algorithms should not crash if used with
improper number of nodes and report the error.
- * SMPI now partially supports MPI_Topologies : MPI_Cart_create, MPI_Cart_shift,
+ * SMPI now partially supports MPI_Topologies: MPI_Cart_create, MPI_Cart_shift,
MPI_Cart_rank, MPI_Cart_get, MPI_Cart_coords, MPI_Cartdim_get,
MPI_Dims_create, MPI_Cart_sub are supported.
* New interface to use SMPI programmatically (still depends on MSG for
- some parts, see examples/smpi/smpi_msg_masterslave) :
+ some parts, see examples/smpi/smpi_msg_masterslave):
- SMPI_app_instance_register(const char *name, xbt_main_func_t code,
int num_processes)
- SMPI_init()
- SMPI_finalize();
* Global variables privatization in MPI executables is now performed at runtime
with the option smpi/privatize_global_variables (default:no).
- Limitations : Linux/BSD only, with mmap enabled. Global variables inside
+ Limitations: Linux/BSD only, with mmap enabled. Global variables inside
dynamic libraries loaded by the application are not privatized (static
linking with these libraries is advised in this case)
- allows to select one in particular with --cfg=smpi/coll_name:algorithm
- allows to use the decision logic of OpenMPI(1.7) or MPICH(3.0.4) by setting
--cfg=smpi/coll_selector:(mpich/ompi)
- * Support for new functions : MPI_Issend, MPI_Ssend, Commutative operations in
+ * Support for new functions: MPI_Issend, MPI_Ssend, Commutative operations in
Reduce
* Add a --cfg:tracing/smpi/internals option, to trace internal communications
happening inside a collective SMPI call.
by a SD_TASK_COMM_E2E typed task. This rate depends on both the nominal
bandwidth on the route onto which the task is scheduled and the amount of
data to transfer.
- To divide the nominal bandwidth by 2, the rate then has to be :
+ To divide the nominal bandwidth by 2, the rate then has to be:
rate = bandwidth/(2*amount)
* Compute tasks that have failed can now be rescheduled and executed again
(from their beginning)
action_free ~> action_unref
action_change_state ~> action_state_set
action_get_state ~> action_state_get
- - Change model methods into functions :
+ - Change model methods into functions:
(model)->common_public->action_use ~> surf_action_ref
* Implement a generic resource; use it as ancestor to specific ones
* 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
+ 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
* 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.
+ 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
* 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 :
+ 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
\subsubsection pf_router <router/>
As said before, <b>router</b> is used only to give some information
-for routing algorithms. So, it does not have any attributes except :
+for routing algorithms. So, it does not have any attributes except:
#### Attributes ####
\verbinclude example_filelist_routing_dijkstra
-Dijkstra example :
+Dijkstra example:
\verbatim
<zone id="zone_2" routing="Dijkstra">
<host id="zone_2_host1" speed="1000000000"/>
\anchor pf_routing_model_full
### Full ###
-Full example :
+Full example:
\verbatim
<zone id="zone0" routing="Full">
<host id="host1" speed="1000000000"/>
<b>bypasszoneroute</b> is the tag you're looking for. It allows to
bypass routes defined between already defined between zone (if you want
to bypass route for a specific host, you should just use byPassRoute).
-The principle is the same as zoneroute : <b>bypasszoneroute</b> contains
+The principle is the same as zoneroute: <b>bypasszoneroute</b> contains
list of links that are in the path between src and dst.
#### Attributes ####
As said before, once you choose
a model, it (most likely; the constant network model, for example, doesn't) calculates routes for you. But maybe you want to
define some of your routes, which will be specific. You may also want
-to bypass some routes defined in lower level zone at an upper stage :
+to bypass some routes defined in lower level zone at an upper stage:
<b>bypassRoute</b> is the tag you're looking for. It allows to bypass
routes defined between <b>host/router</b>. The principle is the same
-as route : <b>bypassRoute</b> contains list of links references of
+as route: <b>bypassRoute</b> contains list of links references of
links that are in the path between src and dst.
#### Attributes ####
defined inside zone_Big. If you choose some shortest-path model,
this route will be computed automatically.
-As said before, there are mainly 2 tags for routing :
+As said before, there are mainly 2 tags for routing:
\li <b>zoneroute</b>: to define routes between two <b>zone</b>
\li <b>route</b>: to define routes between two <b>host/router</b>
routing (as we don't want to bother with defining all routes). As
we're using some shortest path algorithms to route into zone_2, we'll
then have to define some <b>route</b> to gives some topological
-information to SimGrid. Here is a file doing it all :
+information to SimGrid. Here is a file doing it all:
\verbatim
<zone id="zone_Big" routing="Dijkstra">
\subsection pf_exit_zone Exit Zone: why and how
Users that have looked at some of our platforms may have notice a
-non-intuitive schema ... Something like that :
+non-intuitive schema ... Something like that:
\verbatim
Choosing wisely the routing model to use can significantly fasten your
simulation/save your time when writing the platform/save tremendous
disk space. Here is the list of available model and their
-characteristics (lookup : time to resolve a route):
+characteristics (lookup: time to resolve a route):
\li <b>Full</b>: Full routing data (fast, large memory requirements,
fully expressive)
