SMPI is now considered as stable and you can use it in production. You
may probably want to read the scientific publications that detail the
models used and their limits, but this should not be absolutely
-necessary.
-
-If you already fluently write and use MPI applications, SMPI should
-sound very familiar to you. Use smpicc instead of mpicc, and smpirun
-instead of mpirun. Some more information are given below on this page.
+necessary. If you already fluently write and use MPI applications,
+SMPI should sound very familiar to you. Use smpicc instead of mpicc,
+and smpirun instead of mpirun (see below for more details).
Of course, if you don't know what MPI is, the documentation of SMPI
will seem a bit terse to you. You should pick up a good MPI tutorial
-on the Internet (or a course in your university) and come back to SMPI
-once you know a bit more about MPI. Alternatively, you may want to
-turn to the other SimGrid interface such as the \ref MSG_API
-environment, or the \ref SD_API one.
+on the Internet (or a course in your favorite university) and come
+back to SMPI once you know a bit more about MPI. Alternatively, you
+may want to turn to the other SimGrid interfaces such as the
+\ref MSG_API environment, or the \ref SD_API one.
\section SMPI_what What can run within SMPI?
Given the size of the MPI standard, it may well be that we never
implement absolutely all existing primitives. One sided communications
and I/O primitives are not targeted for now. Our current state is
-still very decent: we pass most of the MPICH coverage tests.
+still very decent: we pass most of the MPICH coverage tests.
The full list of not yet implemented functions is documented in the
file <tt>include/smpi/smpi.h</tt> of the archive, between two lines
SimGrid code to help you implementing it, and we'd glad to integrate
it in the main project afterward if you contribute them back.
-\subsection SMPI_what_globals Issues with the globals
+\subsection SMPI_what_globals Global variables
Concerning the globals, the problem comes from the fact that usually,
MPI processes run as real UNIX processes while they are all folded
(note that this article does not deal with SMPI but with a concurrent
solution called AMPI that suffers of the same issue).
-Currently, we have no solution to offer you, because all proposed solutions will
-modify the performance of your application (in the computational
-sections). Sacrificing realism for usability is not very satisfying, so we did
-not implement them yet. You will thus have to modify your application if it uses
-global variables. We are working on another solution, leveraging distributed
-simulation to keep each MPI process within a separate system process, but this
-is far from being ready at the moment.
+A method using dynamic switching of the .data and .bss segments of an
+ELF executable has been introduced in SimGrid 3.11. By using the <tt>smpi/
+privatize_global_variableles</tt> option to yes, SMPI will duplicate
+the segments containing the global variables and when needed, will map
+the right one in memory. This needs ELF executables and mmap on the system
+(Linux and recent BSDs should be compatible). %As no copy is involved,
+performance should not be altered (but memory occupation will be higher).
+
+This solution actually works really good for a good number of MPI
+applications. Its main limitation is that if the application loads dynamic
+libraries, their global variables won't be privatized. This can be avoided
+by linking statically with these libraries (but NOT with libsimgrid, as we
+need SimGrid's own global varibles).
+
\section SMPI_compiling Compiling your code
<tt>examples/smpi/NAS/EP-sampling/ep.c</tt>
-*/
\ No newline at end of file
+\section SMPI_collective_algorithms Simulating collective operations
+
+MPI collective operations can be implemented very differently from one library
+to another. Actually, all existing libraries implement several algorithms
+for each collective operation, and by default select at runtime which one
+should be used for the current operation, depending on the sizes sent, the number
+ of nodes, the communicator, or the communication library being used. These
+decisions are based on empirical results and theoretical complexity estimation,
+but they can sometimes be suboptimal. Manual selection is possible in these cases,
+to allow the user to tune the library and use the better collective if the
+default one is not good enough.
+
+SMPI tries to apply the same logic, regrouping algorithms from OpenMPI, MPICH
+libraries, StarMPI (<a href="http://star-mpi.sourceforge.net/">STAR-MPI</a>), and MVAPICH2 libraries.
+This collection of more than 115 algorithms allows a simple and effective
+ comparison of their behavior and performance, making SMPI a tool of choice for the
+development of such algorithms.
+
+\subsection Tracing_internals Tracing of internal communications
+
+For each collective, default tracing only outputs global data.
+Internal communication operations are not traced to avoid outputting too much data
+to the trace. To debug and compare algorithm, this can be changed with the item
+\b tracing/smpi/internals , which has 0 for default value.
+Here are examples of two alltoall collective algorithms runs on 16 nodes,
+the first one with a ring algorithm, the second with a pairwise one :
+
+\htmlonly
+<a href="smpi_simgrid_alltoall_ring_16.png" border=0><img src="smpi_simgrid_alltoall_ring_16.png" width="30%" border=0 align="center"></a>
+<a href="smpi_simgrid_alltoall_pair_16.png" border=0><img src="smpi_simgrid_alltoall_pair_16.png" width="30%" border=0 align="center"></a>
+<br/>
+\endhtmlonly
+
+\subsection Selectors
+
+The default selection logic implemented by default in OpenMPI (version 1.7)
+and MPICH (version 3.0.4) has been replicated and can be used by setting the
+\b smpi/coll_selector item to either ompi or mpich. A selector based on the selection logic of MVAPICH2 (version 1.9) tuned on the Stampede cluster as also been implemented, as well as a preliminary version of an Intel MPI selector (version 4.1.3, also tuned for the Stampede cluster). Due the closed source nature of Intel MPI, some of the algorithms described in the documentation are not available, and are replaced by mvapich ones.
+
+Values for option \b smpi/coll_selector are :
+ - ompi
+ - mpich
+ - mvapich2
+ - impi
+ - default
+
+The code and details for each
+selector can be found in the <tt>src/smpi/colls/smpi_(openmpi/mpich/mvapich2/impi)_selector.c</tt> file.
+As this is still in development, we do not insure that all algorithms are correctly
+ replicated and that they will behave exactly as the real ones. If you notice a difference,
+please contact <a href="http://lists.gforge.inria.fr/mailman/listinfo/simgrid-devel">SimGrid developers mailing list</a>
+
+The default selector uses the legacy algorithms used in versions of SimGrid
+ previous to the 3.10. they should not be used to perform performance study and
+may be removed in the future, a different selector being used by default.
+
+\subsection algos Available algorithms
+
+For each one of the listed algorithms, several versions are available,
+ either coming from STAR-MPI, MPICH or OpenMPI implementations. Details can be
+ found in the code or in <a href="http://www.cs.arizona.edu/~dkl/research/papers/ics06.pdf">STAR-MPI</a> for STAR-MPI algorithms.
+
+Each collective can be selected using the corresponding configuration item. For example, to use the pairwise alltoall algorithm, one should add \b --cfg=smpi/alltoall:pair to the line. This will override the selector (for this algorithm only) if provided, allowing better flexibility.
+
+Warning: Some collective may require specific conditions to be executed correctly (for instance having a communicator with a power of two number of nodes only), which are currently not enforced by Simgrid. Some crashes can be expected while trying these algorithms with unusual sizes/parameters
+
+\subsubsection MPI_Alltoall
+
+Most of these are best described in <a href="http://www.cs.arizona.edu/~dkl/research/papers/ics06.pdf">STAR-MPI</a>
+
+ - default : naive one, by default
+ - ompi : use openmpi selector for the alltoall operations
+ - mpich : use mpich selector for the alltoall operations
+ - mvapich2 : use mvapich2 selector for the alltoall operations
+ - impi : use intel mpi selector for the alltoall operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - 2dmesh : organizes the nodes as a two dimensional mesh, and perform allgather
+along the dimensions
+ - 3dmesh : adds a third dimension to the previous algorithm
+ - rdb : recursive doubling : extends the mesh to a nth dimension, each one
+containing two nodes
+ - pair : pairwise exchange, only works for power of 2 procs, size-1 steps,
+each process sends and receives from the same process at each step
+ - pair_light_barrier : same, with small barriers between steps to avoid contention
+ - pair_mpi_barrier : same, with MPI_Barrier used
+ - pair_one_barrier : only one barrier at the beginning
+ - ring : size-1 steps, at each step a process send to process (n+i)%size, and receives from (n-i)%size
+ - ring_light_barrier : same, with small barriers between some phases to avoid contention
+ - ring_mpi_barrier : same, with MPI_Barrier used
+ - ring_one_barrier : only one barrier at the beginning
+ - basic_linear : posts all receives and all sends,
+starts the communications, and waits for all communication to finish
+ - mvapich2_scatter_dest : isend/irecv with scattered destinations, posting only a few messages at the same time
+
+\subsubsection MPI_Alltoallv
+
+ - default : naive one, by default
+ - ompi : use openmpi selector for the alltoallv operations
+ - mpich : use mpich selector for the alltoallv operations
+ - mvapich2 : use mvapich2 selector for the alltoallv operations
+ - impi : use intel mpi selector for the alltoallv operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - bruck : same as alltoall
+ - pair : same as alltoall
+ - pair_light_barrier : same as alltoall
+ - pair_mpi_barrier : same as alltoall
+ - pair_one_barrier : same as alltoall
+ - ring : same as alltoall
+ - ring_light_barrier : same as alltoall
+ - ring_mpi_barrier : same as alltoall
+ - ring_one_barrier : same as alltoall
+ - ompi_basic_linear : same as alltoall
+
+
+\subsubsection MPI_Gather
+
+ - default : naive one, by default
+ - ompi : use openmpi selector for the gather operations
+ - mpich : use mpich selector for the gather operations
+ - mvapich2 : use mvapich2 selector for the gather operations
+ - impi : use intel mpi selector for the gather operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+which will iterate over all implemented versions and output the best
+ - ompi_basic_linear : basic linear algorithm from openmpi, each process sends to the root
+ - ompi_binomial : binomial tree algorithm
+ - ompi_linear_sync : same as basic linear, but with a synchronization at the
+ beginning and message cut into two segments.
+ - mvapich2_two_level : SMP-aware version from MVAPICH. Gather first intra-node (defaults to mpich's gather), and then exchange with only one process/node. Use mvapich2 selector to change these to tuned algorithms for Stampede cluster.
+
+\subsubsection MPI_Barrier
+ - default : naive one, by default
+ - ompi : use openmpi selector for the barrier operations
+ - mpich : use mpich selector for the barrier operations
+ - mvapich2 : use mvapich2 selector for the barrier operations
+ - impi : use intel mpi selector for the barrier operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - ompi_basic_linear : all processes send to root
+ - ompi_two_procs : special case for two processes
+ - ompi_bruck : nsteps = sqrt(size), at each step, exchange data with rank-2^k and rank+2^k
+ - ompi_recursivedoubling : recursive doubling algorithm
+ - ompi_tree : recursive doubling type algorithm, with tree structure
+ - ompi_doublering : double ring algorithm
+ - mvapich2_pair : pairwise algorithm
+
+
+\subsubsection MPI_Scatter
+ - default : naive one, by default
+ - ompi : use openmpi selector for the scatter operations
+ - mpich : use mpich selector for the scatter operations
+ - mvapich2 : use mvapich2 selector for the scatter operations
+ - impi : use intel mpi selector for the scatter operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - ompi_basic_linear : basic linear scatter
+ - ompi_binomial : binomial tree scatter
+ - mvapich2_two_level_direct : SMP aware algorithm, with an intra-node stage (default set to mpich selector), and then a basic linear inter node stage. Use mvapich2 selector to change these to tuned algorithms for Stampede cluster.
+ - mvapich2_two_level_binomial : SMP aware algorithm, with an intra-node stage (default set to mpich selector), and then a binomial phase. Use mvapich2 selector to change these to tuned algorithms for Stampede cluster.
+
+
+
+\subsubsection MPI_Reduce
+ - default : naive one, by default
+ - ompi : use openmpi selector for the reduce operations
+ - mpich : use mpich selector for the reduce operations
+ - mvapich2 : use mvapich2 selector for the reduce operations
+ - impi : use intel mpi selector for the reduce operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - arrival_pattern_aware : root exchanges with the first process to arrive
+ - binomial : uses a binomial tree
+ - flat_tree : uses a flat tree
+ - NTSL : Non-topology-specific pipelined linear-bcast function
+ 0->1, 1->2 ,2->3, ....., ->last node : in a pipeline fashion, with segments
+ of 8192 bytes
+ - scatter_gather : scatter then gather
+ - ompi_chain : openmpi reduce algorithms are built on the same basis, but the
+ topology is generated differently for each flavor
+chain = chain with spacing of size/2, and segment size of 64KB
+ - ompi_pipeline : same with pipeline (chain with spacing of 1), segment size
+depends on the communicator size and the message size
+ - ompi_binary : same with binary tree, segment size of 32KB
+ - ompi_in_order_binary : same with binary tree, enforcing order on the
+operations
+ - ompi_binomial : same with binomial algo (redundant with default binomial
+one in most cases)
+ - ompi_basic_linear : basic algorithm, each process sends to root
+ - mvapich2_knomial : k-nomial algorithm. Default factor is 4 (mvapich2 selector adapts it through tuning)
+ - mvapich2_two_level : SMP-aware reduce, with default set to mpich both for intra and inter communicators. Use mvapich2 selector to change these to tuned algorithms for Stampede cluster.
+ - rab : <a href="https://fs.hlrs.de/projects/par/mpi//myreduce.html">Rabenseifner</a>'s reduce algorithm
+
+\subsubsection MPI_Allreduce
+ - default : naive one, by default
+ - ompi : use openmpi selector for the allreduce operations
+ - mpich : use mpich selector for the allreduce operations
+ - mvapich2 : use mvapich2 selector for the allreduce operations
+ - impi : use intel mpi selector for the allreduce operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - lr : logical ring reduce-scatter then logical ring allgather
+ - rab1 : variations of the <a href="https://fs.hlrs.de/projects/par/mpi//myreduce.html">Rabenseifner</a> algorithm : reduce_scatter then allgather
+ - rab2 : variations of the <a href="https://fs.hlrs.de/projects/par/mpi//myreduce.html">Rabenseifner</a> algorithm : alltoall then allgather
+ - rab_rsag : variation of the <a href="https://fs.hlrs.de/projects/par/mpi//myreduce.html">Rabenseifner</a> algorithm : recursive doubling
+reduce_scatter then recursive doubling allgather
+ - rdb : recursive doubling
+ - smp_binomial : binomial tree with smp : binomial intra
+SMP reduce, inter reduce, inter broadcast then intra broadcast
+ - smp_binomial_pipeline : same with segment size = 4096 bytes
+ - smp_rdb : intra : binomial allreduce, inter : Recursive
+doubling allreduce, intra : binomial broadcast
+ - smp_rsag : intra : binomial allreduce, inter : reduce-scatter,
+inter:allgather, intra : binomial broadcast
+ - smp_rsag_lr : intra : binomial allreduce, inter : logical ring
+reduce-scatter, logical ring inter:allgather, intra : binomial broadcast
+ - smp_rsag_rab : intra : binomial allreduce, inter : rab
+reduce-scatter, rab inter:allgather, intra : binomial broadcast
+ - redbcast : reduce then broadcast, using default or tuned algorithms if specified
+ - ompi_ring_segmented : ring algorithm used by OpenMPI
+ - mvapich2_rs : rdb for small messages, reduce-scatter then allgather else
+ - mvapich2_two_level : SMP-aware algorithm, with mpich as intra algoritm, and rdb as inter (Change this behavior by using mvapich2 selector to use tuned values)
+ - rab : default <a href="https://fs.hlrs.de/projects/par/mpi//myreduce.html">Rabenseifner</a> implementation
+
+\subsubsection MPI_Reduce_scatter
+ - default : naive one, by default
+ - ompi : use openmpi selector for the reduce_scatter operations
+ - mpich : use mpich selector for the reduce_scatter operations
+ - mvapich2 : use mvapich2 selector for the reduce_scatter operations
+ - impi : use intel mpi selector for the reduce_scatter operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - ompi_basic_recursivehalving : recursive halving version from OpenMPI
+ - ompi_ring : ring version from OpenMPI
+ - mpich_pair : pairwise exchange version from MPICH
+ - mpich_rdb : recursive doubling version from MPICH
+ - mpich_noncomm : only works for power of 2 procs, recursive doubling for noncommutative ops
+
+
+\subsubsection MPI_Allgather
+
+ - default : naive one, by default
+ - ompi : use openmpi selector for the allgather operations
+ - mpich : use mpich selector for the allgather operations
+ - mvapich2 : use mvapich2 selector for the allgather operations
+ - impi : use intel mpi selector for the allgather operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - 2dmesh : see alltoall
+ - 3dmesh : see alltoall
+ - bruck : Described by Bruck et.al. in <a href="http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=642949">
+Efficient algorithms for all-to-all communications in multiport message-passing systems</a>
+ - GB : Gather - Broadcast (uses tuned version if specified)
+ - loosely_lr : Logical Ring with grouping by core (hardcoded, default
+processes/node: 4)
+ - NTSLR : Non Topology Specific Logical Ring
+ - NTSLR_NB : Non Topology Specific Logical Ring, Non Blocking operations
+ - pair : see alltoall
+ - rdb : see alltoall
+ - rhv : only power of 2 number of processes
+ - ring : see alltoall
+ - SMP_NTS : gather to root of each SMP, then every root of each SMP node
+post INTER-SMP Sendrecv, then do INTRA-SMP Bcast for each receiving message,
+using logical ring algorithm (hardcoded, default processes/SMP: 8)
+ - smp_simple : gather to root of each SMP, then every root of each SMP node
+post INTER-SMP Sendrecv, then do INTRA-SMP Bcast for each receiving message,
+using simple algorithm (hardcoded, default processes/SMP: 8)
+ - spreading_simple : from node i, order of communications is i -> i + 1, i ->
+ i + 2, ..., i -> (i + p -1) % P
+ - ompi_neighborexchange : Neighbor Exchange algorithm for allgather.
+Described by Chen et.al. in <a href="http://ieeexplore.ieee.org/xpl/articleDetails.jsp?tp=&arnumber=1592302">Performance Evaluation of Allgather Algorithms on Terascale Linux Cluster with Fast Ethernet</a>
+ - mvapich2_smp : SMP aware algorithm, performing intra-node gather, inter-node allgather with one process/node, and bcast intra-node
+
+
+\subsubsection MPI_Allgatherv
+ - default : naive one, by default
+ - ompi : use openmpi selector for the allgatherv operations
+ - mpich : use mpich selector for the allgatherv operations
+ - mvapich2 : use mvapich2 selector for the allgatherv operations
+ - impi : use intel mpi selector for the allgatherv operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - GB : Gatherv - Broadcast (uses tuned version if specified, but only for
+Bcast, gatherv is not tuned)
+ - pair : see alltoall
+ - ring : see alltoall
+ - ompi_neighborexchange : see allgather
+ - ompi_bruck : see allgather
+ - mpich_rdb : recursive doubling algorithm from MPICH
+ - mpich_ring : ring algorithm from MPICh - performs differently from the
+one from STAR-MPI
+
+\subsubsection MPI_Bcast
+ - default : naive one, by default
+ - ompi : use openmpi selector for the bcast operations
+ - mpich : use mpich selector for the bcast operations
+ - mvapich2 : use mvapich2 selector for the bcast operations
+ - impi : use intel mpi selector for the bcast operations
+ - automatic (experimental) : use an automatic self-benchmarking algorithm
+ - arrival_pattern_aware : root exchanges with the first process to arrive
+ - arrival_pattern_aware_wait : same with slight variation
+ - binomial_tree : binomial tree exchange
+ - flattree : flat tree exchange
+ - flattree_pipeline : flat tree exchange, message split into 8192 bytes pieces
+ - NTSB : Non-topology-specific pipelined binary tree with 8192 bytes pieces
+ - NTSL : Non-topology-specific pipelined linear with 8192 bytes pieces
+ - NTSL_Isend : Non-topology-specific pipelined linear with 8192 bytes pieces, asynchronous communications
+ - scatter_LR_allgather : scatter followed by logical ring allgather
+ - scatter_rdb_allgather : scatter followed by recursive doubling allgather
+ - arrival_scatter : arrival pattern aware scatter-allgather
+ - SMP_binary : binary tree algorithm with 8 cores/SMP
+ - SMP_binomial : binomial tree algorithm with 8 cores/SMP
+ - SMP_linear : linear algorithm with 8 cores/SMP
+ - ompi_split_bintree : binary tree algorithm from OpenMPI, with message split in 8192 bytes pieces
+ - ompi_pipeline : pipeline algorithm from OpenMPI, with message split in 128KB pieces
+ - mvapich2_inter_node : Inter node default mvapich worker
+ - mvapich2_intra_node : Intra node default mvapich worker
+ - mvapich2_knomial_intra_node : k-nomial intra node default mvapich worker. default factor is 4.
+
+\subsection auto Automatic evaluation
+
+(Warning : This is experimental and may be removed or crash easily)
+
+An automatic version is available for each collective (or even as a selector). This specific
+version will loop over all other implemented algorithm for this particular collective, and apply
+them while benchmarking the time taken for each process. It will then output the quickest for
+each process, and the global quickest. This is still unstable, and a few algorithms which need
+specific number of nodes may crash.
+
+
+\subsection add Add an algorithm
+
+To add a new algorithm, one should check in the src/smpi/colls folder how other algorithms
+are coded. Using plain MPI code inside Simgrid can't be done, so algorithms have to be
+changed to use smpi version of the calls instead (MPI_Send will become smpi_mpi_send). Some functions may have different signatures than their MPI counterpart, please check the other algorithms or contact us using <a href="http://lists.gforge.inria.fr/mailman/listinfo/simgrid-devel">SimGrid developers mailing list</a>.
+
+Example: adding a "pair" version of the Alltoall collective.
+
+ - Implement it in a file called alltoall-pair.c in the src/smpi/colls folder. This file should include colls_private.h.
+
+ - The name of the new algorithm function should be smpi_coll_tuned_alltoall_pair, with the same signature as MPI_Alltoall.
+
+ - Once the adaptation to SMPI code is done, add a reference to the file ("src/smpi/colls/alltoall-pair.c") in the SMPI_SRC part of the DefinePackages.cmake file inside buildtools/cmake, to allow the file to be built and distributed.
+
+ - To register the new version of the algorithm, simply add a line to the corresponding macro in src/smpi/colls/cools.h ( add a "COLL_APPLY(action, COLL_ALLTOALL_SIG, pair)" to the COLL_ALLTOALLS macro ). The algorithm should now be compiled and be selected when using --cfg=smpi/alltoall:pair at runtime.
+
+ - To add a test for the algorithm inside Simgrid's test suite, juste add the new algorithm name in the ALLTOALL_COLL list found inside buildtools/cmake/AddTests.cmake . When running ctest, a test for the new algorithm should be generated and executed. If it does not pass, please check your code or contact us.
+
+ - Feel free to push this new algorithm to the SMPI repository using Git.
+
+
+
+
+*/
