Programming environment for the simulation of MPI applications.
More...
Programming environment for the simulation of MPI applications.
This programming environment enables the study of MPI application by emulating them on top of the SimGrid simulator.
This is particularly interesting to study existing MPI applications within the comfort of the simulator. The motivation for this work is detailed in the reference article (available at http://hal.inria.fr/inria-00527150).
Our goal is to enable the study of unmodified MPI applications, although we are not quite there yet (see What can run within SMPI?). In addition, you can modify your code to speed up your studies or otherwise increase their scalability (see Adapting your MPI code to the use of SMPI).
Who should use SMPI (and who shouldn't)
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 (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 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 MSG environment, or the SimDag one.
What can run within SMPI?
You can run unmodified MPI applications (both C and Fortran) within SMPI, provided that (1) you only use MPI calls that we implemented in MPI and (2) you don't use any globals in your application.
MPI coverage of SMPI
Our coverage of the interface is very decent, but still incomplete; 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.
The full list of not yet implemented functions is documented in the file include/smpi/smpi.h
of the archive, between two lines containing the FIXME
marker. If you really need a missing feature, please get in touch with us: we can guide you though the SimGrid code to help you implementing it, and we'd glad to integrate it in the main project afterward if you contribute them back.
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 into threads of a unique system process in SMPI. Global variables are usually private to each MPI process while they become shared between the processes in SMPI. This point is rather problematic, and currently forces to modify your application to privatize the global variables.
We tried several techniques to work this around. We used to have a script that privatized automatically the globals through static analysis of the source code, but it was not robust enough to be used in production. This issue, as well as several potential solutions, is discussed in this article: "Automatic Handling of Global Variables for
Multi-threaded MPI Programs", available at http://charm.cs.illinois.edu/newPapers/11-23/paper.pdf (note that this article does not deal with SMPI but with a concurrent solution called AMPI that suffers of the same issue).
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 smpi/ privatize_global_variableles
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).
Compiling your code
This is very simply done with the smpicc
script. If you already compiled any MPI code before, you already know how to use it. If not, you should try to get your MPI code running on top of MPI before giving SMPI a spin. Actually, that's very simple even if it's the first time you use MPI code: just use smpicc as a compiler (in replacement of gcc or your usual compiler), and you're set.
Executing your code on top of the simulator
This is done though the smpirun
script as follows. my_hostfile.txt
is a classical MPI hostfile (that is, this file lists the machines on which the processes must be dispatched, one per line) my_platform.xml
is a classical SimGrid platform file. Of course, the hosts of the hostfile must exist in the provided platform. ./program
is the MPI program that you want to simulate (must be compiled by smpicc
) while -arg
is a command-line parameter passed to this program.
smpirun -hostfile my_hostfile.txt -platform my_platform.xml ./program -arg
smpirun accepts other parameters, such as -np
if you don't want to use all the hosts defined in the hostfile, -map
to display on which host each rank gets mapped of -trace
to activate the tracing during the simulation. You can get the full list by running
smpirun -help
Adapting your MPI code to the use of SMPI
As detailed in the reference article (available at http://hal.inria.fr/inria-00527150), you may want to adapt your code to improve the simulation performance. But these tricks may seriously hinder the result quality (or even prevent the app to run) if used wrongly. We assume that if you want to simulate an HPC application, you know what you are doing. Don't prove us wrong!
Reducing your memory footprint
If you get short on memory (the whole app is executed on a single node when simulated), you should have a look at the SMPI_SHARED_MALLOC and SMPI_SHARED_FREE macros. It allows to share memory areas between processes: The purpose of these macro is that the same line malloc on each process will point to the exact same memory area. So if you have a malloc of 2M and you have 16 processes, this macro will change your memory consumption from 2M*16 to 2M only. Only one block for all processes.
If your program is ok with a block containing garbage value because all processes write and read to the same place without any kind of coordination, then this macro can dramatically shrink your memory consumption. For example, that will be very beneficial to a matrix multiplication code, as all blocks will be stored on the same area. Of course, the resulting computations will useless, but you can still study the application behavior this way.
Naturally, this won't work if your code is data-dependent. For example, a Jacobi iterative computation depends on the result computed by the code to detect convergence conditions, so turning them into garbage by sharing the same memory area between processes does not seem very wise. You cannot use the SMPI_SHARED_MALLOC macro in this case, sorry.
This feature is demoed by the example file examples/smpi/NAS/DT-folding/dt.c
Toward faster simulations
If your application is too slow, try using SMPI_SAMPLE_LOCAL, SMPI_SAMPLE_GLOBAL and friends to indicate which computation loops can be sampled. Some of the loop iterations will be executed to measure their duration, and this duration will be used for the subsequent iterations. These samples are done per processor with SMPI_SAMPLE_LOCAL, and shared between all processors with SMPI_SAMPLE_GLOBAL. Of course, none of this will work if the execution time of your loop iteration are not stable.
This feature is demoed by the example file examples/smpi/NAS/EP-sampling/ep.c
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 (STAR-MPI), 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.
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 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 :
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 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 smpi/coll_selector are :
- ompi
- mpich
- mvapich2
- impi
- default
The code and details for each selector can be found in the src/smpi/colls/smpi_(openmpi/mpich/mvapich2/impi)_selector.c
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 SimGrid developers mailing list
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.
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 STAR-MPI 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 –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
MPI_Alltoall
Most of these are best described in STAR-MPI
- 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
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
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.
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
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.
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 : Rabenseifner's reduce algorithm
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 Rabenseifner algorithm : reduce_scatter then allgather
- rab2 : variations of the Rabenseifner algorithm : alltoall then allgather
- rab_rsag : variation of the Rabenseifner 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 Rabenseifner implementation
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
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 Efficient algorithms for all-to-all communications in multiport message-passing systems
- 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 Performance Evaluation of Allgather Algorithms on Terascale Linux Cluster with Fast Ethernet
- mvapich2_smp : SMP aware algorithm, performing intra-node gather, inter-node allgather with one process/node, and bcast intra-node
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
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.
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.
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 SimGrid developers mailing list.
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/Tests.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.