1 /*! \page options Configure SimGrid
3 A number of options can be given at runtime to change the default
4 SimGrid behavior. For a complete list of all configuration options
5 accepted by the SimGrid version used in your simulator, simply pass
6 the --help configuration flag to your program. If some of the options
7 are not documented on this page, this is a bug that you should please
8 report so that we can fix it. Note that some of the options presented
9 here may not be available in your simulators, depending on the
10 @ref install_src_config "compile-time options" that you used.
14 \section options_using Passing configuration options to the simulators
16 There is several way to pass configuration options to the simulators.
17 The most common way is to use the \c --cfg command line argument. For
18 example, to set the item \c Item to the value \c Value, simply
19 type the following: \verbatim
20 my_simulator --cfg=Item:Value (other arguments)
23 Several \c `--cfg` command line arguments can naturally be used. If you
24 need to include spaces in the argument, don't forget to quote the
25 argument. You can even escape the included quotes (write \' for ' if
26 you have your argument between ').
28 Another solution is to use the \c \<config\> tag in the platform file. The
29 only restriction is that this tag must occure before the first
30 platform element (be it \c \<AS\>, \c \<cluster\>, \c \<peer\> or whatever).
31 The \c \<config\> tag takes an \c id attribute, but it is currently
32 ignored so you don't really need to pass it. The important par is that
33 within that tag, you can pass one or several \c \<prop\> tags to specify
34 the configuration to use. For example, setting \c Item to \c Value
35 can be done by adding the following to the beginning of your platform
39 <prop id="Item" value="Value"/>
43 A last solution is to pass your configuration directly using the C
44 interface. If you happen to use the MSG interface, this is very easy
45 with the MSG_config() function. If you do not use MSG, that's a bit
46 more complex, as you have to mess with the internal configuration set
47 directly as follows. Check the \ref XBT_config "relevant page" for
48 details on all the functions you can use in this context, \c
49 _sg_cfg_set being the only configuration set currently used in
53 #include <xbt/config.h>
55 int main(int argc, char *argv[]) {
58 /* Prefer MSG_config() if you use MSG!! */
59 xbt_cfg_set_parse("Item:Value");
65 \section options_index Index of all existing configuration options
68 The full list can be retrieved by passing "--help" and
69 "--help-cfg" to an executable that uses SimGrid.
71 - \c clean-atexit: \ref options_generic_clean_atexit
73 - \c contexts/factory: \ref options_virt_factory
74 - \c contexts/guard-size: \ref options_virt_guard_size
75 - \c contexts/nthreads: \ref options_virt_parallel
76 - \c contexts/parallel_threshold: \ref options_virt_parallel
77 - \c contexts/stack-size: \ref options_virt_stacksize
78 - \c contexts/synchro: \ref options_virt_parallel
80 - \c cpu/maxmin-selective-update: \ref options_model_optim
81 - \c cpu/model: \ref options_model_select
82 - \c cpu/optim: \ref options_model_optim
84 - \c exception/cutpath: \ref options_exception_cutpath
86 - \c host/model: \ref options_model_select
88 - \c maxmin/precision: \ref options_model_precision
89 - \c maxmin/concurrency-limit: \ref options_concurrency_limit
91 - \c msg/debug-multiple-use: \ref options_msg_debug_multiple_use
93 - \c model-check: \ref options_modelchecking
94 - \c model-check/checkpoint: \ref options_modelchecking_steps
95 - \c model-check/communications-determinism: \ref options_modelchecking_comm_determinism
96 - \c model-check/dot-output: \ref options_modelchecking_dot_output
97 - \c model-check/hash: \ref options_modelchecking_hash
98 - \c model-check/property: \ref options_modelchecking_liveness
99 - \c model-check/max-depth: \ref options_modelchecking_max_depth
100 - \c model-check/record: \ref options_modelchecking_recordreplay
101 - \c model-check/reduction: \ref options_modelchecking_reduction
102 - \c model-check/replay: \ref options_modelchecking_recordreplay
103 - \c model-check/send-determinism: \ref options_modelchecking_comm_determinism
104 - \c model-check/sparse-checkpoint: \ref options_modelchecking_sparse_checkpoint
105 - \c model-check/termination: \ref options_modelchecking_termination
106 - \c model-check/timeout: \ref options_modelchecking_timeout
107 - \c model-check/visited: \ref options_modelchecking_visited
109 - \c network/bandwidth-factor: \ref options_model_network_coefs
110 - \c network/crosstraffic: \ref options_model_network_crosstraffic
111 - \c network/latency-factor: \ref options_model_network_coefs
112 - \c network/maxmin-selective-update: \ref options_model_optim
113 - \c network/model: \ref options_model_select
114 - \c network/optim: \ref options_model_optim
115 - \c network/sender_gap: \ref options_model_network_sendergap
116 - \c network/TCP-gamma: \ref options_model_network_gamma
117 - \c network/weight-S: \ref options_model_network_coefs
119 - \c ns3/TcpModel: \ref options_pls
120 - \c path: \ref options_generic_path
121 - \c plugin: \ref options_generic_plugin
123 - \c storage/max_file_descriptors: \ref option_model_storage_maxfd
125 - \c surf/precision: \ref options_model_precision
127 - \c <b>For collective operations of SMPI, please refer to Section \ref options_index_smpi_coll</b>
128 - \c smpi/async-small-thresh: \ref options_model_network_asyncsend
129 - \c smpi/bw-factor: \ref options_model_smpi_bw_factor
130 - \c smpi/coll-selector: \ref options_model_smpi_collectives
131 - \c smpi/comp-adjustment-file: \ref options_model_smpi_adj_file
132 - \c smpi/cpu-threshold: \ref options_smpi_bench
133 - \c smpi/display-timing: \ref options_smpi_timing
134 - \c smpi/grow-injected-times: \ref options_model_smpi_test
135 - \c smpi/host-speed: \ref options_smpi_bench
136 - \c smpi/IB-penalty-factors: \ref options_model_network_coefs
137 - \c smpi/iprobe: \ref options_model_smpi_iprobe
138 - \c smpi/iprobe-cpu-usage: \ref options_model_smpi_iprobe_cpu_usage
139 - \c smpi/init: \ref options_model_smpi_init
140 - \c smpi/keep-temps: \ref options_smpi_temps
141 - \c smpi/lat-factor: \ref options_model_smpi_lat_factor
142 - \c smpi/ois: \ref options_model_smpi_ois
143 - \c smpi/or: \ref options_model_smpi_or
144 - \c smpi/os: \ref options_model_smpi_os
145 - \c smpi/papi-events: \ref options_smpi_papi_events
146 - \c smpi/privatization: \ref options_smpi_privatization
147 - \c smpi/send-is-detached-thresh: \ref options_model_smpi_detached
148 - \c smpi/shared-malloc: \ref options_model_smpi_shared_malloc
149 - \c smpi/shared-malloc-hugepage: \ref options_model_smpi_shared_malloc
150 - \c smpi/simulate-computation: \ref options_smpi_bench
151 - \c smpi/test: \ref options_model_smpi_test
152 - \c smpi/wtime: \ref options_model_smpi_wtime
154 - \c <b>Tracing configuration options can be found in Section \ref tracing_tracing_options</b>.
156 - \c storage/model: \ref options_storage_model
157 - \c verbose-exit: \ref options_generic_exit
159 - \c vm/model: \ref options_vm_model
161 \subsection options_index_smpi_coll Index of SMPI collective algorithms options
163 TODO: All available collective algorithms will be made available via the ``smpirun --help-coll`` command.
165 \section options_model Configuring the platform models
167 \anchor options_storage_model
168 \anchor options_vm_model
169 \subsection options_model_select Selecting the platform models
171 SimGrid comes with several network, CPU and storage models built in, and you
172 can change the used model at runtime by changing the passed
173 configuration. The three main configuration items are given below.
174 For each of these items, passing the special \c help value gives
175 you a short description of all possible values. Also, \c --help-models
176 should provide information about all models for all existing resources.
177 - \b network/model: specify the used network model
178 - \b cpu/model: specify the used CPU model
179 - \b host/model: specify the used host model
180 - \b storage/model: specify the used storage model (there is currently only one such model - this option is hence only useful for future releases)
181 - \b vm/model: specify the model for virtual machines (there is currently only one such model - this option is hence only useful for future releases)
183 As of writing, the following network models are accepted. Over
184 the time new models can be added, and some experimental models can be
185 removed; check the values on your simulators for an uptodate
186 information. Note that the CM02 model is described in the research report
187 <a href="ftp://ftp.ens-lyon.fr/pub/LIP/Rapports/RR/RR2002/RR2002-40.ps.gz">A
188 Network Model for Simulation of Grid Application</a> while LV08 is
190 <a href="http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf">Accuracy Study and Improvement of Network Simulation in the SimGrid Framework</a>.
192 - \b LV08 (default one): Realistic network analytic model
193 (slow-start modeled by multiplying latency by 10.4, bandwidth by
194 .92; bottleneck sharing uses a payload of S=8775 for evaluating RTT)
195 - \anchor options_model_select_network_constant \b Constant: Simplistic network model where all communication
196 take a constant time (one second). This model provides the lowest
197 realism, but is (marginally) faster.
198 - \b SMPI: Realistic network model specifically tailored for HPC
199 settings (accurate modeling of slow start with correction factors on
200 three intervals: < 1KiB, < 64 KiB, >= 64 KiB). See also \ref
201 options_model_network_coefs "this section" for more info.
202 - \b IB: Realistic network model specifically tailored for HPC
203 settings with InfiniBand networks (accurate modeling contention
204 behavior, based on the model explained in
205 http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf).
206 See also \ref options_model_network_coefs "this section" for more info.
207 - \b CM02: Legacy network analytic model (Very similar to LV08, but
208 without corrective factors. The timings of small messages are thus
210 - \b Reno: Model from Steven H. Low using lagrange_solve instead of
211 lmm_solve (experts only; check the code for more info).
212 - \b Reno2: Model from Steven H. Low using lagrange_solve instead of
213 lmm_solve (experts only; check the code for more info).
214 - \b Vegas: Model from Steven H. Low using lagrange_solve instead of
215 lmm_solve (experts only; check the code for more info).
217 If you compiled SimGrid accordingly, you can use packet-level network
218 simulators as network models (see \ref pls_ns3). In that case, you have
219 two extra models, described below, and some
220 \ref options_pls "specific additional configuration flags".
221 - \b NS3: Network pseudo-model using the NS3 tcp model
223 Concerning the CPU, we have only one model for now:
224 - \b Cas01: Simplistic CPU model (time=size/power)
226 The host concept is the aggregation of a CPU with a network
227 card. Three models exists, but actually, only 2 of them are
228 interesting. The "compound" one is simply due to the way our internal
229 code is organized, and can easily be ignored. So at the end, you have
230 two host models: The default one allows to aggregate an
231 existing CPU model with an existing network model, but does not allow
232 parallel tasks because these beasts need some collaboration between
233 the network and CPU model. That is why, ptask_07 is used by default
235 - \b default: Default host model. Currently, CPU:Cas01 and
236 network:LV08 (with cross traffic enabled)
237 - \b compound: Host model that is automatically chosen if
238 you change the network and CPU models
239 - \b ptask_L07: Host model somehow similar to Cas01+CM02 but
240 allowing "parallel tasks", that are intended to model the moldable
241 tasks of the grid scheduling literature.
243 \subsection options_generic_plugin Plugins
245 SimGrid supports the use of plugins; currently, no known plugins
246 can be activated but there are use-cases where you may want to write
247 your own plugin (for instance, for logging).
249 Plugins can for instance define own classes that inherit from
250 existing classes (for instance, a class "CpuEnergy" inherits from
251 "Cpu" to assess energy consumption).
253 The plugin connects to the code by registering callbacks using
254 ``signal.connect(callback)`` (see file ``src/surf/plugins/energy.cpp`` for
262 This option is case-sensitive: Energy and energy are not the same!
264 \subsection options_model_optim Optimization level of the platform models
266 The network and CPU models that are based on lmm_solve (that
267 is, all our analytical models) accept specific optimization
269 - items \b network/optim and \b cpu/optim (both default to 'Lazy'):
270 - \b Lazy: Lazy action management (partial invalidation in lmm +
271 heap in action remaining).
272 - \b TI: Trace integration. Highly optimized mode when using
273 availability traces (only available for the Cas01 CPU model for
275 - \b Full: Full update of remaining and variables. Slow but may be
276 useful when debugging.
277 - items \b network/maxmin-selective-update and
278 \b cpu/maxmin-selective-update: configure whether the underlying
279 should be lazily updated or not. It should have no impact on the
280 computed timings, but should speed up the computation.
282 It is still possible to disable the \c maxmin-selective-update feature
283 because it can reveal counter-productive in very specific scenarios
284 where the interaction level is high. In particular, if all your
285 communication share a given backbone link, you should disable it:
286 without \c maxmin-selective-update, every communications are updated
287 at each step through a simple loop over them. With that feature
288 enabled, every communications will still get updated in this case
289 (because of the dependency induced by the backbone), but through a
290 complicated pattern aiming at following the actual dependencies.
292 \subsection options_model_precision Numerical precision of the platform models
294 The analytical models handle a lot of floating point values. It is
295 possible to change the epsilon used to update and compare them through
296 the \b maxmin/precision item (default value: 0.00001). Changing it
297 may speedup the simulation by discarding very small actions, at the
298 price of a reduced numerical precision.
300 \subsection options_concurrency_limit Concurrency limit
302 The maximum number of variables per resource can be tuned through
303 the \b maxmin/concurrency-limit item. The default value is -1, meaning that
304 there is no such limitation. You can have as many simultaneous actions per
305 resources as you want. If your simulation presents a very high level of
306 concurrency, it may help to use e.g. 100 as a value here. It means that at
307 most 100 actions can consume a resource at a given time. The extraneous actions
308 are queued and wait until the amount of concurrency of the considered resource
309 lowers under the given boundary.
311 Such limitations help both to the simulation speed and simulation accuracy
312 on highly constrained scenarios, but the simulation speed suffers of this
313 setting on regular (less constrained) scenarios so it is off by default.
315 \subsection options_model_network Configuring the Network model
317 \subsubsection options_model_network_gamma Maximal TCP window size
319 The analytical models need to know the maximal TCP window size to take
320 the TCP congestion mechanism into account. This is set to 20000 by
321 default, but can be changed using the \b network/TCP-gamma item.
323 On linux, this value can be retrieved using the following
324 commands. Both give a set of values, and you should use the last one,
325 which is the maximal size.\verbatim
326 cat /proc/sys/net/ipv4/tcp_rmem # gives the sender window
327 cat /proc/sys/net/ipv4/tcp_wmem # gives the receiver window
330 \subsubsection options_model_network_coefs Correcting important network parameters
332 SimGrid can take network irregularities such as a slow startup or
333 changing behavior depending on the message size into account.
334 You should not change these values unless you really know what you're doing.
336 The corresponding values were computed through data fitting one the
337 timings of packet-level simulators.
340 <a href="http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf">Accuracy Study and Improvement of Network Simulation in the SimGrid Framework</a>
341 for more information about these parameters.
343 If you are using the SMPI model, these correction coefficients are
344 themselves corrected by constant values depending on the size of the
345 exchange. Again, only hardcore experts should bother about this fact.
347 InfiniBand network behavior can be modeled through 3 parameters, as explained in
348 <a href="http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf">this PhD thesis</a>.
349 These factors can be changed through the following option:
352 smpi/IB-penalty-factors:"βe;βs;γs"
355 By default SMPI uses factors computed on the Stampede Supercomputer at TACC, with optimal
356 deployment of processes on nodes.
358 \subsubsection options_model_network_crosstraffic Simulating cross-traffic
360 As of SimGrid v3.7, cross-traffic effects can be taken into account in
361 analytical simulations. It means that ongoing and incoming
362 communication flows are treated independently. In addition, the LV08
363 model adds 0.05 of usage on the opposite direction for each new
364 created flow. This can be useful to simulate some important TCP
365 phenomena such as ack compression.
367 For that to work, your platform must have two links for each
368 pair of interconnected hosts. An example of usable platform is
369 available in <tt>examples/platforms/crosstraffic.xml</tt>.
371 This is activated through the \b network/crosstraffic item, that
372 can be set to 0 (disable this feature) or 1 (enable it).
374 Note that with the default host model this option is activated by default.
376 \subsubsection options_model_network_sendergap Simulating sender gap
378 (this configuration item is experimental and may change or disapear)
380 It is possible to specify a timing gap between consecutive emission on
381 the same network card through the \b network/sender-gap item. This
382 is still under investigation as of writting, and the default value is
383 to wait 10 microseconds (1e-5 seconds) between emissions.
385 \subsubsection options_model_network_asyncsend Simulating asyncronous send
387 (this configuration item is experimental and may change or disapear)
389 It is possible to specify that messages below a certain size will be sent
390 as soon as the call to MPI_Send is issued, without waiting for the
391 correspondant receive. This threshold can be configured through the
392 \b smpi/async-small-thresh item. The default value is 0. This behavior can also be
393 manually set for MSG mailboxes, by setting the receiving mode of the mailbox
394 with a call to \ref MSG_mailbox_set_async . For MSG, all messages sent to this
395 mailbox will have this behavior, so consider using two mailboxes if needed.
397 This value needs to be smaller than or equals to the threshold set at
398 \ref options_model_smpi_detached , because asynchronous messages are
399 meant to be detached as well.
401 \subsubsection options_pls Configuring packet-level pseudo-models
403 When using the packet-level pseudo-models, several specific
404 configuration flags are provided to configure the associated tools.
405 There is by far not enough such SimGrid flags to cover every aspects
406 of the associated tools, since we only added the items that we
407 needed ourselves. Feel free to request more items (or even better:
408 provide patches adding more items).
410 When using NS3, the only existing item is \b ns3/TcpModel,
411 corresponding to the ns3::TcpL4Protocol::SocketType configuration item
412 in NS3. The only valid values (enforced on the SimGrid side) are
413 'NewReno' or 'Reno' or 'Tahoe'.
415 \subsection options_model_storage Configuring the Storage model
417 \subsubsection option_model_storage_maxfd Maximum amount of file descriptors per host
419 Each host maintains a fixed-size array of its file descriptors. You
420 can change its size (1024 by default) through the \b
421 storage/max_file_descriptors item to either enlarge it if your
422 application requires it or to reduce it to save memory space.
424 \section options_modelchecking Configuring the Model-Checking
426 To enable the SimGrid model-checking support the program should
427 be executed using the simgrid-mc wrapper:
429 simgrid-mc ./my_program
432 Safety properties are expressed as assertions using the function
434 void MC_assert(int prop);
437 \subsection options_modelchecking_liveness Specifying a liveness property
439 If you want to specify liveness properties (beware, that's
440 experimental), you have to pass them on the command line, specifying
441 the name of the file containing the property, as formatted by the
445 --cfg=model-check/property:<filename>
448 \subsection options_modelchecking_steps Going for stateful verification
450 By default, the system is backtracked to its initial state to explore
451 another path instead of backtracking to the exact step before the fork
452 that we want to explore (this is called stateless verification). This
453 is done this way because saving intermediate states can rapidly
454 exhaust the available memory. If you want, you can change the value of
455 the <tt>model-check/checkpoint</tt> variable. For example, the
456 following configuration will ask to take a checkpoint every step.
457 Beware, this will certainly explode your memory. Larger values are
458 probably better, make sure to experiment a bit to find the right
459 setting for your specific system.
462 --cfg=model-check/checkpoint:1
465 \subsection options_modelchecking_reduction Specifying the kind of reduction
467 The main issue when using the model-checking is the state space
468 explosion. To counter that problem, several exploration reduction
469 techniques can be used. There is unfortunately no silver bullet here,
470 and the most efficient reduction techniques cannot be applied to any
471 properties. In particular, the DPOR method cannot be applied on
472 liveness properties since it may break some cycles in the exploration
473 that are important to the property validity.
476 --cfg=model-check/reduction:<technique>
479 For now, this configuration variable can take 2 values:
480 * none: Do not apply any kind of reduction (mandatory for now for
482 * dpor: Apply Dynamic Partial Ordering Reduction. Only valid if you
483 verify local safety properties (default value for safety checks).
485 \subsection options_modelchecking_visited model-check/visited, Cycle detection
487 In order to detect cycles, the model-checker needs to check if a new explored
488 state is in fact the same state than a previous one. In order to do this,
489 the model-checker can take a snapshot of each visited state: this snapshot is
490 then used to compare it with subsequent states in the exploration graph.
492 The \b model-check/visited is the maximum number of states which are stored in
493 memory. If the maximum number of snapshotted state is reached some states will
494 be removed from the memory and some cycles might be missed.
496 By default, no state is snapshotted and cycles cannot be detected.
498 \subsection options_modelchecking_termination model-check/termination, Non termination detection
500 The \b model-check/termination configuration item can be used to report if a
501 non-termination execution path has been found. This is a path with a cycle
502 which means that the program might never terminate.
504 This only works in safety mode.
506 This options is disabled by default.
508 \subsection options_modelchecking_dot_output model-check/dot-output, Dot output
510 If set, the \b model-check/dot-output configuration item is the name of a file
511 in which to write a dot file of the path leading the found property (safety or
512 liveness violation) as well as the cycle for liveness properties. This dot file
513 can then fed to the graphviz dot tool to generate an corresponding graphical
516 \subsection options_modelchecking_max_depth model-check/max_depth, Depth limit
518 The \b model-checker/max-depth can set the maximum depth of the exploration
519 graph of the model-checker. If this limit is reached, a logging message is
520 sent and the results might not be exact.
522 By default, there is not depth limit.
524 \subsection options_modelchecking_timeout Handling of timeout
526 By default, the model-checker does not handle timeout conditions: the `wait`
527 operations never time out. With the \b model-check/timeout configuration item
528 set to \b yes, the model-checker will explore timeouts of `wait` operations.
530 \subsection options_modelchecking_comm_determinism Communication determinism
532 The \b model-check/communications-determinism and
533 \b model-check/send-determinism items can be used to select the communication
534 determinism mode of the model-checker which checks determinism properties of
535 the communications of an application.
537 \subsection options_modelchecking_sparse_checkpoint Per page checkpoints
539 When the model-checker is configured to take a snapshot of each explored state
540 (with the \b model-checker/visited item), the memory consumption can rapidly
541 reach GiB ou Tib of memory. However, for many workloads, the memory does not
542 change much between different snapshots and taking a complete copy of each
543 snapshot is a waste of memory.
545 The \b model-check/sparse-checkpoint option item can be set to \b yes in order
546 to avoid making a complete copy of each snapshot: instead, each snapshot will be
547 decomposed in blocks which will be stored separately.
548 If multiple snapshots share the same block (or if the same block
549 is used in the same snapshot), the same copy of the block will be shared leading
550 to a reduction of the memory footprint.
552 For many applications, this option considerably reduces the memory consumption.
553 In somes cases, the model-checker might be slightly slower because of the time
554 taken to manage the metadata about the blocks. In other cases however, this
555 snapshotting strategy will be much faster by reducing the cache consumption.
556 When the memory consumption is important, by avoiding to hit the swap or
557 reducing the swap usage, this option might be much faster than the basic
558 snapshotting strategy.
560 This option is currently disabled by default.
562 \subsection options_mc_perf Performance considerations for the model checker
564 The size of the stacks can have a huge impact on the memory
565 consumption when using model-checking. By default, each snapshot will
566 save a copy of the whole stacks and not only of the part which is
567 really meaningful: you should expect the contribution of the memory
568 consumption of the snapshots to be \f$ \mbox{number of processes}
569 \times \mbox{stack size} \times \mbox{number of states} \f$.
571 The \b model-check/sparse-checkpoint can be used to reduce the memory
572 consumption by trying to share memory between the different snapshots.
574 When compiled against the model checker, the stacks are not
575 protected with guards: if the stack size is too small for your
576 application, the stack will silently overflow on other parts of the
577 memory (see \ref options_virt_guard_size).
579 \subsection options_modelchecking_hash Hashing of the state (experimental)
581 Usually most of the time of the model-checker is spent comparing states. This
582 process is complicated and consumes a lot of bandwidth and cache.
583 In order to speedup the state comparison, the experimental \b model-checker/hash
584 configuration item enables the computation of a hash summarizing as much
585 information of the state as possible into a single value. This hash can be used
586 to avoid most of the comparisons: the costly comparison is then only used when
587 the hashes are identical.
589 Currently most of the state is not included in the hash because the
590 implementation was found to be buggy and this options is not as useful as
591 it could be. For this reason, it is currently disabled by default.
593 \subsection options_modelchecking_recordreplay Record/replay (experimental)
595 As the model-checker keeps jumping at different places in the execution graph,
596 it is difficult to understand what happens when trying to debug an application
597 under the model-checker. Event the output of the program is difficult to
598 interpret. Moreover, the model-checker does not behave nicely with advanced
599 debugging tools such as valgrind. For those reason, to identify a trajectory
600 in the execution graph with the model-checker and replay this trajcetory and
601 without the model-checker black-magic but with more standard tools
602 (such as a debugger, valgrind, etc.). For this reason, Simgrid implements an
603 experimental record/replay functionnality in order to record a trajectory with
604 the model-checker and replay it without the model-checker.
606 When the model-checker finds an interesting path in the application execution
607 graph (where a safety or liveness property is violated), it can generate an
608 identifier for this path. In order to enable this behavious the
609 \b model-check/record must be set to \b yes. By default, this behaviour is not
612 This is an example of output:
615 [ 0.000000] (0:@) Check a safety property
616 [ 0.000000] (0:@) **************************
617 [ 0.000000] (0:@) *** PROPERTY NOT VALID ***
618 [ 0.000000] (0:@) **************************
619 [ 0.000000] (0:@) Counter-example execution trace:
620 [ 0.000000] (0:@) Path = 1/3;1/4
621 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(3)
622 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(4)
623 [ 0.000000] (0:@) Expanded states = 27
624 [ 0.000000] (0:@) Visited states = 68
625 [ 0.000000] (0:@) Executed transitions = 46
628 This path can then be replayed outside of the model-checker (and even in
629 non-MC build of simgrid) by setting the \b model-check/replay item to the given
630 path. The other options should be the same (but the model-checker should
633 The format and meaning of the path may change between different releases so
634 the same release of Simgrid should be used for the record phase and the replay
637 \section options_virt Configuring the User Process Virtualization
639 \subsection options_virt_factory Selecting the virtualization factory
641 In SimGrid, the user code is virtualized in a specific mechanism
642 that allows the simulation kernel to control its execution: when a user
643 process requires a blocking action (such as sending a message), it is
644 interrupted, and only gets released when the simulated clock reaches
645 the point where the blocking operation is done. This is explained
646 graphically in the [relevant tutorial, available online](http://simgrid.gforge.inria.fr/tutorials/simgrid-simix-101.pdf).
648 In SimGrid, the containers in which user processes are virtualized are
649 called contexts. Several context factory are provided, and you can
650 select the one you want to use with the \b contexts/factory
651 configuration item. Some of the following may not exist on your
652 machine because of portability issues. In any case, the default one
653 should be the most effcient one (please report bugs if the
654 auto-detection fails for you). They are approximately sorted here from
655 the slowest to the most efficient:
657 - \b thread: very slow factory using full featured threads (either
658 pthreads or windows native threads). They are slow but very
659 standard. Some debuggers or profilers only work with this factory.
660 - \b java: Java applications are virtualized onto java threads (that
661 are regular pthreads registered to the JVM)
662 - \b ucontext: fast factory using System V contexts (Linux and FreeBSD only)
663 - \b boost: This uses the [context implementation](http://www.boost.org/doc/libs/1_59_0/libs/context/doc/html/index.html)
664 of the boost library for a performance that is comparable to our
665 raw implementation.\nInstall the relevant library (e.g. with the
666 libboost-contexts-dev package on Debian/Ubuntu) and recompile
667 SimGrid. Note that our implementation is not compatible with recent
668 implementations of the library, and it will be hard to fix this since
669 the library's author decided to hide an API that we were using.
670 - \b raw: amazingly fast factory using a context switching mechanism
671 of our own, directly implemented in assembly (only available for x86
672 and amd64 platforms for now) and without any unneeded system call.
674 The main reason to change this setting is when the debugging tools get
675 fooled by the optimized context factories. Threads are the most
676 debugging-friendly contextes, as they allow to set breakpoints
677 anywhere with gdb and visualize backtraces for all processes, in order
678 to debug concurrency issues. Valgrind is also more comfortable with
679 threads, but it should be usable with all factories (but the callgrind
680 tool that really don't like raw and ucontext factories).
682 \subsection options_virt_stacksize Adapting the used stack size
684 Each virtualized used process is executed using a specific system
685 stack. The size of this stack has a huge impact on the simulation
686 scalability, but its default value is rather large. This is because
687 the error messages that you get when the stack size is too small are
688 rather disturbing: this leads to stack overflow (overwriting other
689 stacks), leading to segfaults with corrupted stack traces.
691 If you want to push the scalability limits of your code, you might
692 want to reduce the \b contexts/stack-size item. Its default value
693 is 8192 (in KiB), while our Chord simulation works with stacks as small
694 as 16 KiB, for example. For the thread factory, the default value
695 is the one of the system but you can still change it with this parameter.
697 The operating system should only allocate memory for the pages of the
698 stack which are actually used and you might not need to use this in
699 most cases. However, this setting is very important when using the
700 model checker (see \ref options_mc_perf).
702 \subsection options_virt_guard_size Disabling stack guard pages
704 A stack guard page is usually used which prevents the stack of a given
705 actor from overflowing on another stack. But the performance impact
706 may become prohibitive when the amount of actors increases. The
707 option \b contexts:guard-size is the number of stack guard pages used.
708 By setting it to 0, no guard pages will be used: in this case, you
709 should avoid using small stacks (\b stack-size) as the stack will
710 silently overflow on other parts of the memory.
712 When no stack guard page is created, stacks may then silently overflow
713 on other parts of the memory if their size is too small for the
714 application. This happens:
716 - on Windows systems;
717 - when the model checker is enabled;
718 - and of course when guard pages are explicitely disabled (with \b contexts:guard-size=0).
720 \subsection options_virt_parallel Running user code in parallel
722 Parallel execution of the user code is only considered stable in
723 SimGrid v3.7 and higher, and mostly for MSG simulations. SMPI
724 simulations may well fail in parallel mode. It is described in
725 <a href="http://hal.inria.fr/inria-00602216/">INRIA RR-7653</a>.
727 If you are using the \c ucontext or \c raw context factories, you can
728 request to execute the user code in parallel. Several threads are
729 launched, each of them handling as much user contexts at each run. To
730 actiave this, set the \b contexts/nthreads item to the amount of
731 cores that you have in your computer (or lower than 1 to have
732 the amount of cores auto-detected).
734 Even if you asked several worker threads using the previous option,
735 you can request to start the parallel execution (and pay the
736 associated synchronization costs) only if the potential parallelism is
737 large enough. For that, set the \b contexts/parallel-threshold
738 item to the minimal amount of user contexts needed to start the
739 parallel execution. In any given simulation round, if that amount is
740 not reached, the contexts will be run sequentially directly by the
741 main thread (thus saving the synchronization costs). Note that this
742 option is mainly useful when the grain of the user code is very fine,
743 because our synchronization is now very efficient.
745 When parallel execution is activated, you can choose the
746 synchronization schema used with the \b contexts/synchro item,
747 which value is either:
748 - \b futex: ultra optimized synchronisation schema, based on futexes
749 (fast user-mode mutexes), and thus only available on Linux systems.
750 This is the default mode when available.
751 - \b posix: slow but portable synchronisation using only POSIX
753 - \b busy_wait: not really a synchronisation: the worker threads
754 constantly request new contexts to execute. It should be the most
755 efficient synchronisation schema, but it loads all the cores of your
756 machine for no good reason. You probably prefer the other less
759 \section options_tracing Configuring the tracing subsystem
761 The \ref outcomes_vizu "tracing subsystem" can be configured in several
762 different ways depending on the nature of the simulator (MSG, SimDag,
763 SMPI) and the kind of traces that need to be obtained. See the \ref
764 tracing_tracing_options "Tracing Configuration Options subsection" to
765 get a detailed description of each configuration option.
767 We detail here a simple way to get the traces working for you, even if
768 you never used the tracing API.
771 - Any SimGrid-based simulator (MSG, SimDag, SMPI, ...) and raw traces:
773 --cfg=tracing:yes --cfg=tracing/uncategorized:yes --cfg=triva/uncategorized:uncat.plist
775 The first parameter activates the tracing subsystem, the second
776 tells it to trace host and link utilization (without any
777 categorization) and the third creates a graph configuration file
778 to configure Triva when analysing the resulting trace file.
780 - MSG or SimDag-based simulator and categorized traces (you need to declare categories and classify your tasks according to them)
782 --cfg=tracing:yes --cfg=tracing/categorized:yes --cfg=triva/categorized:cat.plist
784 The first parameter activates the tracing subsystem, the second
785 tells it to trace host and link categorized utilization and the
786 third creates a graph configuration file to configure Triva when
787 analysing the resulting trace file.
789 - SMPI simulator and traces for a space/time view:
793 The <i>-trace</i> parameter for the smpirun script runs the
794 simulation with --cfg=tracing:yes and --cfg=tracing/smpi:yes. Check the
795 smpirun's <i>-help</i> parameter for additional tracing options.
797 Sometimes you might want to put additional information on the trace to
798 correctly identify them later, or to provide data that can be used to
799 reproduce an experiment. You have two ways to do that:
801 - Add a string on top of the trace file as comment:
803 --cfg=tracing/comment:my_simulation_identifier
806 - Add the contents of a textual file on top of the trace file as comment:
808 --cfg=tracing/comment-file:my_file_with_additional_information.txt
811 Please, use these two parameters (for comments) to make reproducible
812 simulations. For additional details about this and all tracing
813 options, check See the \ref tracing_tracing_options.
815 \section options_msg Configuring MSG
817 \subsection options_msg_debug_multiple_use Debugging MSG
819 Sometimes your application may try to send a task that is still being
820 executed somewhere else, making it impossible to send this task. However,
821 for debugging purposes, one may want to know what the other host is/was
822 doing. This option shows a backtrace of the other process.
824 Enable this option by adding
827 --cfg=msg/debug-multiple-use:on
830 \section options_smpi Configuring SMPI
832 The SMPI interface provides several specific configuration items.
833 These are uneasy to see since the code is usually launched through the
834 \c smiprun script directly.
836 \subsection options_smpi_bench smpi/bench: Automatic benchmarking of SMPI code
838 In SMPI, the sequential code is automatically benchmarked, and these
839 computations are automatically reported to the simulator. That is to
840 say that if you have a large computation between a \c MPI_Recv() and a
841 \c MPI_Send(), SMPI will automatically benchmark the duration of this
842 code, and create an execution task within the simulator to take this
843 into account. For that, the actual duration is measured on the host
844 machine and then scaled to the power of the corresponding simulated
845 machine. The variable \b smpi/host-speed allows to specify the
846 computational speed of the host machine (in flop/s) to use when
847 scaling the execution times. It defaults to 20000, but you really want
848 to update it to get accurate simulation results.
850 When the code is constituted of numerous consecutive MPI calls, the
851 previous mechanism feeds the simulation kernel with numerous tiny
852 computations. The \b smpi/cpu-threshold item becomes handy when this
853 impacts badly the simulation performance. It specifies a threshold (in
854 seconds) below which the execution chunks are not reported to the
855 simulation kernel (default value: 1e-6).
858 The option smpi/cpu-threshold ignores any computation time spent
859 below this threshold. SMPI does not consider the \a amount of these
860 computations; there is no offset for this. Hence, by using a
861 value that is too low, you may end up with unreliable simulation
864 In some cases, however, one may wish to disable simulation of
865 application computation. This is the case when SMPI is used not to
866 simulate an MPI applications, but instead an MPI code that performs
867 "live replay" of another MPI app (e.g., ScalaTrace's replay tool,
868 various on-line simulators that run an app at scale). In this case the
869 computation of the replay/simulation logic should not be simulated by
870 SMPI. Instead, the replay tool or on-line simulator will issue
871 "computation events", which correspond to the actual MPI simulation
872 being replayed/simulated. At the moment, these computation events can
873 be simulated using SMPI by calling internal smpi_execute*() functions.
875 To disable the benchmarking/simulation of computation in the simulated
876 application, the variable \b smpi/simulate-computation should be set to no.
879 This option just ignores the timings in your simulation; it still executes
880 the computations itself. If you want to stop SMPI from doing that,
881 you should check the SMPI_SAMPLE macros, documented in the section
882 \ref SMPI_adapting_speed.
884 Solution | Computations actually executed? | Computations simulated ?
885 ---------------------------------- | ------------------------------- | ------------------------
886 --cfg=smpi/simulate-computation:no | Yes | No, never
887 --cfg=smpi/cpu-threshold:42 | Yes, in all cases | Only if it lasts more than 42 seconds
888 SMPI_SAMPLE() macro | Only once per loop nest (see @ref SMPI_adapting_speed "documentation") | Always
890 \subsection options_model_smpi_adj_file smpi/comp-adjustment-file: Slow-down or speed-up parts of your code.
892 This option allows you to pass a file that contains two columns: The first column
893 defines the section that will be subject to a speedup; the second column is the speedup.
899 "exchange_1.f:30:exchange_1.f:130",1.18244559422142
902 The first line is the header - you must include it.
903 The following line means that the code between two consecutive MPI calls on
904 line 30 in exchange_1.f and line 130 in exchange_1.f should receive a speedup
905 of 1.18244559422142. The value for the second column is therefore a speedup, if it is
906 larger than 1 and a slow-down if it is smaller than 1. Nothing will be changed if it is
909 Of course, you can set any arbitrary filenames you want (so the start and end don't have to be
910 in the same file), but be aware that this mechanism only supports @em consecutive calls!
913 Please note that you must pass the \b -trace-call-location flag to smpicc
914 or smpiff, respectively! This flag activates some macro definitions in our
915 mpi.h / mpi.f files that help with obtaining the call location.
917 \subsection options_model_smpi_bw_factor smpi/bw-factor: Bandwidth factors
919 The possible throughput of network links is often dependent on the
920 message sizes, as protocols may adapt to different message sizes. With
921 this option, a series of message sizes and factors are given, helping
922 the simulation to be more realistic. For instance, the current
926 65472:0.940694;15424:0.697866;9376:0.58729;5776:1.08739;3484:0.77493;1426:0.608902;732:0.341987;257:0.338112;0:0.812084
929 So, messages with size 65472 and more will get a total of MAX_BANDWIDTH*0.940694,
930 messages of size 15424 to 65471 will get MAX_BANDWIDTH*0.697866 and so on.
931 Here, MAX_BANDWIDTH denotes the bandwidth of the link.
934 The SimGrid-Team has developed a script to help you determine these
935 values. You can find more information and the download here:
936 1. http://simgrid.gforge.inria.fr/contrib/smpi-calibration-doc.html
937 2. http://simgrid.gforge.inria.fr/contrib/smpi-saturation-doc.html
939 \subsection options_smpi_timing smpi/display-timing: Reporting simulation time
941 \b Default: 0 (false)
943 Most of the time, you run MPI code with SMPI to compute the time it
944 would take to run it on a platform. But since the
945 code is run through the \c smpirun script, you don't have any control
946 on the launcher code, making it difficult to report the simulated time
947 when the simulation ends. If you set the \b smpi/display-timing item
948 to 1, \c smpirun will display this information when the simulation ends. \verbatim
949 Simulation time: 1e3 seconds.
952 \subsection options_smpi_temps smpi/keep-temps: not cleaning up after simulation
954 \b Default: 0 (false)
956 Under some conditions, SMPI generates a lot of temporary files. They
957 usually get cleaned, but you may use this option to not erase these
958 files. This is for example useful when debugging or profiling
959 executions using the dlopen privatization schema, as missing binary
960 files tend to fool the debuggers.
962 \subsection options_model_smpi_lat_factor smpi/lat-factor: Latency factors
964 The motivation and syntax for this option is identical to the motivation/syntax
965 of smpi/bw-factor, see \ref options_model_smpi_bw_factor for details.
967 There is an important difference, though: While smpi/bw-factor \a reduces the
968 actual bandwidth (i.e., values between 0 and 1 are valid), latency factors
969 increase the latency, i.e., values larger than or equal to 1 are valid here.
971 This is the default value:
974 65472:11.6436;15424:3.48845;9376:2.59299;5776:2.18796;3484:1.88101;1426:1.61075;732:1.9503;257:1.95341;0:2.01467
978 The SimGrid-Team has developed a script to help you determine these
979 values. You can find more information and the download here:
980 1. http://simgrid.gforge.inria.fr/contrib/smpi-calibration-doc.html
981 2. http://simgrid.gforge.inria.fr/contrib/smpi-saturation-doc.html
983 \subsection options_smpi_papi_events smpi/papi-events: Trace hardware counters with PAPI
986 This option is experimental and will be subject to change.
987 This feature currently requires superuser privileges, as registers are queried.
988 Only use this feature with code you trust! Call smpirun for instance via
989 smpirun -wrapper "sudo " <your-parameters>
990 or run sudo sh -c "echo 0 > /proc/sys/kernel/perf_event_paranoid"
991 In the later case, sudo will not be required.
994 This option is only available when SimGrid was compiled with PAPI support.
996 This option takes the names of PAPI counters and adds their respective values
997 to the trace files. (See Section \ref tracing_tracing_options.)
999 It is planned to make this feature available on a per-process (or per-thread?) basis.
1000 The first draft, however, just implements a "global" (i.e., for all processes) set
1001 of counters, the "default" set.
1004 --cfg=smpi/papi-events:"default:PAPI_L3_LDM:PAPI_L2_LDM"
1007 \subsection options_smpi_privatization smpi/privatization: Automatic privatization of global variables
1009 MPI executables are usually meant to be executed in separated processes, but SMPI is
1010 executed in only one process. Global variables from executables will be placed
1011 in the same memory zone and shared between processes, causing intricate bugs.
1012 Several options are possible to avoid this, as described in the main
1013 <a href="https://hal.inria.fr/hal-01415484">SMPI publication</a>.
1014 SimGrid provides two ways of automatically privatizing the globals,
1015 and this option allows to choose between them.
1017 - <b>no</b> (default): Do not automatically privatize variables.
1018 - <b>mmap</b> or <b>yes</b>: Runtime automatic switching of the data segments.\n
1019 SMPI stores a copy of each global data segment for each process,
1020 and at each context switch replaces the actual data with its copy
1021 from the right process. No copy actually occures as this mechanism
1022 uses mmap for efficiency. As such, it is for now limited to
1023 systems supporting this functionnality (all Linux and most BSD).\n
1024 Another limitation is that SMPI only accounts for global variables
1025 defined in the executable. If the processes use external global
1026 variables from dynamic libraries, they won't be switched
1027 correctly. The easiest way to solve this is to statically link
1028 against the library with these globals (but you should never
1029 statically link against the simgrid library itself).
1030 - <b>dlopen</b>: Link multiple times against the binary.\n
1031 SMPI loads several copy of the same binary in memory, resulting in
1032 the natural duplication global variables. Since the dynamic linker
1033 refuses to link the same file several times, the binary is copied
1034 in a temporary file before being dl-loaded (it is erased right
1036 Note that this feature is somewhat experimental at time of writing
1037 (v3.16) but seems to work.\n
1038 This approach greatly speeds up the context switching, down to
1039 about 40 CPU cycles with our raw contextes, instead of requesting
1040 several syscalls with the \c mmap approach. Another advantage is
1041 that it permits to run the SMPI contexts in parallel, which is
1042 obviously not possible with the \c mmap approach.\n
1043 Further work may be possible to alleviate the memory and disk
1044 overconsumption. It seems that we could
1045 <a href="https://lwn.net/Articles/415889/">punch holes</a>
1046 in the files before dl-loading them to remove the code and
1047 constants, and mmap these area onto a unique copy. This require
1048 to understand the ELF layout of the file, but would
1049 reduce the disk- and memory- usage to the bare minimum. In
1050 addition, this would reduce the pressure on the CPU caches (in
1051 particular on instruction one).
1054 This configuration option cannot be set in your platform file. You can only
1055 pass it as an argument to smpirun.
1057 \subsection options_model_smpi_detached Simulating MPI detached send
1059 This threshold specifies the size in bytes under which the send will return
1060 immediately. This is different from the threshold detailed in \ref options_model_network_asyncsend
1061 because the message is not effectively sent when the send is posted. SMPI still waits for the
1062 correspondant receive to be posted to perform the communication operation. This threshold can be set
1063 by changing the \b smpi/send-is-detached-thresh item. The default value is 65536.
1065 \subsection options_model_smpi_collectives Simulating MPI collective algorithms
1067 SMPI implements more than 100 different algorithms for MPI collective communication, to accurately
1068 simulate the behavior of most of the existing MPI libraries. The \b smpi/coll-selector item can be used
1069 to use the decision logic of either OpenMPI or MPICH libraries (values: ompi or mpich, by default SMPI
1070 uses naive version of collective operations). Each collective operation can be manually selected with a
1071 \b smpi/collective_name:algo_name. Available algorithms are listed in \ref SMPI_use_colls .
1073 \subsection options_model_smpi_iprobe smpi/iprobe: Inject constant times for calls to MPI_Iprobe
1075 \b Default value: 0.0001
1077 The behavior and motivation for this configuration option is identical with \a smpi/test, see
1078 Section \ref options_model_smpi_test for details.
1080 \subsection options_model_smpi_iprobe_cpu_usage smpi/iprobe-cpu-usage: Reduce speed for iprobe calls
1082 \b Default value: 1 (no change from default behavior)
1084 MPI_Iprobe calls can be heavily used in applications. To account correctly for the energy
1085 cores spend probing, it is necessary to reduce the load that these calls cause inside
1088 For instance, we measured a max power consumption of 220 W for a particular application but
1089 only 180 W while this application was probing. Hence, the correct factor that should
1090 be passed to this option would be 180/220 = 0.81.
1092 \subsection options_model_smpi_init smpi/init: Inject constant times for calls to MPI_Init
1096 The behavior for this configuration option is identical with \a smpi/test, see
1097 Section \ref options_model_smpi_test for details.
1099 \subsection options_model_smpi_ois smpi/ois: Inject constant times for asynchronous send operations
1101 This configuration option works exactly as \a smpi/os, see Section \ref options_model_smpi_os.
1102 Of course, \a smpi/ois is used to account for MPI_Isend instead of MPI_Send.
1104 \subsection options_model_smpi_os smpi/os: Inject constant times for send operations
1106 In several network models such as LogP, send (MPI_Send, MPI_Isend) and receive (MPI_Recv)
1107 operations incur costs (i.e., they consume CPU time). SMPI can factor these costs in as well, but the
1108 user has to configure SMPI accordingly as these values may vary by machine.
1109 This can be done by using smpi/os for MPI_Send operations; for MPI_Isend and
1110 MPI_Recv, use \a smpi/ois and \a smpi/or, respectively. These work exactly as
1113 \a smpi/os can consist of multiple sections; each section takes three values, for example:
1119 Here, the sections are divided by ";" (that is, this example contains two sections).
1120 Furthermore, each section consists of three values.
1122 1. The first value denotes the minimum size for this section to take effect;
1123 read it as "if message size is greater than this value (and other section has a larger
1124 first value that is also smaller than the message size), use this".
1125 In the first section above, this value is "1".
1127 2. The second value is the startup time; this is a constant value that will always
1128 be charged, no matter what the size of the message. In the first section above,
1131 3. The third value is the \a per-byte cost. That is, it is charged for every
1132 byte of the message (incurring cost messageSize*cost_per_byte)
1133 and hence accounts also for larger messages. In the first
1134 section of the example above, this value is "2".
1136 Now, SMPI always checks which section it should take for a given message; that is,
1137 if a message of size 11 is sent with the configuration of the example above, only
1138 the second section will be used, not the first, as the first value of the second
1139 section is closer to the message size. Hence, a message of size 11 incurs the
1140 following cost inside MPI_Send:
1146 As 5 is the startup cost and 1 is the cost per byte.
1149 The order of sections can be arbitrary; they will be ordered internally.
1151 \subsection options_model_smpi_or smpi/or: Inject constant times for receive operations
1153 This configuration option works exactly as \a smpi/os, see Section \ref options_model_smpi_os.
1154 Of course, \a smpi/or is used to account for MPI_Recv instead of MPI_Send.
1156 \subsection options_model_smpi_test smpi/test: Inject constant times for calls to MPI_Test
1158 \b Default value: 0.0001
1160 By setting this option, you can control the amount of time a process sleeps
1161 when MPI_Test() is called; this is important, because SimGrid normally only
1162 advances the time while communication is happening and thus,
1163 MPI_Test will not add to the time, resulting in a deadlock if used as a
1170 MPI_Test(request, flag, status);
1176 Internally, in order to speed up execution, we use a counter to keep track
1177 on how often we already checked if the handle is now valid or not. Hence, we
1178 actually use counter*SLEEP_TIME, that is, the time MPI_Test() causes the process
1179 to sleep increases linearly with the number of previously failed tests. This
1180 behavior can be disabled by setting smpi/grow-injected-times to no. This will
1181 also disable this behavior for MPI_Iprobe.
1184 \subsection options_model_smpi_shared_malloc smpi/shared-malloc: Factorize malloc()s
1188 If your simulation consumes too much memory, you may want to modify
1189 your code so that the working areas are shared by all MPI ranks. For
1190 example, in a bloc-cyclic matrix multiplication, you will only
1191 allocate one set of blocs, and every processes will share them.
1192 Naturally, this will lead to very wrong results, but this will save a
1193 lot of memory so this is still desirable for some studies. For more on
1194 the motivation for that feature, please refer to the
1195 <a href="https://simgrid.github.io/SMPI_CourseWare/topic_understanding_performance/matrixmultiplication/">relevant
1196 section</a> of the SMPI CourseWare (see Activity #2.2 of the pointed
1197 assignment). In practice, change the call to malloc() and free() into
1198 SMPI_SHARED_MALLOC() and SMPI_SHARED_FREE().
1200 SMPI provides 2 algorithms for this feature. The first one, called \c
1201 local, allocates one bloc per call to SMPI_SHARED_MALLOC() in your
1202 code (each call location gets its own bloc) and this bloc is shared
1203 amongst all MPI ranks. This is implemented with the shm_* functions
1204 to create a new POSIX shared memory object (kept in RAM, in /dev/shm)
1205 for each shared bloc.
1207 With the \c global algorithm, each call to SMPI_SHARED_MALLOC()
1208 returns a new adress, but it only points to a shadow bloc: its memory
1209 area is mapped on a 1MiB file on disk. If the returned bloc is of size
1210 N MiB, then the same file is mapped N times to cover the whole bloc.
1211 At the end, no matter how many SMPI_SHARED_MALLOC you do, this will
1212 only consume 1 MiB in memory.
1214 You can disable this behavior and come back to regular mallocs (for
1215 example for debugging purposes) using \c "no" as a value.
1217 If you want to keep private some parts of the buffer, for instance if these
1218 parts are used by the application logic and should not be corrupted, you
1219 can use SMPI_PARTIAL_SHARED_MALLOC(size, offsets, offsets_count).
1224 mem = SMPI_PARTIAL_SHARED_MALLOC(500, {27,42 , 100,200}, 2);
1227 will allocate 500 bytes to mem, such that mem[27..41] and mem[100..199]
1228 are shared and other area remain private.
1230 Then, it can be deallocated by calling SMPI_SHARED_FREE(mem).
1232 When smpi/shared-malloc:global is used, the memory consumption problem
1233 is solved, but it may induce too much load on the kernel's pages table.
1234 In this case, you should use huge pages so that we create only one
1235 entry per Mb of malloced data instead of one entry per 4k.
1236 To activate this, you must mount a hugetlbfs on your system and allocate
1237 at least one huge page:
1241 sudo mount none /home/huge -t hugetlbfs -o rw,mode=0777
1242 sudo sh -c 'echo 1 > /proc/sys/vm/nr_hugepages' # echo more if you need more
1245 Then, you can pass the option --cfg=smpi/shared-malloc-hugepage:/home/huge
1246 to smpirun to actually activate the huge page support in shared mallocs.
1248 \subsection options_model_smpi_wtime smpi/wtime: Inject constant times for calls to MPI_Wtime
1252 By setting this option, you can control the amount of time a process sleeps
1253 when MPI_Wtime() is called; this is important, because SimGrid normally only
1254 advances the time while communication is happening and thus,
1255 MPI_Wtime will not add to the time, resulting in a deadlock if used as a
1261 while(MPI_Wtime() < some_time_bound) {
1266 If the time is never advanced, this loop will clearly never end as MPI_Wtime()
1267 always returns the same value. Hence, pass a (small) value to the smpi/wtime
1268 option to force a call to MPI_Wtime to advance the time as well.
1271 \section options_generic Configuring other aspects of SimGrid
1273 \subsection options_generic_clean_atexit Cleanup before termination
1275 The C / C++ standard contains a function called \b [atexit](http://www.cplusplus.com/reference/cstdlib/atexit/).
1276 atexit registers callbacks, which are called just before the program terminates.
1278 By setting the configuration option clean-atexit to 1 (true), a callback
1279 is registered and will clean up some variables and terminate/cleanup the tracing.
1281 TODO: Add when this should be used.
1283 \subsection options_generic_path XML file inclusion path
1285 It is possible to specify a list of directories to search into for the
1286 \<include\> tag in XML files by using the \b path configuration
1287 item. To add several directory to the path, set the configuration
1288 item several times, as in \verbatim
1289 --cfg=path:toto --cfg=path:tutu
1292 \subsection options_generic_exit Behavior on Ctrl-C
1294 By default, when Ctrl-C is pressed, the status of all existing
1295 simulated processes is displayed before exiting the simulation. This is very useful to debug your
1296 code, but it can reveal troublesome in some cases (such as when the
1297 amount of processes becomes really big). This behavior is disabled
1298 when \b verbose-exit is set to 0 (it is to 1 by default).
1300 \subsection options_exception_cutpath Truncate local path from exception backtrace
1303 --cfg=exceptions/cutpath:1
1306 This configuration option is used to remove the path from the
1307 backtrace shown when an exception is thrown. This is mainly useful for
1308 the tests: the full file path makes the tests not reproducible, and
1309 thus failing as we are currently comparing output. Clearly, the path
1310 used on different machines are almost guaranteed to be different and
1311 hence, the output would mismatch, causing the test to fail.
1313 \section options_log Logging Configuration
1315 It can be done by using XBT. Go to \ref XBT_log for more details.