1 /*! \page options Simgrid options and configurations
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.
12 \section options_using Passing configuration options to the simulators
14 There is several way to pass configuration options to the simulators.
15 The most common way is to use the \c --cfg command line argument. For
16 example, to set the item \c Item to the value \c Value, simply
17 type the following: \verbatim
18 my_simulator --cfg=Item:Value (other arguments)
21 Several \c `--cfg` command line arguments can naturally be used. If you
22 need to include spaces in the argument, don't forget to quote the
23 argument. You can even escape the included quotes (write \' for ' if
24 you have your argument between ').
26 Another solution is to use the \c \<config\> tag in the platform file. The
27 only restriction is that this tag must occure before the first
28 platform element (be it \c \<AS\>, \c \<cluster\>, \c \<peer\> or whatever).
29 The \c \<config\> tag takes an \c id attribute, but it is currently
30 ignored so you don't really need to pass it. The important par is that
31 within that tag, you can pass one or several \c \<prop\> tags to specify
32 the configuration to use. For example, setting \c Item to \c Value
33 can be done by adding the following to the beginning of your platform
37 <prop id="Item" value="Value"/>
41 A last solution is to pass your configuration directly using the C
42 interface. If you happen to use the MSG interface, this is very easy
43 with the MSG_config() function. If you do not use MSG, that's a bit
44 more complex, as you have to mess with the internal configuration set
45 directly as follows. Check the \ref XBT_config "relevant page" for
46 details on all the functions you can use in this context, \c
47 _sg_cfg_set being the only configuration set currently used in
51 #include <xbt/config.h>
53 extern xbt_cfg_t _sg_cfg_set;
55 int main(int argc, char *argv[]) {
58 /* Prefer MSG_config() if you use MSG!! */
59 xbt_cfg_set_parse(_sg_cfg_set,"Item:Value");
65 \section options_model Configuring the platform models
67 \anchor options_storage_model
68 \anchor options_vm_model
69 \subsection options_model_select Selecting the platform models
71 SimGrid comes with several network, CPU and storage models built in, and you
72 can change the used model at runtime by changing the passed
73 configuration. The three main configuration items are given below.
74 For each of these items, passing the special \c help value gives
75 you a short description of all possible values. Also, \c --help-models
76 should provide information about all models for all existing resources.
77 - \b network/model: specify the used network model
78 - \b cpu/model: specify the used CPU model
79 - \b host/model: specify the used host model
80 - \b storage/model: specify the used storage model (there is currently only one such model - this option is hence only useful for future releases)
81 - \b vm_workstation/model: specify the workstation model for virtual machines (there is currently only one such model - this option is hence only useful for future releases)
83 %As of writing, the following network models are accepted. Over
84 the time new models can be added, and some experimental models can be
85 removed; check the values on your simulators for an uptodate
86 information. Note that the CM02 model is described in the research report
87 <a href="ftp://ftp.ens-lyon.fr/pub/LIP/Rapports/RR/RR2002/RR2002-40.ps.gz">A
88 Network Model for Simulation of Grid Application</a> while LV08 is
90 <a href="http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf">Accuracy Study and Improvement of Network Simulation in the SimGrid Framework</a>.
92 - \b LV08 (default one): Realistic network analytic model
93 (slow-start modeled by multiplying latency by 10.4, bandwidth by
94 .92; bottleneck sharing uses a payload of S=8775 for evaluating RTT)
95 - \b Constant: Simplistic network model where all communication
96 take a constant time (one second). This model provides the lowest
97 realism, but is (marginally) faster.
98 - \b SMPI: Realistic network model specifically tailored for HPC
99 settings (accurate modeling of slow start with correction factors on
100 three intervals: < 1KiB, < 64 KiB, >= 64 KiB). See also \ref
101 options_model_network_coefs "this section" for more info.
102 - \b IB: Realistic network model specifically tailored for HPC
103 settings with InfiniBand networks (accurate modeling contention
104 behavior, based on the model explained in
105 http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf).
106 See also \ref options_model_network_coefs "this section" for more info.
107 - \b CM02: Legacy network analytic model (Very similar to LV08, but
108 without corrective factors. The timings of small messages are thus
110 - \b Reno: Model from Steven H. Low using lagrange_solve instead of
111 lmm_solve (experts only; check the code for more info).
112 - \b Reno2: Model from Steven H. Low using lagrange_solve instead of
113 lmm_solve (experts only; check the code for more info).
114 - \b Vegas: Model from Steven H. Low using lagrange_solve instead of
115 lmm_solve (experts only; check the code for more info).
117 If you compiled SimGrid accordingly, you can use packet-level network
118 simulators as network models (see \ref pls). In that case, you have
119 two extra models, described below, and some \ref options_pls "specific
120 additional configuration flags".
121 - \b GTNets: Network pseudo-model using the GTNets simulator instead
123 - \b NS3: Network pseudo-model using the NS3 tcp model instead of an
126 Concerning the CPU, we have only one model for now:
127 - \b Cas01: Simplistic CPU model (time=size/power)
129 The host concept is the aggregation of a CPU with a network
130 card. Three models exists, but actually, only 2 of them are
131 interesting. The "compound" one is simply due to the way our internal
132 code is organized, and can easily be ignored. So at the end, you have
133 two host models: The default one allows to aggregate an
134 existing CPU model with an existing network model, but does not allow
135 parallel tasks because these beasts need some collaboration between
136 the network and CPU model. That is why, ptask_07 is used by default
138 - \b default: Default host model. Currently, CPU:Cas01 and
139 network:LV08 (with cross traffic enabled)
140 - \b compound: Host model that is automatically chosen if
141 you change the network and CPU models
142 - \b ptask_L07: Host model somehow similar to Cas01+CM02 but
143 allowing parallel tasks
145 \subsection options_generic_plugin Plugins
147 SimGrid supports the use of plugins; currently, no known plugins
148 can be activated but there are use-cases where you may want to write
149 your own plugin (for instance, for logging).
151 Plugins can for instance define own classes that inherit from
152 existing classes (for instance, a class "CpuEnergy" inherits from
153 "Cpu" to assess energy consumption).
155 The plugin connects to the code by registering callbacks using
156 ``surf_callback_register`` (see file ``src/surf/plugins/energy.cpp`` for details).
163 This option is case-sensitive: Energy and energy are not the same!
165 \subsection options_model_optim Optimization level of the platform models
167 The network and CPU models that are based on lmm_solve (that
168 is, all our analytical models) accept specific optimization
170 - items \b network/optim and \b CPU/optim (both default to 'Lazy'):
171 - \b Lazy: Lazy action management (partial invalidation in lmm +
172 heap in action remaining).
173 - \b TI: Trace integration. Highly optimized mode when using
174 availability traces (only available for the Cas01 CPU model for
176 - \b Full: Full update of remaining and variables. Slow but may be
177 useful when debugging.
178 - items \b network/maxmin_selective_update and
179 \b cpu/maxmin_selective_update: configure whether the underlying
180 should be lazily updated or not. It should have no impact on the
181 computed timings, but should speed up the computation.
183 It is still possible to disable the \c maxmin_selective_update feature
184 because it can reveal counter-productive in very specific scenarios
185 where the interaction level is high. In particular, if all your
186 communication share a given backbone link, you should disable it:
187 without \c maxmin_selective_update, every communications are updated
188 at each step through a simple loop over them. With that feature
189 enabled, every communications will still get updated in this case
190 (because of the dependency induced by the backbone), but through a
191 complicated pattern aiming at following the actual dependencies.
193 \subsection options_model_precision Numerical precision of the platform models
195 The analytical models handle a lot of floating point values. It is
196 possible to change the epsilon used to update and compare them through
197 the \b maxmin/precision item (default value: 0.00001). Changing it
198 may speedup the simulation by discarding very small actions, at the
199 price of a reduced numerical precision.
201 \subsection options_model_nthreads Parallel threads for model updates
203 By default, Surf computes the analytical models sequentially to share their
204 resources and update their actions. It is possible to run them in parallel,
205 using the \b surf/nthreads item (default value: 1). If you use a
206 negative or null value, the amount of available cores is automatically
207 detected and used instead.
209 Depending on the workload of the models and their complexity, you may get a
210 speedup or a slowdown because of the synchronization costs of threads.
212 \subsection options_model_network Configuring the Network model
214 \subsubsection options_model_network_gamma Maximal TCP window size
216 The analytical models need to know the maximal TCP window size to take
217 the TCP congestion mechanism into account. This is set to 20000 by
218 default, but can be changed using the \b network/TCP_gamma item.
220 On linux, this value can be retrieved using the following
221 commands. Both give a set of values, and you should use the last one,
222 which is the maximal size.\verbatim
223 cat /proc/sys/net/ipv4/tcp_rmem # gives the sender window
224 cat /proc/sys/net/ipv4/tcp_wmem # gives the receiver window
227 \subsubsection options_model_network_coefs Correcting important network parameters
229 SimGrid can take network irregularities such as a slow startup or
230 changing behavior depending on the message size into account.
231 You should not change these values unless you really know what you're doing.
233 The corresponding values were computed through data fitting one the
234 timings of packet-level simulators.
237 <a href="http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf">Accuracy Study and Improvement of Network Simulation in the SimGrid Framework</a>
238 for more information about these parameters.
240 If you are using the SMPI model, these correction coefficients are
241 themselves corrected by constant values depending on the size of the
242 exchange. Again, only hardcore experts should bother about this fact.
244 InfiniBand network behavior can be modeled through 3 parameters, as explained in
245 <a href="http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf">this PhD thesis</a>.
246 These factors can be changed through the following option:
249 smpi/IB_penalty_factors:"βe;βs;γs"
252 By default SMPI uses factors computed on the Stampede Supercomputer at TACC, with optimal
253 deployment of processes on nodes.
255 \subsubsection options_model_network_crosstraffic Simulating cross-traffic
257 %As of SimGrid v3.7, cross-traffic effects can be taken into account in
258 analytical simulations. It means that ongoing and incoming
259 communication flows are treated independently. In addition, the LV08
260 model adds 0.05 of usage on the opposite direction for each new
261 created flow. This can be useful to simulate some important TCP
262 phenomena such as ack compression.
264 For that to work, your platform must have two links for each
265 pair of interconnected hosts. An example of usable platform is
266 available in <tt>examples/msg/gtnets/crosstraffic-p.xml</tt>.
268 This is activated through the \b network/crosstraffic item, that
269 can be set to 0 (disable this feature) or 1 (enable it).
271 Note that with the default host model this option is activated by default.
273 \subsubsection options_model_network_coord Coordinated-based network models
275 When you want to use network coordinates, as it happens when you use
276 an \<AS\> in your platform file with \c Vivaldi as a routing, you must
277 set the \b network/coordinates to \c yes so that all mandatory
278 initialization are done in the simulator.
280 \subsubsection options_model_network_sendergap Simulating sender gap
282 (this configuration item is experimental and may change or disapear)
284 It is possible to specify a timing gap between consecutive emission on
285 the same network card through the \b network/sender_gap item. This
286 is still under investigation as of writting, and the default value is
287 to wait 10 microseconds (1e-5 seconds) between emissions.
289 \subsubsection options_model_network_asyncsend Simulating asyncronous send
291 (this configuration item is experimental and may change or disapear)
293 It is possible to specify that messages below a certain size will be sent
294 as soon as the call to MPI_Send is issued, without waiting for the
295 correspondant receive. This threshold can be configured through the
296 \b smpi/async_small_thres item. The default value is 0. This behavior can also be
297 manually set for MSG mailboxes, by setting the receiving mode of the mailbox
298 with a call to \ref MSG_mailbox_set_async . For MSG, all messages sent to this
299 mailbox will have this behavior, so consider using two mailboxes if needed.
301 This value needs to be smaller than or equals to the threshold set at
302 \ref options_model_smpi_detached , because asynchronous messages are
303 meant to be detached as well.
305 \subsubsection options_pls Configuring packet-level pseudo-models
307 When using the packet-level pseudo-models, several specific
308 configuration flags are provided to configure the associated tools.
309 There is by far not enough such SimGrid flags to cover every aspects
310 of the associated tools, since we only added the items that we
311 needed ourselves. Feel free to request more items (or even better:
312 provide patches adding more items).
314 When using NS3, the only existing item is \b ns3/TcpModel,
315 corresponding to the ns3::TcpL4Protocol::SocketType configuration item
316 in NS3. The only valid values (enforced on the SimGrid side) are
317 'NewReno' or 'Reno' or 'Tahoe'.
319 When using GTNeTS, two items exist:
320 - \b gtnets/jitter, that is a double value to oscillate
321 the link latency, uniformly in random interval
322 [-latency*gtnets_jitter,latency*gtnets_jitter). It defaults to 0.
323 - \b gtnets/jitter_seed, the positive seed used to reproduce jitted
324 results. Its value must be in [1,1e8] and defaults to 10.
326 \section options_modelchecking Configuring the Model-Checking
328 To enable the experimental SimGrid model-checking support the program should
329 be executed with the command line argument
334 Safety properties are expressed as assertions using the function
336 void MC_assert(int prop);
339 \subsection options_modelchecking_liveness Specifying a liveness property
341 If you want to specify liveness properties (beware, that's
342 experimental), you have to pass them on the command line, specifying
343 the name of the file containing the property, as formatted by the
347 --cfg=model-check/property:<filename>
350 Of course, specifying a liveness property enables the model-checking
351 so that you don't have to give <tt>--cfg=model-check:1</tt> in
354 \subsection options_modelchecking_steps Going for stateful verification
356 By default, the system is backtracked to its initial state to explore
357 another path instead of backtracking to the exact step before the fork
358 that we want to explore (this is called stateless verification). This
359 is done this way because saving intermediate states can rapidly
360 exhaust the available memory. If you want, you can change the value of
361 the <tt>model-check/checkpoint</tt> variable. For example, the
362 following configuration will ask to take a checkpoint every step.
363 Beware, this will certainly explode your memory. Larger values are
364 probably better, make sure to experiment a bit to find the right
365 setting for your specific system.
368 --cfg=model-check/checkpoint:1
371 Of course, specifying this option enables the model-checking so that
372 you don't have to give <tt>--cfg=model-check:1</tt> in addition.
374 \subsection options_modelchecking_reduction Specifying the kind of reduction
376 The main issue when using the model-checking is the state space
377 explosion. To counter that problem, several exploration reduction
378 techniques can be used. There is unfortunately no silver bullet here,
379 and the most efficient reduction techniques cannot be applied to any
380 properties. In particular, the DPOR method cannot be applied on
381 liveness properties since it may break some cycles in the exploration
382 that are important to the property validity.
385 --cfg=model-check/reduction:<technique>
388 For now, this configuration variable can take 2 values:
389 * none: Do not apply any kind of reduction (mandatory for now for
391 * dpor: Apply Dynamic Partial Ordering Reduction. Only valid if you
392 verify local safety properties.
394 Of course, specifying a reduction technique enables the model-checking
395 so that you don't have to give <tt>--cfg=model-check:1</tt> in
398 \subsection options_modelchecking_visited model-check/visited, Cycle detection
400 In order to detect cycles, the model-checker needs to check if a new explored
401 state is in fact the same state than a previous one. In order to do this,
402 the model-checker can take a snapshot of each visited state: this snapshot is
403 then used to compare it with subsequent states in the exploration graph.
405 The \b model-check/visited is the maximum number of states which are stored in
406 memory. If the maximum number of snapshotted state is reached some states will
407 be removed from the memory and some cycles might be missed.
409 By default, no state is snapshotted and cycles cannot be detected.
411 \subsection options_modelchecking_termination model-check/termination, Non termination detection
413 The \b model-check/termination configuration item can be used to report if a
414 non-termination execution path has been found. This is a path with a cycle
415 which means that the program might never terminate.
417 This only works in safety mode.
419 This options is disabled by default.
421 \subsection options_modelchecking_dot_output model-check/dot_output, Dot output
423 If set, the \b model-check/dot_output configuration item is the name of a file
424 in which to write a dot file of the path leading the found property (safety or
425 liveness violation) as well as the cycle for liveness properties. This dot file
426 can then fed to the graphviz dot tool to generate an corresponding graphical
429 \subsection options_modelchecking_max_depth model-check/max_depth, Depth limit
431 The \b model-checker/max_depth can set the maximum depth of the exploration
432 graph of the model-checker. If this limit is reached, a logging message is
433 sent and the results might not be exact.
435 By default, there is not depth limit.
437 \subsection options_modelchecking_timeout Handling of timeout
439 By default, the model-checker does not handle timeout conditions: the `wait`
440 operations never time out. With the \b model-check/timeout configuration item
441 set to \b yes, the model-checker will explore timeouts of `wait` operations.
443 \subsection options_modelchecking_comm_determinism Communication determinism
445 The \b model-check/communications_determinism and
446 \b model-check/send_determinism items can be used to select the communication
447 determinism mode of the model-checker which checks determinism properties of
448 the communications of an application.
450 \subsection options_modelchecking_sparse_checkpoint Per page checkpoints
452 When the model-checker is configured to take a snapshot of each explored state
453 (with the \b model-checker/visited item), the memory consumption can rapidly
454 reach GiB ou Tib of memory. However, for many workloads, the memory does not
455 change much between different snapshots and taking a complete copy of each
456 snapshot is a waste of memory.
458 The \b model-check/sparse-checkpoint option item can be set to \b yes in order
459 to avoid making a complete copy of each snapshot: instead, each snapshot will be
460 decomposed in blocks which will be stored separately.
461 If multiple snapshots share the same block (or if the same block
462 is used in the same snapshot), the same copy of the block will be shared leading
463 to a reduction of the memory footprint.
465 For many applications, this option considerably reduces the memory consumption.
466 In somes cases, the model-checker might be slightly slower because of the time
467 taken to manage the metadata about the blocks. In other cases however, this
468 snapshotting strategy will be much faster by reducing the cache consumption.
469 When the memory consumption is important, by avoiding to hit the swap or
470 reducing the swap usage, this option might be much faster than the basic
471 snapshotting strategy.
473 This option is currently disabled by default.
475 \subsection options_mc_perf Performance considerations for the model checker
477 The size of the stacks can have a huge impact on the memory
478 consumption when using model-checking. By default, each snapshot will
479 save a copy of the whole stacks and not only of the part which is
480 really meaningful: you should expect the contribution of the memory
481 consumption of the snapshots to be \f$ \mbox{number of processes}
482 \times \mbox{stack size} \times \mbox{number of states} \f$.
484 The \b model-check/sparse-checkpoint can be used to reduce the memory
485 consumption by trying to share memory between the different snapshots.
487 When compiled against the model checker, the stacks are not
488 protected with guards: if the stack size is too small for your
489 application, the stack will silently overflow on other parts of the
492 \subsection options_modelchecking_hash Hashing of the state (experimental)
494 Usually most of the time of the model-checker is spent comparing states. This
495 process is complicated and consumes a lot of bandwidth and cache.
496 In order to speedup the state comparison, the experimental \b model-checker/hash
497 configuration item enables the computation of a hash summarizing as much
498 information of the state as possible into a single value. This hash can be used
499 to avoid most of the comparisons: the costly comparison is then only used when
500 the hashes are identical.
502 Currently most of the state is not included in the hash because the
503 implementation was found to be buggy and this options is not as useful as
504 it could be. For this reason, it is currently disabled by default.
506 \subsection options_modelchecking_recordreplay Record/replay (experimental)
508 As the model-checker keeps jumping at different places in the execution graph,
509 it is difficult to understand what happens when trying to debug an application
510 under the model-checker. Event the output of the program is difficult to
511 interpret. Moreover, the model-checker does not behave nicely with advanced
512 debugging tools such as valgrind. For those reason, to identify a trajectory
513 in the execution graph with the model-checker and replay this trajcetory and
514 without the model-checker black-magic but with more standard tools
515 (such as a debugger, valgrind, etc.). For this reason, Simgrid implements an
516 experimental record/replay functionnality in order to record a trajectory with
517 the model-checker and replay it without the model-checker.
519 When the model-checker finds an interesting path in the application execution
520 graph (where a safety or liveness property is violated), it can generate an
521 identifier for this path. In order to enable this behavious the
522 \b model-check/record must be set to \b yes. By default, this behaviour is not
525 This is an example of output:
528 [ 0.000000] (0:@) Check a safety property
529 [ 0.000000] (0:@) **************************
530 [ 0.000000] (0:@) *** PROPERTY NOT VALID ***
531 [ 0.000000] (0:@) **************************
532 [ 0.000000] (0:@) Counter-example execution trace:
533 [ 0.000000] (0:@) Path = 1/3;1/4
534 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(3)
535 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(4)
536 [ 0.000000] (0:@) Expanded states = 27
537 [ 0.000000] (0:@) Visited states = 68
538 [ 0.000000] (0:@) Executed transitions = 46
541 This path can then be replayed outside of the model-checker (and even in
542 non-MC build of simgrid) by setting the \b model-check/replay item to the given
543 path. The other options should be the same (but the model-checker should
546 The format and meaning of the path may change between different releases so
547 the same release of Simgrid should be used for the record phase and the replay
550 \section options_virt Configuring the User Process Virtualization
552 \subsection options_virt_factory Selecting the virtualization factory
554 In SimGrid, the user code is virtualized in a specific mecanism
555 allowing the simulation kernel to control its execution: when a user
556 process requires a blocking action (such as sending a message), it is
557 interrupted, and only gets released when the simulated clock reaches
558 the point where the blocking operation is done.
560 In SimGrid, the containers in which user processes are virtualized are
561 called contexts. Several context factory are provided, and you can
562 select the one you want to use with the \b contexts/factory
563 configuration item. Some of the following may not exist on your
564 machine because of portability issues. In any case, the default one
565 should be the most effcient one (please report bugs if the
566 auto-detection fails for you). They are sorted here from the slowest
568 - \b thread: very slow factory using full featured threads (either
569 pthreads or windows native threads)
570 - \b ucontext: fast factory using System V contexts (or a portability
571 layer of our own on top of Windows fibers)
572 - \b raw: amazingly fast factory using a context switching mecanism
573 of our own, directly implemented in assembly (only available for x86
574 and amd64 platforms for now)
576 The only reason to change this setting is when the debugging tools get
577 fooled by the optimized context factories. Threads are the most
578 debugging-friendly contextes, as they allow to set breakpoints anywhere with gdb
579 and visualize backtraces for all processes, in order to debug concurrency issues.
580 Valgrind is also more comfortable with threads, but it should be usable with all factories.
582 \subsection options_virt_stacksize Adapting the used stack size
584 Each virtualized used process is executed using a specific system
585 stack. The size of this stack has a huge impact on the simulation
586 scalability, but its default value is rather large. This is because
587 the error messages that you get when the stack size is too small are
588 rather disturbing: this leads to stack overflow (overwriting other
589 stacks), leading to segfaults with corrupted stack traces.
591 If you want to push the scalability limits of your code, you might
592 want to reduce the \b contexts/stack_size item. Its default value
593 is 8192 (in KiB), while our Chord simulation works with stacks as small
594 as 16 KiB, for example. For the thread factory, the default value
595 is the one of the system, if it is too large/small, it has to be set
598 The operating system should only allocate memory for the pages of the
599 stack which are actually used and you might not need to use this in
600 most cases. However, this setting is very important when using the
601 model checker (see \ref options_mc_perf).
603 In some cases, no stack guard page is used and the stack will silently
604 overflow on other parts of the memory if the stack size is too small
605 for your application. This happens :
607 - on Windows systems;
608 - when the model checker is enabled;
609 - when stack guard pages are explicitely disabled (see \ref options_perf_guard_size).
611 \subsection options_virt_parallel Running user code in parallel
613 Parallel execution of the user code is only considered stable in
614 SimGrid v3.7 and higher. It is described in
615 <a href="http://hal.inria.fr/inria-00602216/">INRIA RR-7653</a>.
617 If you are using the \c ucontext or \c raw context factories, you can
618 request to execute the user code in parallel. Several threads are
619 launched, each of them handling as much user contexts at each run. To
620 actiave this, set the \b contexts/nthreads item to the amount of
621 cores that you have in your computer (or lower than 1 to have
622 the amount of cores auto-detected).
624 Even if you asked several worker threads using the previous option,
625 you can request to start the parallel execution (and pay the
626 associated synchronization costs) only if the potential parallelism is
627 large enough. For that, set the \b contexts/parallel_threshold
628 item to the minimal amount of user contexts needed to start the
629 parallel execution. In any given simulation round, if that amount is
630 not reached, the contexts will be run sequentially directly by the
631 main thread (thus saving the synchronization costs). Note that this
632 option is mainly useful when the grain of the user code is very fine,
633 because our synchronization is now very efficient.
635 When parallel execution is activated, you can choose the
636 synchronization schema used with the \b contexts/synchro item,
637 which value is either:
638 - \b futex: ultra optimized synchronisation schema, based on futexes
639 (fast user-mode mutexes), and thus only available on Linux systems.
640 This is the default mode when available.
641 - \b posix: slow but portable synchronisation using only POSIX
643 - \b busy_wait: not really a synchronisation: the worker threads
644 constantly request new contexts to execute. It should be the most
645 efficient synchronisation schema, but it loads all the cores of your
646 machine for no good reason. You probably prefer the other less
649 \section options_tracing Configuring the tracing subsystem
651 The \ref tracing "tracing subsystem" can be configured in several
652 different ways depending on the nature of the simulator (MSG, SimDag,
653 SMPI) and the kind of traces that need to be obtained. See the \ref
654 tracing_tracing_options "Tracing Configuration Options subsection" to
655 get a detailed description of each configuration option.
657 We detail here a simple way to get the traces working for you, even if
658 you never used the tracing API.
661 - Any SimGrid-based simulator (MSG, SimDag, SMPI, ...) and raw traces:
663 --cfg=tracing:yes --cfg=tracing/uncategorized:yes --cfg=triva/uncategorized:uncat.plist
665 The first parameter activates the tracing subsystem, the second
666 tells it to trace host and link utilization (without any
667 categorization) and the third creates a graph configuration file
668 to configure Triva when analysing the resulting trace file.
670 - MSG or SimDag-based simulator and categorized traces (you need to declare categories and classify your tasks according to them)
672 --cfg=tracing:yes --cfg=tracing/categorized:yes --cfg=triva/categorized:cat.plist
674 The first parameter activates the tracing subsystem, the second
675 tells it to trace host and link categorized utilization and the
676 third creates a graph configuration file to configure Triva when
677 analysing the resulting trace file.
679 - SMPI simulator and traces for a space/time view:
683 The <i>-trace</i> parameter for the smpirun script runs the
684 simulation with --cfg=tracing:yes and --cfg=tracing/smpi:yes. Check the
685 smpirun's <i>-help</i> parameter for additional tracing options.
687 Sometimes you might want to put additional information on the trace to
688 correctly identify them later, or to provide data that can be used to
689 reproduce an experiment. You have two ways to do that:
691 - Add a string on top of the trace file as comment:
693 --cfg=tracing/comment:my_simulation_identifier
696 - Add the contents of a textual file on top of the trace file as comment:
698 --cfg=tracing/comment_file:my_file_with_additional_information.txt
701 Please, use these two parameters (for comments) to make reproducible
702 simulations. For additional details about this and all tracing
703 options, check See the \ref tracing_tracing_options.
705 \section options_msg Configuring MSG
707 \subsection options_msg_debug_multiple_use Debugging MSG
709 Sometimes your application may try to send a task that is still being
710 executed somewhere else, making it impossible to send this task. However,
711 for debugging purposes, one may want to know what the other host is/was
712 doing. This option shows a backtrace of the other process.
714 Enable this option by adding
717 --cfg=msg/debug_multiple_use:on
720 \section options_smpi Configuring SMPI
722 The SMPI interface provides several specific configuration items.
723 These are uneasy to see since the code is usually launched through the
724 \c smiprun script directly.
726 \subsection options_smpi_bench smpi/bench: Automatic benchmarking of SMPI code
728 In SMPI, the sequential code is automatically benchmarked, and these
729 computations are automatically reported to the simulator. That is to
730 say that if you have a large computation between a \c MPI_Recv() and a
731 \c MPI_Send(), SMPI will automatically benchmark the duration of this
732 code, and create an execution task within the simulator to take this
733 into account. For that, the actual duration is measured on the host
734 machine and then scaled to the power of the corresponding simulated
735 machine. The variable \b smpi/running_power allows to specify the
736 computational power of the host machine (in flop/s) to use when
737 scaling the execution times. It defaults to 20000, but you really want
738 to update it to get accurate simulation results.
740 When the code is constituted of numerous consecutive MPI calls, the
741 previous mechanism feeds the simulation kernel with numerous tiny
742 computations. The \b smpi/cpu_threshold item becomes handy when this
743 impacts badly the simulation performance. It specifies a threshold (in
744 seconds) below which the execution chunks are not reported to the
745 simulation kernel (default value: 1e-6).
749 The option smpi/cpu_threshold ignores any computation time spent
750 below this threshold. SMPI does not consider the \a amount of these
751 computations; there is no offset for this. Hence, by using a
752 value that is too low, you may end up with unreliable simulation
755 In some cases, however, one may wish to disable simulation of
756 application computation. This is the case when SMPI is used not to
757 simulate an MPI applications, but instead an MPI code that performs
758 "live replay" of another MPI app (e.g., ScalaTrace's replay tool,
759 various on-line simulators that run an app at scale). In this case the
760 computation of the replay/simulation logic should not be simulated by
761 SMPI. Instead, the replay tool or on-line simulator will issue
762 "computation events", which correspond to the actual MPI simulation
763 being replayed/simulated. At the moment, these computation events can
764 be simulated using SMPI by calling internal smpi_execute*() functions.
766 To disable the benchmarking/simulation of computation in the simulated
767 application, the variable \b
768 smpi/simulation_computation should be set to no
770 \subsection options_model_smpi_bw_factor smpi/bw_factor: Bandwidth factors
772 The possible throughput of network links is often dependent on the
773 message sizes, as protocols may adapt to different message sizes. With
774 this option, a series of message sizes and factors are given, helping
775 the simulation to be more realistic. For instance, the current
779 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
782 So, messages with size 65472 and more will get a total of MAX_BANDWIDTH*0.940694,
783 messages of size 15424 to 65471 will get MAX_BANDWIDTH*0.697866 and so on.
784 Here, MAX_BANDWIDTH denotes the bandwidth of the link.
787 The SimGrid-Team has developed a script to help you determine these
788 values. You can find more information and the download here:
789 1. http://simgrid.gforge.inria.fr/contrib/smpi-calibration-doc.html
790 2. http://simgrid.gforge.inria.fr/contrib/smpi-saturation-doc.html
792 \subsection options_smpi_timing smpi/display_timing: Reporting simulation time
794 \b Default: 0 (false)
796 Most of the time, you run MPI code through SMPI to compute the time it
797 would take to run it on a platform that you don't have. But since the
798 code is run through the \c smpirun script, you don't have any control
799 on the launcher code, making difficult to report the simulated time
800 when the simulation ends. If you set the \b smpi/display_timing item
801 to 1, \c smpirun will display this information when the simulation ends. \verbatim
802 Simulation time: 1e3 seconds.
805 \subsection options_model_smpi_lat_factor smpi/lat_factor: Latency factors
807 The motivation and syntax for this option is identical to the motivation/syntax
808 of smpi/bw_factor, see \ref options_model_smpi_bw_factor for details.
810 There is an important difference, though: While smpi/bw_factor \a reduces the
811 actual bandwidth (i.e., values between 0 and 1 are valid), latency factors
812 increase the latency, i.e., values larger than or equal to 1 are valid here.
814 This is the default value:
817 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
821 The SimGrid-Team has developed a script to help you determine these
822 values. You can find more information and the download here:
823 1. http://simgrid.gforge.inria.fr/contrib/smpi-calibration-doc.html
824 2. http://simgrid.gforge.inria.fr/contrib/smpi-saturation-doc.html
826 \subsection options_smpi_global smpi/privatize_global_variables: Automatic privatization of global variables
828 MPI executables are meant to be executed in separated processes, but SMPI is
829 executed in only one process. Global variables from executables will be placed
830 in the same memory zone and shared between processes, causing hard to find bugs.
831 To avoid this, several options are possible :
832 - Manual edition of the code, for example to add __thread keyword before data
833 declaration, which allows the resulting code to work with SMPI, but only
834 if the thread factory (see \ref options_virt_factory) is used, as global
835 variables are then placed in the TLS (thread local storage) segment.
836 - Source-to-source transformation, to add a level of indirection
837 to the global variables. SMPI does this for F77 codes compiled with smpiff,
838 and used to provide coccinelle scripts for C codes, which are not functional anymore.
839 - Compilation pass, to have the compiler automatically put the data in
841 - Runtime automatic switching of the data segments. SMPI stores a copy of
842 each global data segment for each process, and at each context switch replaces
843 the actual data with its copy from the right process. This mechanism uses mmap,
844 and is for now limited to systems supporting this functionnality (all Linux
845 and some BSD should be compatible).
846 Another limitation is that SMPI only accounts for global variables defined in
847 the executable. If the processes use external global variables from dynamic
848 libraries, they won't be switched correctly. To avoid this, using static
849 linking is advised (but not with the simgrid library, to avoid replicating
850 its own global variables).
852 To use this runtime automatic switching, the variable \b smpi/privatize_global_variables
857 \subsection options_model_smpi_detached Simulating MPI detached send
859 This threshold specifies the size in bytes under which the send will return
860 immediately. This is different from the threshold detailed in \ref options_model_network_asyncsend
861 because the message is not effectively sent when the send is posted. SMPI still waits for the
862 correspondant receive to be posted to perform the communication operation. This threshold can be set
863 by changing the \b smpi/send_is_detached item. The default value is 65536.
865 \subsection options_model_smpi_collectives Simulating MPI collective algorithms
867 SMPI implements more than 100 different algorithms for MPI collective communication, to accurately
868 simulate the behavior of most of the existing MPI libraries. The \b smpi/coll_selector item can be used
869 to use the decision logic of either OpenMPI or MPICH libraries (values: ompi or mpich, by default SMPI
870 uses naive version of collective operations). Each collective operation can be manually selected with a
871 \b smpi/collective_name:algo_name. Available algorithms are listed in \ref SMPI_collective_algorithms .
873 \subsection options_model_smpi_iprobe smpi/iprobe: Inject constant times for calls to MPI_Iprobe
875 \b Default value: 0.0001
877 The behavior and motivation for this configuration option is identical with \a smpi/test, see
878 Section \ref options_model_smpi_test for details.
880 \subsection options_model_smpi_ois smpi/ois: Inject constant times for asynchronous send operations
882 This configuration option works exactly as \a smpi/os, see Section \ref options_model_smpi_os.
883 Of course, \a smpi/ois is used to account for MPI_Isend instead of MPI_Send.
885 \subsection options_model_smpi_os smpi/os: Inject constant times for send operations
887 In several network models such as LogP, send (MPI_Send, MPI_Isend) and receive (MPI_Recv)
888 operations incur costs (i.e., they consume CPU time). SMPI can factor these costs in as well, but the
889 user has to configure SMPI accordingly as these values may vary by machine.
890 This can be done by using smpi/os for MPI_Send operations; for MPI_Isend and
891 MPI_Recv, use \a smpi/ois and \a smpi/or, respectively. These work exactly as
894 \a smpi/os can consist of multiple sections; each section takes three values, for example:
900 Here, the sections are divided by ";" (that is, this example contains two sections).
901 Furthermore, each section consists of three values.
903 1. The first value denotes the minimum size for this section to take effect;
904 read it as "if message size is greater than this value (and other section has a larger
905 first value that is also smaller than the message size), use this".
906 In the first section above, this value is "1".
908 2. The second value is the startup time; this is a constant value that will always
909 be charged, no matter what the size of the message. In the first section above,
912 3. The third value is the \a per-byte cost. That is, it is charged for every
913 byte of the message (incurring cost messageSize*cost_per_byte)
914 and hence accounts also for larger messages. In the first
915 section of the example above, this value is "2".
917 Now, SMPI always checks which section it should take for a given message; that is,
918 if a message of size 11 is sent with the configuration of the example above, only
919 the second section will be used, not the first, as the first value of the second
920 section is closer to the message size. Hence, a message of size 11 incurs the
921 following cost inside MPI_Send:
927 %As 5 is the startup cost and 1 is the cost per byte.
930 The order of sections can be arbitrary; they will be ordered internally.
932 \subsection options_model_smpi_or smpi/or: Inject constant times for receive operations
934 This configuration option works exactly as \a smpi/os, see Section \ref options_model_smpi_os.
935 Of course, \a smpi/or is used to account for MPI_Recv instead of MPI_Send.
937 \subsection options_model_smpi_test smpi/test: Inject constant times for calls to MPI_Test
939 \b Default value: 0.0001
941 By setting this option, you can control the amount of time a process sleeps
942 when MPI_Test() is called; this is important, because SimGrid normally only
943 advances the time while communication is happening and thus,
944 MPI_Test will not add to the time, resulting in a deadlock if used as a
951 MPI_Test(request, flag, status);
957 Internally, in order to speed up execution, we use a counter to keep track
958 on how often we already checked if the handle is now valid or not. Hence, we
959 actually use counter*SLEEP_TIME, that is, the time MPI_Test() causes the process
960 to sleep increases linearly with the number of previously failed testk.
963 \subsection options_model_smpi_use_shared_malloc smpi/use_shared_malloc: Use shared memory
967 SMPI can use shared memory by calling shm_* functions; this might speed up the simulation.
968 This opens or creates a new POSIX shared memory object, kept in RAM, in /dev/shm.
970 If you want to disable this behavior, set the value to 0.
972 \subsection options_model_smpi_wtime smpi/wtime: Inject constant times for calls to MPI_Wtime
976 By setting this option, you can control the amount of time a process sleeps
977 when MPI_Wtime() is called; this is important, because SimGrid normally only
978 advances the time while communication is happening and thus,
979 MPI_Wtime will not add to the time, resulting in a deadlock if used as a
985 while(MPI_Wtime() < some_time_bound) {
990 If the time is never advanced, this loop will clearly never end as MPI_Wtime()
991 always returns the same value. Hence, pass a (small) value to the smpi/wtime
992 option to force a call to MPI_Wtime to advance the time as well.
995 \section options_generic Configuring other aspects of SimGrid
997 \subsection options_generic_clean_atexit Cleanup before termination
999 The C / C++ standard contains a function called \b [atexit](http://www.cplusplus.com/reference/cstdlib/atexit/).
1000 atexit registers callbacks, which are called just before the program terminates.
1002 By setting the configuration option clean_atexit to 1 (true), a callback
1003 is registered and will clean up some variables and terminate/cleanup the tracing.
1005 TODO: Add when this should be used.
1007 \subsection options_generic_path XML file inclusion path
1009 It is possible to specify a list of directories to search into for the
1010 \<include\> tag in XML files by using the \b path configuration
1011 item. To add several directory to the path, set the configuration
1012 item several times, as in \verbatim
1013 --cfg=path:toto --cfg=path:tutu
1016 \subsection options_generic_exit Behavior on Ctrl-C
1018 By default, when Ctrl-C is pressed, the status of all existing
1019 simulated processes is displayed before exiting the simulation. This is very useful to debug your
1020 code, but it can reveal troublesome in some cases (such as when the
1021 amount of processes becomes really big). This behavior is disabled
1022 when \b verbose-exit is set to 0 (it is to 1 by default).
1024 \subsection options_exception_cutpath Truncate local path from exception backtrace
1026 <b>This configuration option is an internal option and should normally not be used
1027 by the user.</b> It is used to remove the path from the backtrace
1028 shown when an exception is thrown; if we didn't remove this part, the tests
1029 testing the exception parts of simgrid would fail on most machines, as we are
1030 currently comparing output. Clearly, the path used on different machines are almost
1031 guaranteed to be different and hence, the output would
1032 mismatch, causing the test to fail.
1034 \section options_log Logging Configuration
1036 It can be done by using XBT. Go to \ref XBT_log for more details.
1038 \section options_perf Performance optimizations
1040 \subsection options_perf_context Context factory
1042 In order to achieve higher performance, you might want to use the raw
1043 context factory which avoids any system call when switching between
1044 tasks. If it is not possible you might use ucontext instead.
1046 \subsection options_perf_guard_size Disabling stack guard pages
1048 A stack guard page is usually used which prevents the stack from
1049 overflowing on other parts of the memory. However this might have a
1050 performance impact if a huge number of processes is created. The
1051 option \b contexts:guard_size is the number of stack guard pages
1052 used. By setting it to 0, no guard pages will be used: in this case,
1053 you should avoid using small stacks (\b stack_size) as the stack will
1054 silently overflow on other parts of the memory.
1056 \section options_index Index of all existing configuration options
1059 Almost all options are defined in <i>src/simgrid/sg_config.c</i>. You may
1060 want to check this file, too, but this index should be somewhat complete
1061 for the moment (May 2015).
1064 \b Please \b note: You can also pass the command-line option "--help" and
1065 "--help-cfg" to an executable that uses simgrid.
1067 - \c clean_atexit: \ref options_generic_clean_atexit
1069 - \c contexts/factory: \ref options_virt_factory
1070 - \c contexts/guard_size: \ref options_virt_parallel
1071 - \c contexts/nthreads: \ref options_virt_parallel
1072 - \c contexts/parallel_threshold: \ref options_virt_parallel
1073 - \c contexts/stack_size: \ref options_virt_stacksize
1074 - \c contexts/synchro: \ref options_virt_parallel
1076 - \c cpu/maxmin_selective_update: \ref options_model_optim
1077 - \c cpu/model: \ref options_model_select
1078 - \c cpu/optim: \ref options_model_optim
1080 - \c exception/cutpath: \ref options_exception_cutpath
1082 - \c gtnets/jitter: \ref options_pls
1083 - \c gtnets/jitter_seed: \ref options_pls
1085 - \c host/model: \ref options_model_select
1087 - \c maxmin/precision: \ref options_model_precision
1089 - \c msg/debug_multiple_use: \ref options_msg_debug_multiple_use
1091 - \c model-check: \ref options_modelchecking
1092 - \c model-check/checkpoint: \ref options_modelchecking_steps
1093 - \c model-check/communications_determinism: \ref options_modelchecking_comm_determinism
1094 - \c model-check/send_determinism: \ref options_modelchecking_comm_determinism
1095 - \c model-check/dot_output: \ref options_modelchecking_dot_output
1096 - \c model-check/hash: \ref options_modelchecking_hash
1097 - \c model-check/property: \ref options_modelchecking_liveness
1098 - \c model-check/max_depth: \ref options_modelchecking_max_depth
1099 - \c model-check/record: \ref options_modelchecking_recordreplay
1100 - \c model-check/reduction: \ref options_modelchecking_reduction
1101 - \c model-check/replay: \ref options_modelchecking_recordreplay
1102 - \c model-check/send_determinism: \ref options_modelchecking_sparse_checkpoint
1103 - \c model-check/sparse-checkpoint: \ref options_modelchecking_sparse_checkpoint
1104 - \c model-check/termination: \ref options_modelchecking_termination
1105 - \c model-check/timeout: \ref options_modelchecking_timeout
1106 - \c model-check/visited: \ref options_modelchecking_visited
1108 - \c network/bandwidth_factor: \ref options_model_network_coefs
1109 - \c network/coordinates: \ref options_model_network_coord
1110 - \c network/crosstraffic: \ref options_model_network_crosstraffic
1111 - \c network/latency_factor: \ref options_model_network_coefs
1112 - \c network/maxmin_selective_update: \ref options_model_optim
1113 - \c network/model: \ref options_model_select
1114 - \c network/optim: \ref options_model_optim
1115 - \c network/sender_gap: \ref options_model_network_sendergap
1116 - \c network/TCP_gamma: \ref options_model_network_gamma
1117 - \c network/weight_S: \ref options_model_network_coefs
1119 - \c ns3/TcpModel: \ref options_pls
1120 - \c path: \ref options_generic_path
1121 - \c plugin: \ref options_generic_plugin
1123 - \c surf/nthreads: \ref options_model_nthreads
1124 - \c surf/precision: \ref options_model_precision
1126 - \c <b>For collective operations of SMPI, please refer to Section \ref options_index_smpi_coll</b>
1127 - \c smpi/async_small_thres: \ref options_model_network_asyncsend
1128 - \c smpi/bw_factor: \ref options_model_smpi_bw_factor
1129 - \c smpi/coll_selector: \ref options_model_smpi_collectives
1130 - \c smpi/cpu_threshold: \ref options_smpi_bench
1131 - \c smpi/display_timing: \ref options_smpi_timing
1132 - \c smpi/lat_factor: \ref options_model_smpi_lat_factor
1133 - \c smpi/IB_penalty_factors: \ref options_model_network_coefs
1134 - \c smpi/iprobe: \ref options_model_smpi_iprobe
1135 - \c smpi/ois: \ref options_model_smpi_ois
1136 - \c smpi/or: \ref options_model_smpi_or
1137 - \c smpi/os: \ref options_model_smpi_os
1138 - \c smpi/privatize_global_variables: \ref options_smpi_global
1139 - \c smpi/running_power: \ref options_smpi_bench
1140 - \c smpi/send_is_detached_thresh: \ref options_model_smpi_detached
1141 - \c smpi/simulation_computation: \ref options_smpi_bench
1142 - \c smpi/test: \ref options_model_smpi_test
1143 - \c smpi/use_shared_malloc: \ref options_model_smpi_use_shared_malloc
1144 - \c smpi/wtime: \ref options_model_smpi_wtime
1146 - \c <b>Tracing configuration options can be found in Section \ref tracing_tracing_options</b>.
1148 - \c storage/model: \ref options_storage_model
1149 - \c verbose-exit: \ref options_generic_exit
1151 - \c vm_workstation/model: \ref options_vm_workstation_model
1153 \subsection options_index_smpi_coll Index of SMPI collective algorithms options
1155 TODO: All available collective algorithms will be made available via the ``smpirun --help-coll`` command.