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 \subsection options_model_select Selecting the platform models
69 SimGrid comes with several network and CPU models built in, and you
70 can change the used model at runtime by changing the passed
71 configuration. The three main configuration items are given below.
72 For each of these items, passing the special \c help value gives
73 you a short description of all possible values. Also, \c --help-models
74 should provide information about all models for all existing resources.
75 - \b network/model: specify the used network model
76 - \b cpu/model: specify the used CPU model
77 - \b workstation/model: specify the used workstation model
79 %As of writing, the following network models are accepted. Over
80 the time new models can be added, and some experimental models can be
81 removed; check the values on your simulators for an uptodate
82 information. Note that the CM02 model is described in the research report
83 <a href="ftp://ftp.ens-lyon.fr/pub/LIP/Rapports/RR/RR2002/RR2002-40.ps.gz">A
84 Network Model for Simulation of Grid Application</a> while LV08 is
86 <a href="http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf">Accuracy Study and Improvement of Network Simulation in the SimGrid Framework</a>.
88 - \b LV08 (default one): Realistic network analytic model
89 (slow-start modeled by multiplying latency by 10.4, bandwidth by
90 .92; bottleneck sharing uses a payload of S=8775 for evaluating RTT)
91 - \b Constant: Simplistic network model where all communication
92 take a constant time (one second). This model provides the lowest
93 realism, but is (marginally) faster.
94 - \b SMPI: Realistic network model specifically tailored for HPC
95 settings (accurate modeling of slow start with correction factors on
96 three intervals: < 1KiB, < 64 KiB, >= 64 KiB). See also \ref
97 options_model_network_coefs "this section" for more info.
98 - \b IB: Realistic network model specifically tailored for HPC
99 settings with InfiniBand networks (accurate modeling contention
100 behavior, based on the model explained in
101 http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf).
102 See also \ref options_model_network_coefs "this section" for more info.
103 - \b CM02: Legacy network analytic model (Very similar to LV08, but
104 without corrective factors. The timings of small messages are thus
106 - \b Reno: Model from Steven H. Low using lagrange_solve instead of
107 lmm_solve (experts only; check the code for more info).
108 - \b Reno2: Model from Steven H. Low using lagrange_solve instead of
109 lmm_solve (experts only; check the code for more info).
110 - \b Vegas: Model from Steven H. Low using lagrange_solve instead of
111 lmm_solve (experts only; check the code for more info).
113 If you compiled SimGrid accordingly, you can use packet-level network
114 simulators as network models (see \ref pls). In that case, you have
115 two extra models, described below, and some \ref options_pls "specific
116 additional configuration flags".
117 - \b GTNets: Network pseudo-model using the GTNets simulator instead
119 - \b NS3: Network pseudo-model using the NS3 tcp model instead of an
122 Concerning the CPU, we have only one model for now:
123 - \b Cas01: Simplistic CPU model (time=size/power)
125 The workstation concept is the aggregation of a CPU with a network
126 card. Three models exists, but actually, only 2 of them are
127 interesting. The "compound" one is simply due to the way our internal
128 code is organized, and can easily be ignored. So at the end, you have
129 two workstation models: The default one allows to aggregate an
130 existing CPU model with an existing network model, but does not allow
131 parallel tasks because these beasts need some collaboration between
132 the network and CPU model. That is why, ptask_07 is used by default
134 - \b default: Default workstation model. Currently, CPU:Cas01 and
135 network:LV08 (with cross traffic enabled)
136 - \b compound: Workstation model that is automatically chosen if
137 you change the network and CPU models
138 - \b ptask_L07: Workstation model somehow similar to Cas01+CM02 but
139 allowing parallel tasks
141 \subsection options_model_optim Optimization level of the platform models
143 The network and CPU models that are based on lmm_solve (that
144 is, all our analytical models) accept specific optimization
146 - items \b network/optim and \b CPU/optim (both default to 'Lazy'):
147 - \b Lazy: Lazy action management (partial invalidation in lmm +
148 heap in action remaining).
149 - \b TI: Trace integration. Highly optimized mode when using
150 availability traces (only available for the Cas01 CPU model for
152 - \b Full: Full update of remaining and variables. Slow but may be
153 useful when debugging.
154 - items \b network/maxmin_selective_update and
155 \b cpu/maxmin_selective_update: configure whether the underlying
156 should be lazily updated or not. It should have no impact on the
157 computed timings, but should speed up the computation.
159 It is still possible to disable the \c maxmin_selective_update feature
160 because it can reveal counter-productive in very specific scenarios
161 where the interaction level is high. In particular, if all your
162 communication share a given backbone link, you should disable it:
163 without \c maxmin_selective_update, every communications are updated
164 at each step through a simple loop over them. With that feature
165 enabled, every communications will still get updated in this case
166 (because of the dependency induced by the backbone), but through a
167 complicated pattern aiming at following the actual dependencies.
169 \subsection options_model_precision Numerical precision of the platform models
171 The analytical models handle a lot of floating point values. It is
172 possible to change the epsilon used to update and compare them through
173 the \b maxmin/precision item (default value: 0.00001). Changing it
174 may speedup the simulation by discarding very small actions, at the
175 price of a reduced numerical precision.
177 \subsection options_model_nthreads Parallel threads for model updates
179 By default, Surf computes the analytical models sequentially to share their
180 resources and update their actions. It is possible to run them in parallel,
181 using the \b surf/nthreads item (default value: 1). If you use a
182 negative or null value, the amount of available cores is automatically
183 detected and used instead.
185 Depending on the workload of the models and their complexity, you may get a
186 speedup or a slowdown because of the synchronization costs of threads.
188 \subsection options_model_network Configuring the Network model
190 \subsubsection options_model_network_gamma Maximal TCP window size
192 The analytical models need to know the maximal TCP window size to take
193 the TCP congestion mechanism into account. This is set to 20000 by
194 default, but can be changed using the \b network/TCP_gamma item.
196 On linux, this value can be retrieved using the following
197 commands. Both give a set of values, and you should use the last one,
198 which is the maximal size.\verbatim
199 cat /proc/sys/net/ipv4/tcp_rmem # gives the sender window
200 cat /proc/sys/net/ipv4/tcp_wmem # gives the receiver window
203 \subsubsection options_model_network_coefs Correcting important network parameters
205 SimGrid can take network irregularities such as a slow startup or
206 changing behavior depending on the message size into account.
207 You should not change these values unless you really know what you're doing.
209 The corresponding values were computed through data fitting one the
210 timings of packet-level simulators.
213 <a href="http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf">Accuracy Study and Improvement of Network Simulation in the SimGrid Framework</a>
214 for more information about these parameters.
216 If you are using the SMPI model, these correction coefficients are
217 themselves corrected by constant values depending on the size of the
218 exchange. Again, only hardcore experts should bother about this fact.
220 InfiniBand network behavior can be modeled through 3 parameters, as explained in
221 <a href="http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf">this PhD thesis</a>.
222 These factors can be changed through the following option:
225 smpi/IB_penalty_factors:"βe;βs;γs"
228 By default SMPI uses factors computed on the Stampede Supercomputer at TACC, with optimal
229 deployment of processes on nodes.
231 \subsubsection options_model_network_crosstraffic Simulating cross-traffic
233 %As of SimGrid v3.7, cross-traffic effects can be taken into account in
234 analytical simulations. It means that ongoing and incoming
235 communication flows are treated independently. In addition, the LV08
236 model adds 0.05 of usage on the opposite direction for each new
237 created flow. This can be useful to simulate some important TCP
238 phenomena such as ack compression.
240 For that to work, your platform must have two links for each
241 pair of interconnected hosts. An example of usable platform is
242 available in <tt>examples/msg/gtnets/crosstraffic-p.xml</tt>.
244 This is activated through the \b network/crosstraffic item, that
245 can be set to 0 (disable this feature) or 1 (enable it).
247 Note that with the default workstation model this option is activated by default.
249 \subsubsection options_model_network_coord Coordinated-based network models
251 When you want to use network coordinates, as it happens when you use
252 an \<AS\> in your platform file with \c Vivaldi as a routing, you must
253 set the \b network/coordinates to \c yes so that all mandatory
254 initialization are done in the simulator.
256 \subsubsection options_model_network_sendergap Simulating sender gap
258 (this configuration item is experimental and may change or disapear)
260 It is possible to specify a timing gap between consecutive emission on
261 the same network card through the \b network/sender_gap item. This
262 is still under investigation as of writting, and the default value is
263 to wait 10 microseconds (1e-5 seconds) between emissions.
265 \subsubsection options_model_network_asyncsend Simulating asyncronous send
267 (this configuration item is experimental and may change or disapear)
269 It is possible to specify that messages below a certain size will be sent
270 as soon as the call to MPI_Send is issued, without waiting for the
271 correspondant receive. This threshold can be configured through the
272 \b smpi/async_small_thres item. The default value is 0. This behavior can also be
273 manually set for MSG mailboxes, by setting the receiving mode of the mailbox
274 with a call to \ref MSG_mailbox_set_async . For MSG, all messages sent to this
275 mailbox will have this behavior, so consider using two mailboxes if needed.
277 This value needs to be smaller than or equals to the threshold set at
278 \ref options_model_smpi_detached , because asynchronous messages are
279 meant to be detached as well.
281 \subsubsection options_pls Configuring packet-level pseudo-models
283 When using the packet-level pseudo-models, several specific
284 configuration flags are provided to configure the associated tools.
285 There is by far not enough such SimGrid flags to cover every aspects
286 of the associated tools, since we only added the items that we
287 needed ourselves. Feel free to request more items (or even better:
288 provide patches adding more items).
290 When using NS3, the only existing item is \b ns3/TcpModel,
291 corresponding to the ns3::TcpL4Protocol::SocketType configuration item
292 in NS3. The only valid values (enforced on the SimGrid side) are
293 'NewReno' or 'Reno' or 'Tahoe'.
295 When using GTNeTS, two items exist:
296 - \b gtnets/jitter, that is a double value to oscillate
297 the link latency, uniformly in random interval
298 [-latency*gtnets_jitter,latency*gtnets_jitter). It defaults to 0.
299 - \b gtnets/jitter_seed, the positive seed used to reproduce jitted
300 results. Its value must be in [1,1e8] and defaults to 10.
302 \section options_modelchecking Configuring the Model-Checking
304 To enable the experimental SimGrid model-checking support the program should
305 be executed with the command line argument
310 Safety properties are expressed as assertions using the function
312 void MC_assert(int prop);
315 \subsection options_modelchecking_liveness Specifying a liveness property
317 If you want to specify liveness properties (beware, that's
318 experimental), you have to pass them on the command line, specifying
319 the name of the file containing the property, as formatted by the
323 --cfg=model-check/property:<filename>
326 Of course, specifying a liveness property enables the model-checking
327 so that you don't have to give <tt>--cfg=model-check:1</tt> in
330 \subsection options_modelchecking_steps Going for stateful verification
332 By default, the system is backtracked to its initial state to explore
333 another path instead of backtracking to the exact step before the fork
334 that we want to explore (this is called stateless verification). This
335 is done this way because saving intermediate states can rapidly
336 exhaust the available memory. If you want, you can change the value of
337 the <tt>model-check/checkpoint</tt> variable. For example, the
338 following configuration will ask to take a checkpoint every step.
339 Beware, this will certainly explode your memory. Larger values are
340 probably better, make sure to experiment a bit to find the right
341 setting for your specific system.
344 --cfg=model-check/checkpoint:1
347 Of course, specifying this option enables the model-checking so that
348 you don't have to give <tt>--cfg=model-check:1</tt> in addition.
350 \subsection options_modelchecking_reduction Specifying the kind of reduction
352 The main issue when using the model-checking is the state space
353 explosion. To counter that problem, several exploration reduction
354 techniques can be used. There is unfortunately no silver bullet here,
355 and the most efficient reduction techniques cannot be applied to any
356 properties. In particular, the DPOR method cannot be applied on
357 liveness properties since it may break some cycles in the exploration
358 that are important to the property validity.
361 --cfg=model-check/reduction:<technique>
364 For now, this configuration variable can take 2 values:
365 * none: Do not apply any kind of reduction (mandatory for now for
367 * dpor: Apply Dynamic Partial Ordering Reduction. Only valid if you
368 verify local safety properties.
370 Of course, specifying a reduction technique enables the model-checking
371 so that you don't have to give <tt>--cfg=model-check:1</tt> in
374 \subsection options_modelchecking_visited model-check/visited, Cycle detection
376 In order to detect cycles, the model-checker needs to check if a new explored
377 state is in fact the same state than a previous one. In order to do this,
378 the model-checker can take a snapshot of each visited state: this snapshot is
379 then used to compare it with subsequent states in the exploration graph.
381 The \b model-check/visited is the maximum number of states which are stored in
382 memory. If the maximum number of snapshotted state is reached some states will
383 be removed from the memory and some cycles might be missed.
385 By default, no state is snapshotted and cycles cannot be detected.
387 \subsection options_modelchecking_termination, model-check/termination, Non termination detection
389 The \b model-check/termination configuration item can be used to report if a
390 non-termination execution path has been found. This is a path with a cycle
391 which means that the program might never terminate.
393 This only works in safety mode.
395 This options is disabled by default.
397 \subsection options_modelchecking_max_depth model-check/max_depth, Depth limit
399 The \b model-checker/max_depth can set the maximum depth of the exploration
400 graph of the model-checker. If this limit is reached, a logging message is
401 sent and the results might not be exact.
403 By default, there is not depth limit.
405 \subsection options_modelchecking_timeout Handling of timeout
407 By default, the model-checker does not handle timeout conditions: the `wait`
408 operations never time out. With the \b model-check/timeout configuration item
409 set to \b yes, the model-checker will explore timeouts of `wait` operations.
411 \subsection options_modelchecking_comm_determinism Communication determinism
413 The \b model-check/communications_determinism and
414 \b model-check/send_determinism items can be used to select the communication
415 determinism mode of the model-checker which checks determinism properties of
416 the communications of an application.
418 \subsection options_modelchecking_sparse_checkpoint Per page checkpoints
420 When the model-checker is configured to take a snapshot of each explored state
421 (with the \b model-checker/visited item), the memory consumption can rapidly
422 reach GiB ou Tib of memory. However, for many workloads, the memory does not
423 change much between different snapshots and taking a complete copy of each
424 snapshot is a waste of memory.
426 The \b model-check/sparse-checkpoint option item can be set to \b yes in order
427 to avoid making a complete copy of each snapshot: instead, each snapshot will be
428 decomposed in blocks which will be stored separately.
429 If multiple snapshots share the same block (or if the same block
430 is used in the same snapshot), the same copy of the block will be shared leading
431 to a reduction of the memory footprint.
433 For many applications, this option considerably reduces the memory consumption.
434 In somes cases, the model-checker might be slightly slower because of the time
435 taken to manage the metadata about the blocks. In other cases however, this
436 snapshotting strategy will be much faster by reducing the cache consumption.
437 When the memory consumption is important, by avoiding to hit the swap or
438 reducing the swap usage, this option might be much faster than the basic
439 snapshotting strategy.
441 This option is currently disabled by default.
443 \subsection options_mc_perf Performance considerations for the model checker
445 The size of the stacks can have a huge impact on the memory
446 consumption when using model-checking. By default, each snapshot will
447 save a copy of the whole stacks and not only of the part which is
448 really meaningful: you should expect the contribution of the memory
449 consumption of the snapshots to be \f$ \mbox{number of processes}
450 \times \mbox{stack size} \times \mbox{number of states} \f$.
452 The \b model-check/sparse-checkpoint can be used to reduce the memory
453 consumption by trying to share memory between the different snapshots.
455 When compiled against the model checker, the stacks are not
456 protected with guards: if the stack size is too small for your
457 application, the stack will silently overflow on other parts of the
460 \subsection options_modelchecking_hash Hashing of the state (experimental)
462 Usually most of the time of the model-checker is spent comparing states. This
463 process is complicated and consumes a lot of bandwidth and cache.
464 In order to speedup the state comparison, the experimental \b model-checker/hash
465 configuration item enables the computation of a hash summarizing as much
466 information of the state as possible into a single value. This hash can be used
467 to avoid most of the comparisons: the costly comparison is then only used when
468 the hashes are identical.
470 Currently most of the state is not included in the hash because the
471 implementation was found to be buggy and this options is not as useful as
472 it could be. For this reason, it is currently disabled by default.
474 \subsection options_recordreplay Record/replay (experimental)
476 As the model-checker keeps jumping at different places in the execution graph,
477 it is difficult to understand what happens when trying to debug an application
478 under the model-checker. Event the output of the program is difficult to
479 interpret. Moreover, the model-checker does not behave nicely with advanced
480 debugging tools such as valgrind. For those reason, to identify a trajectory
481 in the execution graph with the model-checker and replay this trajcetory and
482 without the model-checker black-magic but with more standard tools
483 (such as a debugger, valgrind, etc.). For this reason, Simgrid implements an
484 experimental record/replay functionnality in order to record a trajectory with
485 the model-checker and replay it without the model-checker.
487 When the model-checker finds an interesting path in the application execution
488 graph (where a safety or liveness property is violated), it can generate an
489 identifier for this path. In order to enable this behavious the
490 \b model-check/record must be set to \b yes. By default, this behaviour is not
493 This is an example of output:
496 [ 0.000000] (0:@) Check a safety property
497 [ 0.000000] (0:@) **************************
498 [ 0.000000] (0:@) *** PROPERTY NOT VALID ***
499 [ 0.000000] (0:@) **************************
500 [ 0.000000] (0:@) Counter-example execution trace:
501 [ 0.000000] (0:@) Path = 1/3;1/4
502 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(3)
503 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(4)
504 [ 0.000000] (0:@) Expanded states = 27
505 [ 0.000000] (0:@) Visited states = 68
506 [ 0.000000] (0:@) Executed transitions = 46
509 This path can then be replayed outside of the model-checker (and even in
510 non-MC build of simgrid) by setting the \b model-check/replay item to the given
511 path. The other options should be the same (but the model-checker should
514 The format and meaning of the path may change between different releases so
515 the same release of Simgrid should be used for the record phase and the replay
518 \section options_virt Configuring the User Process Virtualization
520 \subsection options_virt_factory Selecting the virtualization factory
522 In SimGrid, the user code is virtualized in a specific mecanism
523 allowing the simulation kernel to control its execution: when a user
524 process requires a blocking action (such as sending a message), it is
525 interrupted, and only gets released when the simulated clock reaches
526 the point where the blocking operation is done.
528 In SimGrid, the containers in which user processes are virtualized are
529 called contexts. Several context factory are provided, and you can
530 select the one you want to use with the \b contexts/factory
531 configuration item. Some of the following may not exist on your
532 machine because of portability issues. In any case, the default one
533 should be the most effcient one (please report bugs if the
534 auto-detection fails for you). They are sorted here from the slowest
536 - \b thread: very slow factory using full featured threads (either
537 pthreads or windows native threads)
538 - \b ucontext: fast factory using System V contexts (or a portability
539 layer of our own on top of Windows fibers)
540 - \b raw: amazingly fast factory using a context switching mecanism
541 of our own, directly implemented in assembly (only available for x86
542 and amd64 platforms for now)
544 The only reason to change this setting is when the debugging tools get
545 fooled by the optimized context factories. Threads are the most
546 debugging-friendly contextes, as they allow to set breakpoints anywhere with gdb
547 and visualize backtraces for all processes, in order to debug concurrency issues.
548 Valgrind is also more comfortable with threads, but it should be usable with all factories.
550 \subsection options_virt_stacksize Adapting the used stack size
552 Each virtualized used process is executed using a specific system
553 stack. The size of this stack has a huge impact on the simulation
554 scalability, but its default value is rather large. This is because
555 the error messages that you get when the stack size is too small are
556 rather disturbing: this leads to stack overflow (overwriting other
557 stacks), leading to segfaults with corrupted stack traces.
559 If you want to push the scalability limits of your code, you might
560 want to reduce the \b contexts/stack_size item. Its default value
561 is 8192 (in KiB), while our Chord simulation works with stacks as small
562 as 16 KiB, for example. For the thread factory, the default value
563 is the one of the system, if it is too large/small, it has to be set
566 The operating system should only allocate memory for the pages of the
567 stack which are actually used and you might not need to use this in
568 most cases. However, this setting is very important when using the
569 model checker (see \ref options_mc_perf).
571 In some cases, no stack guard page is used and the stack will silently
572 overflow on other parts of the memory if the stack size is too small
573 for your application. This happens :
575 - on Windows systems;
576 - when the model checker is enabled;
577 - when stack guard pages are explicitely disabled (see \ref options_perf_guard_size).
579 \subsection options_virt_parallel Running user code in parallel
581 Parallel execution of the user code is only considered stable in
582 SimGrid v3.7 and higher. It is described in
583 <a href="http://hal.inria.fr/inria-00602216/">INRIA RR-7653</a>.
585 If you are using the \c ucontext or \c raw context factories, you can
586 request to execute the user code in parallel. Several threads are
587 launched, each of them handling as much user contexts at each run. To
588 actiave this, set the \b contexts/nthreads item to the amount of
589 cores that you have in your computer (or lower than 1 to have
590 the amount of cores auto-detected).
592 Even if you asked several worker threads using the previous option,
593 you can request to start the parallel execution (and pay the
594 associated synchronization costs) only if the potential parallelism is
595 large enough. For that, set the \b contexts/parallel_threshold
596 item to the minimal amount of user contexts needed to start the
597 parallel execution. In any given simulation round, if that amount is
598 not reached, the contexts will be run sequentially directly by the
599 main thread (thus saving the synchronization costs). Note that this
600 option is mainly useful when the grain of the user code is very fine,
601 because our synchronization is now very efficient.
603 When parallel execution is activated, you can choose the
604 synchronization schema used with the \b contexts/synchro item,
605 which value is either:
606 - \b futex: ultra optimized synchronisation schema, based on futexes
607 (fast user-mode mutexes), and thus only available on Linux systems.
608 This is the default mode when available.
609 - \b posix: slow but portable synchronisation using only POSIX
611 - \b busy_wait: not really a synchronisation: the worker threads
612 constantly request new contexts to execute. It should be the most
613 efficient synchronisation schema, but it loads all the cores of your
614 machine for no good reason. You probably prefer the other less
617 \section options_tracing Configuring the tracing subsystem
619 The \ref tracing "tracing subsystem" can be configured in several
620 different ways depending on the nature of the simulator (MSG, SimDag,
621 SMPI) and the kind of traces that need to be obtained. See the \ref
622 tracing_tracing_options "Tracing Configuration Options subsection" to
623 get a detailed description of each configuration option.
625 We detail here a simple way to get the traces working for you, even if
626 you never used the tracing API.
629 - Any SimGrid-based simulator (MSG, SimDag, SMPI, ...) and raw traces:
631 --cfg=tracing:yes --cfg=tracing/uncategorized:yes --cfg=triva/uncategorized:uncat.plist
633 The first parameter activates the tracing subsystem, the second
634 tells it to trace host and link utilization (without any
635 categorization) and the third creates a graph configuration file
636 to configure Triva when analysing the resulting trace file.
638 - MSG or SimDag-based simulator and categorized traces (you need to declare categories and classify your tasks according to them)
640 --cfg=tracing:yes --cfg=tracing/categorized:yes --cfg=triva/categorized:cat.plist
642 The first parameter activates the tracing subsystem, the second
643 tells it to trace host and link categorized utilization and the
644 third creates a graph configuration file to configure Triva when
645 analysing the resulting trace file.
647 - SMPI simulator and traces for a space/time view:
651 The <i>-trace</i> parameter for the smpirun script runs the
652 simulation with --cfg=tracing:yes and --cfg=tracing/smpi:yes. Check the
653 smpirun's <i>-help</i> parameter for additional tracing options.
655 Sometimes you might want to put additional information on the trace to
656 correctly identify them later, or to provide data that can be used to
657 reproduce an experiment. You have two ways to do that:
659 - Add a string on top of the trace file as comment:
661 --cfg=tracing/comment:my_simulation_identifier
664 - Add the contents of a textual file on top of the trace file as comment:
666 --cfg=tracing/comment_file:my_file_with_additional_information.txt
669 Please, use these two parameters (for comments) to make reproducible
670 simulations. For additional details about this and all tracing
671 options, check See the \ref tracing_tracing_options.
673 \section options_smpi Configuring SMPI
675 The SMPI interface provides several specific configuration items.
676 These are uneasy to see since the code is usually launched through the
677 \c smiprun script directly.
679 \subsection options_smpi_bench smpi/bench: Automatic benchmarking of SMPI code
681 In SMPI, the sequential code is automatically benchmarked, and these
682 computations are automatically reported to the simulator. That is to
683 say that if you have a large computation between a \c MPI_Recv() and a
684 \c MPI_Send(), SMPI will automatically benchmark the duration of this
685 code, and create an execution task within the simulator to take this
686 into account. For that, the actual duration is measured on the host
687 machine and then scaled to the power of the corresponding simulated
688 machine. The variable \b smpi/running_power allows to specify the
689 computational power of the host machine (in flop/s) to use when
690 scaling the execution times. It defaults to 20000, but you really want
691 to update it to get accurate simulation results.
693 When the code is constituted of numerous consecutive MPI calls, the
694 previous mechanism feeds the simulation kernel with numerous tiny
695 computations. The \b smpi/cpu_threshold item becomes handy when this
696 impacts badly the simulation performance. It specify a threshold (in
697 second) under which the execution chunks are not reported to the
698 simulation kernel (default value: 1e-6). Please note that in some
699 circonstances, this optimization can hinder the simulation accuracy.
701 In some cases, however, one may wish to disable simulation of
702 application computation. This is the case when SMPI is used not to
703 simulate an MPI applications, but instead an MPI code that performs
704 "live replay" of another MPI app (e.g., ScalaTrace's replay tool,
705 various on-line simulators that run an app at scale). In this case the
706 computation of the replay/simulation logic should not be simulated by
707 SMPI. Instead, the replay tool or on-line simulator will issue
708 "computation events", which correspond to the actual MPI simulation
709 being replayed/simulated. At the moment, these computation events can
710 be simulated using SMPI by calling internal smpi_execute*() functions.
712 To disable the benchmarking/simulation of computation in the simulated
713 application, the variable \b
714 smpi/simulation_computation should be set to no
716 \subsection options_model_smpi_bw_factor smpi/bw_factor: Bandwidth factors
718 The possible throughput of network links is often dependent on the
719 message sizes, as protocols may adapt to different message sizes. With
720 this option, a series of message sizes and factors are given, helping
721 the simulation to be more realistic. For instance, the current
725 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
728 So, messages with size 65472 and more will get a total of MAX_BANDWIDTH*0.940694,
729 messages of size 15424 to 65471 will get MAX_BANDWIDTH*0.697866 and so on.
730 Here, MAX_BANDWIDTH denotes the bandwidth of the link.
732 \subsection options_smpi_timing smpi/display_timing: Reporting simulation time
734 Most of the time, you run MPI code through SMPI to compute the time it
735 would take to run it on a platform that you don't have. But since the
736 code is run through the \c smpirun script, you don't have any control
737 on the launcher code, making difficult to report the simulated time
738 when the simulation ends. If you set the \b smpi/display_timing item
739 to 1, \c smpirun will display this information when the simulation ends. \verbatim
740 Simulation time: 1e3 seconds.
743 \subsection options_model_smpi_lat_factor smpi/lat_factor: Latency factors
745 The motivation and syntax for this option is identical to the motivation/syntax
746 of smpi/bw_factor, see \ref options_model_smpi_bw_factor for details.
748 There is an important difference, though: While smpi/bw_factor \a reduces the
749 actual bandwidth (i.e., values between 0 and 1 are valid), latency factors
750 increase the latency, i.e., values larger than or equal to 1 are valid here.
752 This is the default value:
755 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
758 \subsection options_smpi_global smpi/privatize_global_variables: Automatic privatization of global variables
760 MPI executables are meant to be executed in separated processes, but SMPI is
761 executed in only one process. Global variables from executables will be placed
762 in the same memory zone and shared between processes, causing hard to find bugs.
763 To avoid this, several options are possible :
764 - Manual edition of the code, for example to add __thread keyword before data
765 declaration, which allows the resulting code to work with SMPI, but only
766 if the thread factory (see \ref options_virt_factory) is used, as global
767 variables are then placed in the TLS (thread local storage) segment.
768 - Source-to-source transformation, to add a level of indirection
769 to the global variables. SMPI does this for F77 codes compiled with smpiff,
770 and used to provide coccinelle scripts for C codes, which are not functional anymore.
771 - Compilation pass, to have the compiler automatically put the data in
773 - Runtime automatic switching of the data segments. SMPI stores a copy of
774 each global data segment for each process, and at each context switch replaces
775 the actual data with its copy from the right process. This mechanism uses mmap,
776 and is for now limited to systems supporting this functionnality (all Linux
777 and some BSD should be compatible).
778 Another limitation is that SMPI only accounts for global variables defined in
779 the executable. If the processes use external global variables from dynamic
780 libraries, they won't be switched correctly. To avoid this, using static
781 linking is advised (but not with the simgrid library, to avoid replicating
782 its own global variables).
784 To use this runtime automatic switching, the variable \b smpi/privatize_global_variables
789 \subsection options_model_smpi_detached Simulating MPI detached send
791 This threshold specifies the size in bytes under which the send will return
792 immediately. This is different from the threshold detailed in \ref options_model_network_asyncsend
793 because the message is not effectively sent when the send is posted. SMPI still waits for the
794 correspondant receive to be posted to perform the communication operation. This threshold can be set
795 by changing the \b smpi/send_is_detached item. The default value is 65536.
797 \subsection options_model_smpi_collectives Simulating MPI collective algorithms
799 SMPI implements more than 100 different algorithms for MPI collective communication, to accurately
800 simulate the behavior of most of the existing MPI libraries. The \b smpi/coll_selector item can be used
801 to use the decision logic of either OpenMPI or MPICH libraries (values: ompi or mpich, by default SMPI
802 uses naive version of collective operations). Each collective operation can be manually selected with a
803 \b smpi/collective_name:algo_name. Available algorithms are listed in \ref SMPI_collective_algorithms .
805 \subsection options_model_smpi_iprobe smpi/iprobe: Inject constant times for calls to MPI_Iprobe
807 \b Default value: 0.0001
809 The behavior and motivation for this configuration option is identical with \a smpi/test, see
810 Section \ref options_model_smpi_test for details.
812 \subsection options_model_smpi_test smpi/test: Inject constant times for calls to MPI_Test
814 \b Default value: 0.0001
816 By setting this option, you can control the amount of time a process sleeps
817 when MPI_Test() is called; this is important, because SimGrid normally only
818 advances the time while communication is happening and thus,
819 MPI_Test will not add to the time, resulting in a deadlock if used as a
826 MPI_Test(request, flag, status);
832 Internally, in order to speed up execution, we use a counter to keep track
833 on how often we already checked if the handle is now valid or not. Hence, we
834 actually use counter*SLEEP_TIME, that is, the time MPI_Test() causes the process
835 to sleep increases linearly with the number of previously failed testk.
838 \subsection options_model_smpi_wtime smpi/wtime: Inject constant times for calls to MPI_Wtime
842 By setting this option, you can control the amount of time a process sleeps
843 when MPI_Wtime() is called; this is important, because SimGrid normally only
844 advances the time while communication is happening and thus,
845 MPI_Wtime will not add to the time, resulting in a deadlock if used as a
851 while(MPI_Wtime() < some_time_bound) {
856 If the time is never advanced, this loop will clearly never end as MPI_Wtime()
857 always returns the same value. Hence, pass a (small) value to the smpi/wtime
858 option to force a call to MPI_Wtime to advance the time as well.
861 \section options_generic Configuring other aspects of SimGrid
863 \subsection options_generic_clean_atexit Cleanup before termination
865 The C / C++ standard contains a function called \b [atexit](http://www.cplusplus.com/reference/cstdlib/atexit/).
866 atexit registers callbacks, which are called just before the program terminates.
868 By setting the configuration option clean_atexit to 1 (true), a callback
869 is registered and will clean up some variables and terminate/cleanup the tracing.
871 TODO: Add when this should be used.
873 \subsection options_generic_path XML file inclusion path
875 It is possible to specify a list of directories to search into for the
876 \<include\> tag in XML files by using the \b path configuration
877 item. To add several directory to the path, set the configuration
878 item several times, as in \verbatim
879 --cfg=path:toto --cfg=path:tutu
882 \subsection options_generic_exit Behavior on Ctrl-C
884 By default, when Ctrl-C is pressed, the status of all existing
885 simulated processes is displayed before exiting the simulation. This is very useful to debug your
886 code, but it can reveal troublesome in some cases (such as when the
887 amount of processes becomes really big). This behavior is disabled
888 when \b verbose-exit is set to 0 (it is to 1 by default).
890 \subsection options_exception_cutpath Truncate local path from exception backtrace
892 <b>This configuration option is an internal option and should normally not be used
893 by the user.</b> It is used to remove the path from the backtrace
894 shown when an exception is thrown; if we didn't remove this part, the tests
895 testing the exception parts of simgrid would fail on most machines, as we are
896 currently comparing output. Clearly, the path used on different machines are almost
897 guaranteed to be different and hence, the output would
898 mismatch, causing the test to fail.
900 \section options_log Logging Configuration
902 It can be done by using XBT. Go to \ref XBT_log for more details.
904 \section options_perf Performance optimizations
906 \subsection options_perf_context Context factory
908 In order to achieve higher performance, you might want to use the raw
909 context factory which avoids any system call when switching between
910 tasks. If it is not possible you might use ucontext instead.
912 \subsection options_perf_guard_size Disabling stack guard pages
914 A stack guard page is usually used which prevents the stack from
915 overflowing on other parts of the memory. However this might have a
916 performance impact if a huge number of processes is created. The
917 option \b contexts:guard_size is the number of stack guard pages
918 used. By setting it to 0, no guard pages will be used: in this case,
919 you should avoid using small stacks (\b stack_size) as the stack will
920 silently overflow on other parts of the memory.
922 \section options_index Index of all existing configuration options
925 Almost all options are defined in <i>src/simgrid/sg_config.c</i>. You may
926 want to check this file, too, but this index should be somewhat complete
927 for the moment (May 2015).
930 \b Please \b note: You can also pass the command-line option "\b --help" and
931 "--help-cfg" to an executable that uses simgrid.
933 - \c clean_atexit: \ref options_generic_clean_atexit
935 - \c contexts/factory: \ref options_virt_factory
936 - \c contexts/guard_size: \ref options_virt_parallel
937 - \c contexts/nthreads: \ref options_virt_parallel
938 - \c contexts/parallel_threshold: \ref options_virt_parallel
939 - \c contexts/stack_size: \ref options_virt_stacksize
940 - \c contexts/synchro: \ref options_virt_parallel
942 - \c cpu/maxmin_selective_update: \ref options_model_optim
943 - \c cpu/model: \ref options_model_select
944 - \c cpu/optim: \ref options_model_optim
946 - \c exception/cutpath: \ref options_exception_cutpath
948 - \c gtnets/jitter: \ref options_pls
949 - \c gtnets/jitter_seed: \ref options_pls
951 - \c maxmin/precision: \ref options_model_precision
953 - \c msg/debug_multiple_use: \ref options_msg_debug_multiple_use
955 - \c model-check: \ref options_modelchecking
956 - \c model-check/checkpoint: \ref options_modelchecking_steps
957 - \c model-check/communications_determinism: \ref options_modelchecking_comm_determinism
958 - \c model-check/communications_determinism: \ref options_modelchecking_send_determinism
959 - \c model-check/dot_output: \ref options_modelchecking_dot_output
960 - \c model-check/hash: \ref options_modelchecking_hash
961 - \c model-check/property: \ref options_modelchecking_liveness
962 - \c model-check/max_depth: \ref options_modelchecking_max_depth
963 - \c model-check/record: \ref options_modelchecking_recordreplay
964 - \c model-check/reduction: \ref options_modelchecking_reduction
965 - \c model-check/replay: \ref options_modelchecking_recordreplay
966 - \c model-check/send_determinism: \ref options_modelchecking_sparse_checkpoint
967 - \c model-check/sparse-checkpoint: \ref options_modelchecking_sparse_checkpoint
968 - \c model-check/termination: \ref options_modelchecking_termination
969 - \c model-check/timeout: \ref options_modelchecking_timeout
970 - \c model-check/visited: \ref options_modelchecking_visited
972 - \c network/bandwidth_factor: \ref options_model_network_coefs
973 - \c network/coordinates: \ref options_model_network_coord
974 - \c network/crosstraffic: \ref options_model_network_crosstraffic
975 - \c network/latency_factor: \ref options_model_network_coefs
976 - \c network/maxmin_selective_update: \ref options_model_optim
977 - \c network/model: \ref options_model_select
978 - \c network/optim: \ref options_model_optim
979 - \c network/sender_gap: \ref options_model_network_sendergap
980 - \c network/TCP_gamma: \ref options_model_network_gamma
981 - \c network/weight_S: \ref options_model_network_coefs
983 - \c ns3/TcpModel: \ref options_pls
985 - \c surf/nthreads: \ref options_model_nthreads
986 - \c surf/precision: \ref options_model_precision
988 - \c <b>For collective operations of SMPI, please refer to Section \ref options_index_smpi_coll</b>
989 - \c smpi/async_small_thres: \ref options_model_network_asyncsend
990 - \c smpi/bw_factor: \ref options_model_smpi_bw_factor
991 - \c smpi/coll_selector: \ref options_model_smpi_collectives
992 - \c smpi/cpu_threshold: \ref options_smpi_bench
993 - \c smpi/display_timing: \ref options_smpi_timing
994 - \c smpi/lat_factor: \ref options_model_smpi_lat_factor
995 - \c smpi/IB_penalty_factors: \ref options_model_network_coefs
996 - \c smpi/iprobe: \ref options_model_smpi_iprobe
997 - \c smpi/ois: \ref options_model_smpi_ois
998 - \c smpi/or: \ref options_model_smpi_or
999 - \c smpi/os: \ref options_model_smpi_os
1000 - \c smpi/privatize_global_variables: \ref options_smpi_global
1001 - \c smpi/running_power: \ref options_smpi_bench
1002 - \c smpi/send_is_detached_thresh: \ref options_model_smpi_detached
1003 - \c smpi/simulation_computation: \ref options_smpi_bench
1004 - \c smpi/test: \ref options_model_smpi_test
1005 - \c smpi/use_shared_malloc: \ref options_model_smpi_use_shared_malloc
1006 - \c smpi/wtime: \ref options_model_smpi_wtime
1008 - \c <b>Tracing configuration options can be found in Section \ref tracing_tracing_options</b>.
1010 - \c storage/model: \ref options_storage_model
1011 - \c path: \ref options_generic_path
1012 - \c plugin: \ref options_generic_plugin
1013 - \c verbose-exit: \ref options_generic_exit
1015 - \c vm_workstation/model: \ref options_vm_workstation_model
1016 - \c workstation/model: \ref options_model_select
1018 \subsection options_index_smpi_coll Index of SMPI collective algorithms options
1019 - \c smpi/allgather: \ref options_model_smpi_coll_allgather
1020 - \c smpi/allgatherv: \ref options_model_smpi_coll_allgatherv
1021 - \c smpi/allreduce: \ref options_model_smpi_coll_allreduce
1022 - \c smpi/alltoall: \ref options_model_smpi_coll_alltoall
1023 - \c smpi/alltoallv: \ref options_model_smpi_coll_alltoallv
1024 - \c smpi/barrier: \ref options_model_smpi_coll_barrier
1025 - \c smpi/bcast: \ref options_model_smpi_coll_bcast
1026 - \c smpi/gather: \ref options_model_smpi_coll_gather
1027 - \c smpi/reduce: \ref options_model_smpi_coll_reduce
1028 - \c smpi/reduce_scatter: \ref options_model_smpi_coll_reduce_scatter
1029 - \c smpi/scatter: \ref options_model_smpi_coll_scatter