8 <object id="TOC" data="graphical-toc.svg" width="100%" type="image/svg+xml"></object>
10 window.onload=function() { // Wait for the SVG to be loaded before changing it
11 var elem=document.querySelector("#TOC").contentDocument.getElementById("ConfigBox")
12 elem.style="opacity:0.93999999;fill:#ff0000;fill-opacity:0.1;stroke:#000000;stroke-width:0.35277778;stroke-linecap:round;stroke-linejoin:round;stroke-miterlimit:4;stroke-dasharray:none;stroke-dashoffset:0;stroke-opacity:1";
18 A number of options can be given at runtime to change the default
19 SimGrid behavior. For a complete list of all configuration options
20 accepted by the SimGrid version used in your simulator, simply pass
21 the --help configuration flag to your program. If some of the options
22 are not documented on this page, this is a bug that you should please
23 report so that we can fix it. Note that some of the options presented
24 here may not be available in your simulators, depending on the
25 :ref:`compile-time options <install_src_config>` that you used.
27 Setting Configuration Items
28 ---------------------------
30 There is several way to pass configuration options to the simulators.
31 The most common way is to use the ``--cfg`` command line argument. For
32 example, to set the item ``Item`` to the value ``Value``, simply
33 type the following on the command-line:
37 my_simulator --cfg=Item:Value (other arguments)
39 Several ``--cfg`` command line arguments can naturally be used. If you
40 need to include spaces in the argument, don't forget to quote the
41 argument. You can even escape the included quotes (write @' for ' if
42 you have your argument between ').
44 Another solution is to use the ``<config>`` tag in the platform file. The
45 only restriction is that this tag must occur before the first
46 platform element (be it ``<zone>``, ``<cluster>``, ``<peer>`` or whatever).
47 The ``<config>`` tag takes an ``id`` attribute, but it is currently
48 ignored so you don't really need to pass it. The important part is that
49 within that tag, you can pass one or several ``<prop>`` tags to specify
50 the configuration to use. For example, setting ``Item`` to ``Value``
51 can be done by adding the following to the beginning of your platform
57 <prop id="Item" value="Value"/>
60 A last solution is to pass your configuration directly in your program
61 with :cpp:func:`simgrid::s4u::Engine::set_config` or :cpp:func:`MSG_config`.
65 #include <simgrid/s4u.hpp>
67 int main(int argc, char *argv[]) {
68 simgrid::s4u::Engine e(&argc, argv);
70 e->set_config("Item:Value");
77 Existing Configuration Items
78 ----------------------------
81 The full list can be retrieved by passing ``--help`` and
82 ``--help-cfg`` to an executable that uses SimGrid. Try passing
83 ``help`` as a value to get the list of values accepted by a given
84 option. For example, ``--cfg=plugin:help`` will give you the list
85 of plugins available in your installation of SimGrid.
87 - **contexts/factory:** :ref:`cfg=contexts/factory`
88 - **contexts/guard-size:** :ref:`cfg=contexts/guard-size`
89 - **contexts/nthreads:** :ref:`cfg=contexts/nthreads`
90 - **contexts/parallel-threshold:** :ref:`cfg=contexts/parallel-threshold`
91 - **contexts/stack-size:** :ref:`cfg=contexts/stack-size`
92 - **contexts/synchro:** :ref:`cfg=contexts/synchro`
94 - **cpu/maxmin-selective-update:** :ref:`Cpu Optimization Level <options_model_optim>`
95 - **cpu/model:** :ref:`options_model_select`
96 - **cpu/optim:** :ref:`Cpu Optimization Level <options_model_optim>`
98 - **debug/breakpoint:** :ref:`cfg=debug/breakpoint`
99 - **debug/clean-atexit:** :ref:`cfg=debug/clean-atexit`
100 - **debug/verbose-exit:** :ref:`cfg=debug/verbose-exit`
102 - **exception/cutpath:** :ref:`cfg=exception/cutpath`
104 - **host/model:** :ref:`options_model_select`
106 - **maxmin/precision:** :ref:`cfg=maxmin/precision`
107 - **maxmin/concurrency-limit:** :ref:`cfg=maxmin/concurrency-limit`
109 - **msg/debug-multiple-use:** :ref:`cfg=msg/debug-multiple-use`
111 - **model-check:** :ref:`options_modelchecking`
112 - **model-check/checkpoint:** :ref:`cfg=model-check/checkpoint`
113 - **model-check/communications-determinism:** :ref:`cfg=model-check/communications-determinism`
114 - **model-check/dot-output:** :ref:`cfg=model-check/dot-output`
115 - **model-check/max-depth:** :ref:`cfg=model-check/max-depth`
116 - **model-check/property:** :ref:`cfg=model-check/property`
117 - **model-check/reduction:** :ref:`cfg=model-check/reduction`
118 - **model-check/replay:** :ref:`cfg=model-check/replay`
119 - **model-check/send-determinism:** :ref:`cfg=model-check/send-determinism`
120 - **model-check/termination:** :ref:`cfg=model-check/termination`
121 - **model-check/timeout:** :ref:`cfg=model-check/timeout`
122 - **model-check/visited:** :ref:`cfg=model-check/visited`
124 - **network/bandwidth-factor:** :ref:`cfg=network/bandwidth-factor`
125 - **network/crosstraffic:** :ref:`cfg=network/crosstraffic`
126 - **network/latency-factor:** :ref:`cfg=network/latency-factor`
127 - **network/maxmin-selective-update:** :ref:`Network Optimization Level <options_model_optim>`
128 - **network/model:** :ref:`options_model_select`
129 - **network/optim:** :ref:`Network Optimization Level <options_model_optim>`
130 - **network/TCP-gamma:** :ref:`cfg=network/TCP-gamma`
131 - **network/weight-S:** :ref:`cfg=network/weight-S`
133 - **ns3/TcpModel:** :ref:`options_pls`
134 - **path:** :ref:`cfg=path`
135 - **plugin:** :ref:`cfg=plugin`
137 - **storage/max_file_descriptors:** :ref:`cfg=storage/max_file_descriptors`
139 - **surf/precision:** :ref:`cfg=surf/precision`
141 - **For collective operations of SMPI,** please refer to Section :ref:`cfg=smpi/coll-selector`
142 - **smpi/async-small-thresh:** :ref:`cfg=smpi/async-small-thresh`
143 - **smpi/buffering:** :ref:`cfg=smpi/buffering`
144 - **smpi/bw-factor:** :ref:`cfg=smpi/bw-factor`
145 - **smpi/coll-selector:** :ref:`cfg=smpi/coll-selector`
146 - **smpi/comp-adjustment-file:** :ref:`cfg=smpi/comp-adjustment-file`
147 - **smpi/cpu-threshold:** :ref:`cfg=smpi/cpu-threshold`
148 - **smpi/display-timing:** :ref:`cfg=smpi/display-timing`
149 - **smpi/grow-injected-times:** :ref:`cfg=smpi/grow-injected-times`
150 - **smpi/host-speed:** :ref:`cfg=smpi/host-speed`
151 - **smpi/IB-penalty-factors:** :ref:`cfg=smpi/IB-penalty-factors`
152 - **smpi/iprobe:** :ref:`cfg=smpi/iprobe`
153 - **smpi/iprobe-cpu-usage:** :ref:`cfg=smpi/iprobe-cpu-usage`
154 - **smpi/init:** :ref:`cfg=smpi/init`
155 - **smpi/keep-temps:** :ref:`cfg=smpi/keep-temps`
156 - **smpi/lat-factor:** :ref:`cfg=smpi/lat-factor`
157 - **smpi/ois:** :ref:`cfg=smpi/ois`
158 - **smpi/or:** :ref:`cfg=smpi/or`
159 - **smpi/os:** :ref:`cfg=smpi/os`
160 - **smpi/papi-events:** :ref:`cfg=smpi/papi-events`
161 - **smpi/privatization:** :ref:`cfg=smpi/privatization`
162 - **smpi/privatize-libs:** :ref:`cfg=smpi/privatize-libs`
163 - **smpi/send-is-detached-thresh:** :ref:`cfg=smpi/send-is-detached-thresh`
164 - **smpi/shared-malloc:** :ref:`cfg=smpi/shared-malloc`
165 - **smpi/shared-malloc-hugepage:** :ref:`cfg=smpi/shared-malloc-hugepage`
166 - **smpi/simulate-computation:** :ref:`cfg=smpi/simulate-computation`
167 - **smpi/test:** :ref:`cfg=smpi/test`
168 - **smpi/wtime:** :ref:`cfg=smpi/wtime`
170 - **Tracing configuration options** can be found in Section :ref:`tracing_tracing_options`
172 - **storage/model:** :ref:`options_model_select`
174 - **vm/model:** :ref:`options_model_select`
178 Configuring the Platform Models
179 -------------------------------
181 .. _options_model_select:
183 Choosing the Platform Models
184 ............................
186 SimGrid comes with several network, CPU and storage models built in,
187 and you can change the used model at runtime by changing the passed
188 configuration. The three main configuration items are given below.
189 For each of these items, passing the special ``help`` value gives you
190 a short description of all possible values (for example,
191 ``--cfg=network/model:help`` will present all provided network
192 models). Also, ``--help-models`` should provide information about all
193 models for all existing resources.
195 - ``network/model``: specify the used network model. Possible values:
197 - **LV08 (default one):** Realistic network analytic model
198 (slow-start modeled by multiplying latency by 13.01, bandwidth by
199 .97; bottleneck sharing uses a payload of S=20537 for evaluating
200 RTT). Described in `Accuracy Study and Improvement of Network
201 Simulation in the SimGrid Framework
202 <http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf>`_.
203 - **Constant:** Simplistic network model where all communication
204 take a constant time (one second). This model provides the lowest
205 realism, but is (marginally) faster.
206 - **SMPI:** Realistic network model specifically tailored for HPC
207 settings (accurate modeling of slow start with correction factors on
208 three intervals: < 1KiB, < 64 KiB, >= 64 KiB). This model can be
209 :ref:`further configured <options_model_network>`.
210 - **IB:** Realistic network model specifically tailored for HPC
211 settings with InfiniBand networks (accurate modeling contention
212 behavior, based on the model explained in `this PhD work
213 <http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf>`_.
214 This model can be :ref:`further configured <options_model_network>`.
215 - **CM02:** Legacy network analytic model. Very similar to LV08, but
216 without corrective factors. The timings of small messages are thus
217 poorly modeled. This model is described in `A Network Model for
218 Simulation of Grid Application
219 <https://hal.inria.fr/inria-00071989/document>`_.
220 - **ns-3** (only available if you compiled SimGrid accordingly):
221 Use the packet-level network
222 simulators as network models (see :ref:`model_ns3`).
223 This model can be :ref:`further configured <options_pls>`.
225 - ``cpu/model``: specify the used CPU model. We have only one model
228 - **Cas01:** Simplistic CPU model (time=size/power)
230 - ``host/model``: The host concept is the aggregation of a CPU with a
231 network card. Three models exists, but actually, only 2 of them are
232 interesting. The "compound" one is simply due to the way our
233 internal code is organized, and can easily be ignored. So at the
234 end, you have two host models: The default one allows aggregation of
235 an existing CPU model with an existing network model, but does not
236 allow parallel tasks because these beasts need some collaboration
237 between the network and CPU model. That is why, ptask_07 is used by
238 default when using SimDag.
240 - **default:** Default host model. Currently, CPU:Cas01 and
241 network:LV08 (with cross traffic enabled)
242 - **compound:** Host model that is automatically chosen if
243 you change the network and CPU models
244 - **ptask_L07:** Host model somehow similar to Cas01+CM02 but
245 allowing "parallel tasks", that are intended to model the moldable
246 tasks of the grid scheduling literature.
248 - ``storage/model``: specify the used storage model. Only one model is
250 - ``vm/model``: specify the model for virtual machines. Only one model
253 .. todo: make 'compound' the default host model.
255 .. _options_model_optim:
260 The network and CPU models that are based on lmm_solve (that
261 is, all our analytical models) accept specific optimization
264 - items ``network/optim`` and ``cpu/optim`` (both default to 'Lazy'):
266 - **Lazy:** Lazy action management (partial invalidation in lmm +
267 heap in action remaining).
268 - **TI:** Trace integration. Highly optimized mode when using
269 availability traces (only available for the Cas01 CPU model for
271 - **Full:** Full update of remaining and variables. Slow but may be
272 useful when debugging.
274 - items ``network/maxmin-selective-update`` and
275 ``cpu/maxmin-selective-update``: configure whether the underlying
276 should be lazily updated or not. It should have no impact on the
277 computed timings, but should speed up the computation. |br| It is
278 still possible to disable this feature because it can reveal
279 counter-productive in very specific scenarios where the
280 interaction level is high. In particular, if all your
281 communication share a given backbone link, you should disable it:
282 without it, a simple regular loop is used to update each
283 communication. With it, each of them is still updated (because of
284 the dependency induced by the backbone), but through a complicated
285 and slow pattern that follows the actual dependencies.
287 .. _cfg=maxmin/precision:
288 .. _cfg=surf/precision:
293 **Option** ``maxmin/precision`` **Default:** 0.00001 (in flops or bytes) |br|
294 **Option** ``surf/precision`` **Default:** 0.00001 (in seconds)
296 The analytical models handle a lot of floating point values. It is
297 possible to change the epsilon used to update and compare them through
298 this configuration item. Changing it may speedup the simulation by
299 discarding very small actions, at the price of a reduced numerical
300 precision. You can modify separately the precision used to manipulate
301 timings (in seconds) and the one used to manipulate amounts of work
304 .. _cfg=maxmin/concurrency-limit:
309 **Option** ``maxmin/concurrency-limit`` **Default:** -1 (no limit)
311 The maximum number of variables per resource can be tuned through this
312 option. You can have as many simultaneous actions per resources as you
313 want. If your simulation presents a very high level of concurrency, it
314 may help to use e.g. 100 as a value here. It means that at most 100
315 actions can consume a resource at a given time. The extraneous actions
316 are queued and wait until the amount of concurrency of the considered
317 resource lowers under the given boundary.
319 Such limitations help both to the simulation speed and simulation accuracy
320 on highly constrained scenarios, but the simulation speed suffers of this
321 setting on regular (less constrained) scenarios so it is off by default.
323 .. _options_model_network:
325 Configuring the Network Model
326 .............................
328 .. _cfg=network/TCP-gamma:
330 Maximal TCP Window Size
331 ^^^^^^^^^^^^^^^^^^^^^^^
333 **Option** ``network/TCP-gamma`` **Default:** 4194304
335 The analytical models need to know the maximal TCP window size to take
336 the TCP congestion mechanism into account. On Linux, this value can
337 be retrieved using the following commands. Both give a set of values,
338 and you should use the last one, which is the maximal size.
340 .. code-block:: shell
342 cat /proc/sys/net/ipv4/tcp_rmem # gives the sender window
343 cat /proc/sys/net/ipv4/tcp_wmem # gives the receiver window
345 .. _cfg=smpi/IB-penalty-factors:
346 .. _cfg=network/bandwidth-factor:
347 .. _cfg=network/latency-factor:
348 .. _cfg=network/weight-S:
350 Correcting Important Network Parameters
351 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
353 SimGrid can take network irregularities such as a slow startup or
354 changing behavior depending on the message size into account. You
355 should not change these values unless you really know what you're
356 doing. The corresponding values were computed through data fitting
357 one the timings of packet-level simulators, as described in `Accuracy
358 Study and Improvement of Network Simulation in the SimGrid Framework
359 <http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf>`_.
362 If you are using the SMPI model, these correction coefficients are
363 themselves corrected by constant values depending on the size of the
364 exchange. By default SMPI uses factors computed on the Stampede
365 Supercomputer at TACC, with optimal deployment of processes on
366 nodes. Again, only hardcore experts should bother about this fact.
368 InfiniBand network behavior can be modeled through 3 parameters
369 ``smpi/IB-penalty-factors:"βe;βs;γs"``, as explained in `this PhD
371 <http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf>`_.
373 .. todo:: This section should be rewritten, and actually explain the
374 options network/bandwidth-factor, network/latency-factor,
377 .. _cfg=network/crosstraffic:
379 Simulating Cross-Traffic
380 ^^^^^^^^^^^^^^^^^^^^^^^^
382 Since SimGrid v3.7, cross-traffic effects can be taken into account in
383 analytical simulations. It means that ongoing and incoming
384 communication flows are treated independently. In addition, the LV08
385 model adds 0.05 of usage on the opposite direction for each new
386 created flow. This can be useful to simulate some important TCP
387 phenomena such as ack compression.
389 For that to work, your platform must have two links for each
390 pair of interconnected hosts. An example of usable platform is
391 available in ``examples/platforms/crosstraffic.xml``.
393 This is activated through the ``network/crosstraffic`` item, that
394 can be set to 0 (disable this feature) or 1 (enable it).
396 Note that with the default host model this option is activated by default.
398 .. _cfg=smpi/async-small-thresh:
400 Simulating Asynchronous Send
401 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
403 (this configuration item is experimental and may change or disappear)
405 It is possible to specify that messages below a certain size will be
406 sent as soon as the call to MPI_Send is issued, without waiting for
407 the correspondant receive. This threshold can be configured through
408 the ``smpi/async-small-thresh`` item. The default value is 0. This
409 behavior can also be manually set for mailboxes, by setting the
410 receiving mode of the mailbox with a call to
411 :cpp:func:`MSG_mailbox_set_async`. After this, all messages sent to
412 this mailbox will have this behavior regardless of the message size.
414 This value needs to be smaller than or equals to the threshold set at
415 :ref:`cfg=smpi/send-is-detached-thresh`, because asynchronous messages
416 are meant to be detached as well.
423 **Option** ``ns3/TcpModel`` **Default:** "default" (ns-3 default)
425 When using ns-3, there is an extra item ``ns3/TcpModel``, corresponding
426 to the ``ns3::TcpL4Protocol::SocketType`` configuration item in
427 ns-3. The only valid values (enforced on the SimGrid side) are
428 'default' (no change to the ns-3 configuration), 'NewReno' or 'Reno' or
431 Configuring the Storage model
432 .............................
434 .. _cfg=storage/max_file_descriptors:
436 File Descriptor Cound per Host
437 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
439 **Option** ``storage/max_file_descriptors`` **Default:** 1024
441 Each host maintains a fixed-size array of its file descriptors. You
442 can change its size through this item to either enlarge it if your
443 application requires it or to reduce it to save memory space.
450 SimGrid plugins allow one to extend the framework without changing its
451 source code directly. Read the source code of the existing plugins to
452 learn how to do so (in ``src/plugins``), and ask your questions to the
453 usual channels (Stack Overflow, Mailing list, IRC). The basic idea is
454 that plugins usually register callbacks to some signals of interest.
455 If they need to store some information about a given object (Link, CPU
456 or Actor), they do so through the use of a dedicated object extension.
458 Some of the existing plugins can be activated from the command line,
459 meaning that you can activate them from the command line without any
460 modification to your simulation code. For example, you can activate
461 the host energy plugin by adding ``--cfg=plugin:host_energy`` to your
464 Here is the full list of plugins that can be activated this way:
466 - **host_energy:** keeps track of the energy dissipated by
467 computations. More details in @ref plugin_energy.
468 - **link_energy:** keeps track of the energy dissipated by
469 communications. More details in @ref SURF_plugin_energy.
470 - **host_load:** keeps track of the computational load.
471 More details in @ref plugin_load.
473 .. _options_modelchecking:
475 Configuring the Model-Checking
476 ------------------------------
478 To enable the SimGrid model-checking support the program should
479 be executed using the simgrid-mc wrapper:
481 .. code-block:: shell
483 simgrid-mc ./my_program
485 Safety properties are expressed as assertions using the function
486 :cpp:func:`void MC_assert(int prop)`.
488 .. _cfg=smpi/buffering:
490 Specifying the MPI buffering behavior
491 .....................................
493 **Option** ``smpi/buffering`` **Default:** infty
495 Buffering in MPI has a huge impact on the communication semantic. For example,
496 standard blocking sends are synchronous calls when the system buffers are full
497 while these calls can complete immediately without even requiring a matching
498 receive call for small messages sent when the system buffers are empty.
500 In SMPI, this depends on the message size, that is compared against two thresholds:
502 - if (size < :ref:`smpi/async-small-thresh <cfg=smpi/async-small-thresh>`) then
503 MPI_Send returns immediately, even if the corresponding receive has not be issued yet.
504 - if (:ref:`smpi/async-small-thresh <cfg=smpi/async-small-thresh>` < size < :ref:`smpi/send-is-detached-thresh <cfg=smpi/send-is-detached-thresh>`) then
505 MPI_Send returns as soon as the corresponding receive has been issued. This is known as the eager mode.
506 - if (:ref:`smpi/send-is-detached-thresh <cfg=smpi/send-is-detached-thresh>` < size) then
507 MPI_Send returns only when the message has actually been sent over the network. This is known as the rendez-vous mode.
509 The ``smpi/buffering`` option gives an easier interface to choose between these semantics. It can take two values:
511 - **zero:** means that buffering should be disabled. Blocking communications are actually blocking.
512 - **infty:** means that buffering should be made infinite. Blocking communications are non-blocking.
514 .. _cfg=model-check/property:
516 Specifying a liveness property
517 ..............................
519 **Option** ``model-check/property`` **Default:** unset
521 If you want to specify liveness properties, you have to pass them on
522 the command line, specifying the name of the file containing the
523 property, as formatted by the ltl2ba program. Note that ltl2ba is not
524 part of SimGrid and must be installed separatly.
526 .. code-block:: shell
528 simgrid-mc ./my_program --cfg=model-check/property:<filename>
530 .. _cfg=model-check/checkpoint:
532 Going for Stateful Verification
533 ...............................
535 By default, the system is backtracked to its initial state to explore
536 another path instead of backtracking to the exact step before the fork
537 that we want to explore (this is called stateless verification). This
538 is done this way because saving intermediate states can rapidly
539 exhaust the available memory. If you want, you can change the value of
540 the ``model-check/checkpoint`` item. For example,
541 ``--cfg=model-check/checkpoint:1`` asks to take a checkpoint every
542 step. Beware, this will certainly explode your memory. Larger values
543 are probably better, make sure to experiment a bit to find the right
544 setting for your specific system.
546 .. _cfg=model-check/reduction:
548 Specifying the kind of reduction
549 ................................
551 The main issue when using the model-checking is the state space
552 explosion. To counter that problem, you can chose a exploration
553 reduction techniques with
554 ``--cfg=model-check/reduction:<technique>``. For now, this
555 configuration variable can take 2 values:
557 - **none:** Do not apply any kind of reduction (mandatory for now for
559 - **dpor:** Apply Dynamic Partial Ordering Reduction. Only valid if
560 you verify local safety properties (default value for safety
563 There is unfortunately no silver bullet here, and the most efficient
564 reduction techniques cannot be applied to any properties. In
565 particular, the DPOR method cannot be applied on liveness properties
566 since our implementation of DPOR may break some cycles, while cycles
567 are very important to the soundness of the exploration for liveness
570 .. _cfg=model-check/visited:
572 Size of Cycle Detection Set
573 ...........................
575 In order to detect cycles, the model checker needs to check if a new
576 explored state is in fact the same state than a previous one. For
577 that, the model checker can take a snapshot of each visited state:
578 this snapshot is then used to compare it with subsequent states in the
581 The ``model-check/visited`` item is the maximum number of states which
582 are stored in memory. If the maximum number of snapshotted state is
583 reached, some states will be removed from the memory and some cycles
584 might be missed. Small values can lead to incorrect verifications, but
585 large value can exhaust your memory, so choose carefully.
587 By default, no state is snapshotted and cycles cannot be detected.
589 .. _cfg=model-check/termination:
591 Non-Termination Detection
592 .........................
594 The ``model-check/termination`` configuration item can be used to
595 report if a non-termination execution path has been found. This is a
596 path with a cycle which means that the program might never terminate.
598 This only works in safety mode, not in liveness mode.
600 This options is disabled by default.
602 .. _cfg=model-check/dot-output:
607 If set, the ``model-check/dot-output`` configuration item is the name
608 of a file in which to write a dot file of the path leading the found
609 property (safety or liveness violation) as well as the cycle for
610 liveness properties. This dot file can then fed to the graphviz dot
611 tool to generate an corresponding graphical representation.
613 .. _cfg=model-check/max-depth:
615 Exploration Depth Limit
616 .......................
618 The ``model-checker/max-depth`` can set the maximum depth of the
619 exploration graph of the model checker. If this limit is reached, a
620 logging message is sent and the results might not be exact.
622 By default, there is not depth limit.
624 .. _cfg=model-check/timeout:
629 By default, the model checker does not handle timeout conditions: the `wait`
630 operations never time out. With the ``model-check/timeout`` configuration item
631 set to **yes**, the model checker will explore timeouts of `wait` operations.
633 .. _cfg=model-check/communications-determinism:
634 .. _cfg=model-check/send-determinism:
636 Communication Determinism
637 .........................
639 The ``model-check/communications-determinism`` and
640 ``model-check/send-determinism`` items can be used to select the
641 communication determinism mode of the model checker which checks
642 determinism properties of the communications of an application.
644 Verification Performance Considerations
645 .......................................
647 The size of the stacks can have a huge impact on the memory
648 consumption when using model-checking. By default, each snapshot will
649 save a copy of the whole stacks and not only of the part which is
650 really meaningful: you should expect the contribution of the memory
651 consumption of the snapshots to be @f$ @mbox{number of processes}
652 @times @mbox{stack size} @times @mbox{number of states} @f$.
654 When compiled against the model checker, the stacks are not
655 protected with guards: if the stack size is too small for your
656 application, the stack will silently overflow on other parts of the
657 memory (see :ref:`contexts/guard-size <cfg=contexts/guard-size>`).
659 .. _cfg=model-check/replay:
661 Replaying buggy execution paths from the model checker
662 ......................................................
664 Debugging the problems reported by the model checker is challenging: First, the
665 application under verification cannot be debugged with gdb because the
666 model checker already traces it. Then, the model checker may explore several
667 execution paths before encountering the issue, making it very difficult to
668 understand the outputs. Fortunately, SimGrid provides the execution path leading
669 to any reported issue so that you can replay this path out of the model checker,
670 enabling the usage of classical debugging tools.
672 When the model checker finds an interesting path in the application
673 execution graph (where a safety or liveness property is violated), it
674 generates an identifier for this path. Here is an example of output:
676 .. code-block:: shell
678 [ 0.000000] (0:@) Check a safety property
679 [ 0.000000] (0:@) **************************
680 [ 0.000000] (0:@) *** PROPERTY NOT VALID ***
681 [ 0.000000] (0:@) **************************
682 [ 0.000000] (0:@) Counter-example execution trace:
683 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(3)
684 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(4)
685 [ 0.000000] (0:@) Path = 1/3;1/4
686 [ 0.000000] (0:@) Expanded states = 27
687 [ 0.000000] (0:@) Visited states = 68
688 [ 0.000000] (0:@) Executed transitions = 46
690 The interesting line is ``Path = 1/3;1/4``, which means that you should use
691 ``--cfg=model-check/replay:1/3;1/4`` to replay your application on the buggy
692 execution path. All options (but the model checker related ones) must
693 remain the same. In particular, if you ran your application with
694 ``smpirun -wrapper simgrid-mc``, then do it again. Remove all
695 MC-related options, keep the other ones and add
696 ``--cfg=model-check/replay``.
698 Currently, if the path is of the form ``X;Y;Z``, each number denotes
699 the actor's pid that is selected at each indecision point. If it's of
700 the form ``X/a;Y/b``, the X and Y are the selected pids while the a
701 and b are the return values of their simcalls. In the previouse
702 example, ``1/3;1/4``, you can see from the full output that the actor
703 1 is doing MC_RANDOM simcalls, so the 3 and 4 simply denote the values
704 that these simcall return.
706 Configuring the User Code Virtualization
707 ----------------------------------------
709 .. _cfg=contexts/factory:
711 Selecting the Virtualization Factory
712 ....................................
714 **Option** contexts/factory **Default:** "raw"
716 In SimGrid, the user code is virtualized in a specific mechanism that
717 allows the simulation kernel to control its execution: when a user
718 process requires a blocking action (such as sending a message), it is
719 interrupted, and only gets released when the simulated clock reaches
720 the point where the blocking operation is done. This is explained
721 graphically in the `relevant tutorial, available online
722 <https://simgrid.org/tutorials/simgrid-simix-101.pdf>`_.
724 In SimGrid, the containers in which user processes are virtualized are
725 called contexts. Several context factory are provided, and you can
726 select the one you want to use with the ``contexts/factory``
727 configuration item. Some of the following may not exist on your
728 machine because of portability issues. In any case, the default one
729 should be the most effcient one (please report bugs if the
730 auto-detection fails for you). They are approximately sorted here from
731 the slowest to the most efficient:
733 - **thread:** very slow factory using full featured threads (either
734 pthreads or windows native threads). They are slow but very
735 standard. Some debuggers or profilers only work with this factory.
736 - **java:** Java applications are virtualized onto java threads (that
737 are regular pthreads registered to the JVM)
738 - **ucontext:** fast factory using System V contexts (Linux and FreeBSD only)
739 - **boost:** This uses the `context
740 implementation <http://www.boost.org/doc/libs/1_59_0/libs/context/doc/html/index.html>`_
741 of the boost library for a performance that is comparable to our
743 |br| Install the relevant library (e.g. with the
744 libboost-contexts-dev package on Debian/Ubuntu) and recompile
746 - **raw:** amazingly fast factory using a context switching mechanism
747 of our own, directly implemented in assembly (only available for x86
748 and amd64 platforms for now) and without any unneeded system call.
750 The main reason to change this setting is when the debugging tools get
751 fooled by the optimized context factories. Threads are the most
752 debugging-friendly contexts, as they allow one to set breakpoints
753 anywhere with gdb and visualize backtraces for all processes, in order
754 to debug concurrency issues. Valgrind is also more comfortable with
755 threads, but it should be usable with all factories (Exception: the
756 callgrind tool really dislikes raw and ucontext factories).
758 .. _cfg=contexts/stack-size:
760 Adapting the Stack Size
761 .......................
763 **Option** ``contexts/stack-size`` **Default:** 8192 KiB
765 Each virtualized used process is executed using a specific system
766 stack. The size of this stack has a huge impact on the simulation
767 scalability, but its default value is rather large. This is because
768 the error messages that you get when the stack size is too small are
769 rather disturbing: this leads to stack overflow (overwriting other
770 stacks), leading to segfaults with corrupted stack traces.
772 If you want to push the scalability limits of your code, you might
773 want to reduce the ``contexts/stack-size`` item. Its default value is
774 8192 (in KiB), while our Chord simulation works with stacks as small
775 as 16 KiB, for example. This *setting is ignored* when using the
776 thread factory. Instead, you should compile SimGrid and your
777 application with ``-fsplit-stack``. Note that this compilation flag is
778 not compatible with the model checker right now.
780 The operating system should only allocate memory for the pages of the
781 stack which are actually used and you might not need to use this in
782 most cases. However, this setting is very important when using the
783 model checker (see :ref:`options_mc_perf`).
785 .. _cfg=contexts/guard-size:
787 Disabling Stack Guard Pages
788 ...........................
790 **Option** ``contexts/guard-size`` **Default** 1 page in most case (0 pages on Windows or with MC)
792 Unless you use the threads context factory (see
793 :ref:`cfg=contexts/factory`), a stack guard page is usually used
794 which prevents the stack of a given actor from overflowing on another
795 stack. But the performance impact may become prohibitive when the
796 amount of actors increases. The option ``contexts/guard-size`` is the
797 number of stack guard pages used. By setting it to 0, no guard pages
798 will be used: in this case, you should avoid using small stacks (with
799 :ref:`contexts/stack-size <cfg=contexts/stack-size>`) as the stack
800 will silently overflow on other parts of the memory.
802 When no stack guard page is created, stacks may then silently overflow
803 on other parts of the memory if their size is too small for the
806 .. _cfg=contexts/nthreads:
807 .. _cfg=contexts/parallel-threshold:
808 .. _cfg=contexts/synchro:
810 Running User Code in Parallel
811 .............................
813 Parallel execution of the user code is only considered stable in
814 SimGrid v3.7 and higher, and mostly for MSG simulations. SMPI
815 simulations may well fail in parallel mode. It is described in
816 `INRIA RR-7653 <http://hal.inria.fr/inria-00602216/>`_.
818 If you are using the **ucontext** or **raw** context factories, you can
819 request to execute the user code in parallel. Several threads are
820 launched, each of them handling the same number of user contexts at each
821 run. To activate this, set the ``contexts/nthreads`` item to the amount
822 of cores that you have in your computer (or lower than 1 to have the
823 amount of cores auto-detected).
825 Even if you asked several worker threads using the previous option,
826 you can request to start the parallel execution (and pay the
827 associated synchronization costs) only if the potential parallelism is
828 large enough. For that, set the ``contexts/parallel-threshold``
829 item to the minimal amount of user contexts needed to start the
830 parallel execution. In any given simulation round, if that amount is
831 not reached, the contexts will be run sequentially directly by the
832 main thread (thus saving the synchronization costs). Note that this
833 option is mainly useful when the grain of the user code is very fine,
834 because our synchronization is now very efficient.
836 When parallel execution is activated, you can choose the
837 synchronization schema used with the ``contexts/synchro`` item,
838 which value is either:
840 - **futex:** ultra optimized synchronisation schema, based on futexes
841 (fast user-mode mutexes), and thus only available on Linux systems.
842 This is the default mode when available.
843 - **posix:** slow but portable synchronisation using only POSIX
845 - **busy_wait:** not really a synchronisation: the worker threads
846 constantly request new contexts to execute. It should be the most
847 efficient synchronisation schema, but it loads all the cores of
848 your machine for no good reason. You probably prefer the other less
851 Configuring the Tracing
852 -----------------------
854 The :ref:`tracing subsystem <outcomes_vizu>` can be configured in
855 several different ways depending on the nature of the simulator (MSG,
856 SimDag, SMPI) and the kind of traces that need to be obtained. See the
857 :ref:`Tracing Configuration Options subsection
858 <tracing_tracing_options>` to get a detailed description of each
859 configuration option.
861 We detail here a simple way to get the traces working for you, even if
862 you never used the tracing API.
865 - Any SimGrid-based simulator (MSG, SimDag, SMPI, ...) and raw traces:
867 .. code-block:: shell
869 --cfg=tracing:yes --cfg=tracing/uncategorized:yes --cfg=triva/uncategorized:uncat.plist
871 The first parameter activates the tracing subsystem, the second
872 tells it to trace host and link utilization (without any
873 categorization) and the third creates a graph configuration file to
874 configure Triva when analysing the resulting trace file.
876 - MSG or SimDag-based simulator and categorized traces (you need to
877 declare categories and classify your tasks according to them)
879 .. code-block:: shell
881 --cfg=tracing:yes --cfg=tracing/categorized:yes --cfg=triva/categorized:cat.plist
883 The first parameter activates the tracing subsystem, the second
884 tells it to trace host and link categorized utilization and the
885 third creates a graph configuration file to configure Triva when
886 analysing the resulting trace file.
888 - SMPI simulator and traces for a space/time view:
890 .. code-block:: shell
894 The `-trace` parameter for the smpirun script runs the simulation
895 with ``--cfg=tracing:yes --cfg=tracing/smpi:yes``. Check the
896 smpirun's `-help` parameter for additional tracing options.
898 Sometimes you might want to put additional information on the trace to
899 correctly identify them later, or to provide data that can be used to
900 reproduce an experiment. You have two ways to do that:
902 - Add a string on top of the trace file as comment:
904 .. code-block:: shell
906 --cfg=tracing/comment:my_simulation_identifier
908 - Add the contents of a textual file on top of the trace file as comment:
910 .. code-block:: shell
912 --cfg=tracing/comment-file:my_file_with_additional_information.txt
914 Please, use these two parameters (for comments) to make reproducible
915 simulations. For additional details about this and all tracing
916 options, check See the :ref:`tracing_tracing_options`.
921 .. _cfg=msg/debug-multiple-use:
926 **Option** ``msg/debug-multiple-use`` **Default:** off
928 Sometimes your application may try to send a task that is still being
929 executed somewhere else, making it impossible to send this task. However,
930 for debugging purposes, one may want to know what the other host is/was
931 doing. This option shows a backtrace of the other process.
936 The SMPI interface provides several specific configuration items.
937 These are uneasy to see since the code is usually launched through the
938 ``smiprun`` script directly.
940 .. _cfg=smpi/host-speed:
941 .. _cfg=smpi/cpu-threshold:
942 .. _cfg=smpi/simulate-computation:
944 Automatic Benchmarking of SMPI Code
945 ...................................
947 In SMPI, the sequential code is automatically benchmarked, and these
948 computations are automatically reported to the simulator. That is to
949 say that if you have a large computation between a ``MPI_Recv()`` and
950 a ``MPI_Send()``, SMPI will automatically benchmark the duration of
951 this code, and create an execution task within the simulator to take
952 this into account. For that, the actual duration is measured on the
953 host machine and then scaled to the power of the corresponding
954 simulated machine. The variable ``smpi/host-speed`` allows one to specify
955 the computational speed of the host machine (in flop/s) to use when
956 scaling the execution times. It defaults to 20000, but you really want
957 to update it to get accurate simulation results.
959 When the code is constituted of numerous consecutive MPI calls, the
960 previous mechanism feeds the simulation kernel with numerous tiny
961 computations. The ``smpi/cpu-threshold`` item becomes handy when this
962 impacts badly the simulation performance. It specifies a threshold (in
963 seconds) below which the execution chunks are not reported to the
964 simulation kernel (default value: 1e-6).
966 .. note:: The option ``smpi/cpu-threshold`` ignores any computation
967 time spent below this threshold. SMPI does not consider the
968 `amount` of these computations; there is no offset for this. Hence,
969 a value that is too small, may lead to unreliable simulation
972 In some cases, however, one may wish to disable simulation of
973 application computation. This is the case when SMPI is used not to
974 simulate an MPI applications, but instead an MPI code that performs
975 "live replay" of another MPI app (e.g., ScalaTrace's replay tool,
976 various on-line simulators that run an app at scale). In this case the
977 computation of the replay/simulation logic should not be simulated by
978 SMPI. Instead, the replay tool or on-line simulator will issue
979 "computation events", which correspond to the actual MPI simulation
980 being replayed/simulated. At the moment, these computation events can
981 be simulated using SMPI by calling internal smpi_execute*() functions.
983 To disable the benchmarking/simulation of computation in the simulated
984 application, the variable ``smpi/simulate-computation`` should be set
985 to no. This option just ignores the timings in your simulation; it
986 still executes the computations itself. If you want to stop SMPI from
987 doing that, you should check the SMPI_SAMPLE macros, documented in
988 Section :ref:`SMPI_adapting_speed`.
990 +------------------------------------+-------------------------+-----------------------------+
991 | Solution | Computations executed? | Computations simulated? |
992 +====================================+=========================+=============================+
993 | --cfg=smpi/simulate-computation:no | Yes | Never |
994 +------------------------------------+-------------------------+-----------------------------+
995 | --cfg=smpi/cpu-threshold:42 | Yes, in all cases | If it lasts over 42 seconds |
996 +------------------------------------+-------------------------+-----------------------------+
997 | SMPI_SAMPLE() macro | Only once per loop nest | Always |
998 +------------------------------------+-------------------------+-----------------------------+
1000 .. _cfg=smpi/comp-adjustment-file:
1002 Slow-down or speed-up parts of your code
1003 ........................................
1005 **Option** ``smpi/comp-adjustment-file:`` **Default:** unset
1007 This option allows you to pass a file that contains two columns: The
1008 first column defines the section that will be subject to a speedup;
1009 the second column is the speedup. For instance:
1011 .. code-block:: shell
1013 "start:stop","ratio"
1014 "exchange_1.f:30:exchange_1.f:130",1.18244559422142
1016 The first line is the header - you must include it. The following
1017 line means that the code between two consecutive MPI calls on line 30
1018 in exchange_1.f and line 130 in exchange_1.f should receive a speedup
1019 of 1.18244559422142. The value for the second column is therefore a
1020 speedup, if it is larger than 1 and a slow-down if it is smaller
1021 than 1. Nothing will be changed if it is equal to 1.
1023 Of course, you can set any arbitrary filenames you want (so the start
1024 and end don't have to be in the same file), but be aware that this
1025 mechanism only supports `consecutive calls!`
1027 Please note that you must pass the ``-trace-call-location`` flag to
1028 smpicc or smpiff, respectively. This flag activates some internal
1029 macro definitions that help with obtaining the call location.
1031 .. _cfg=smpi/bw-factor:
1036 **Option** ``smpi/bw-factor``
1037 |br| **Default:** 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
1039 The possible throughput of network links is often dependent on the
1040 message sizes, as protocols may adapt to different message sizes. With
1041 this option, a series of message sizes and factors are given, helping
1042 the simulation to be more realistic. For instance, the current default
1043 value means that messages with size 65472 and more will get a total of
1044 MAX_BANDWIDTH*0.940694, messages of size 15424 to 65471 will get
1045 MAX_BANDWIDTH*0.697866 and so on (where MAX_BANDWIDTH denotes the
1046 bandwidth of the link).
1048 An experimental script to compute these factors is available online. See
1049 https://framagit.org/simgrid/platform-calibration/
1050 https://simgrid.org/contrib/smpi-saturation-doc.html
1052 .. _cfg=smpi/display-timing:
1054 Reporting Simulation Time
1055 .........................
1057 **Option** ``smpi/display-timing`` **Default:** 0 (false)
1059 Most of the time, you run MPI code with SMPI to compute the time it
1060 would take to run it on a platform. But since the code is run through
1061 the ``smpirun`` script, you don't have any control on the launcher
1062 code, making it difficult to report the simulated time when the
1063 simulation ends. If you enable the ``smpi/display-timing`` item,
1064 ``smpirun`` will display this information when the simulation
1067 .. _cfg=smpi/keep-temps:
1069 Keeping temporary files after simulation
1070 ........................................
1072 **Option** ``smpi/keep-temps`` **default:** 0 (false)
1074 SMPI usually generates a lot of temporary files that are cleaned after
1075 use. This option request to preserve them, for example to debug or
1076 profile your code. Indeed, the binary files are removed very early
1077 under the dlopen privatization schema, which tend to fool the
1080 .. _cfg=smpi/lat-factor:
1085 **Option** ``smpi/lat-factor`` |br|
1086 **default:** 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
1088 The motivation and syntax for this option is identical to the motivation/syntax
1089 of :ref:`cfg=smpi/bw-factor`.
1091 There is an important difference, though: While smpi/bw-factor `reduces` the
1092 actual bandwidth (i.e., values between 0 and 1 are valid), latency factors
1093 increase the latency, i.e., values larger than or equal to 1 are valid here.
1095 .. _cfg=smpi/papi-events:
1097 Trace hardware counters with PAPI
1098 .................................
1100 **Option** ``smpi/papi-events`` **default:** unset
1102 When the PAPI support was compiled in SimGrid, this option takes the
1103 names of PAPI counters and adds their respective values to the trace
1104 files (See Section :ref:`tracing_tracing_options`).
1108 This feature currently requires superuser privileges, as registers
1109 are queried. Only use this feature with code you trust! Call
1110 smpirun for instance via ``smpirun -wrapper "sudo "
1111 <your-parameters>`` or run ``sudo sh -c "echo 0 >
1112 /proc/sys/kernel/perf_event_paranoid"`` In the later case, sudo
1113 will not be required.
1115 It is planned to make this feature available on a per-process (or per-thread?) basis.
1116 The first draft, however, just implements a "global" (i.e., for all processes) set
1117 of counters, the "default" set.
1119 .. code-block:: shell
1121 --cfg=smpi/papi-events:"default:PAPI_L3_LDM:PAPI_L2_LDM"
1123 .. _cfg=smpi/privatization:
1125 Automatic Privatization of Global Variables
1126 ...........................................
1128 **Option** ``smpi/privatization`` **default:** "dlopen" (when using smpirun)
1130 MPI executables are usually meant to be executed in separated
1131 processes, but SMPI is executed in only one process. Global variables
1132 from executables will be placed in the same memory zone and shared
1133 between processes, causing intricate bugs. Several options are
1134 possible to avoid this, as described in the main `SMPI publication
1135 <https://hal.inria.fr/hal-01415484>`_ and in the :ref:`SMPI
1136 documentation <SMPI_what_globals>`. SimGrid provides two ways of
1137 automatically privatizing the globals, and this option allows one to
1138 choose between them.
1140 - **no** (default when not using smpirun): Do not automatically
1141 privatize variables. Pass ``-no-privatize`` to smpirun to disable
1143 - **dlopen** or **yes** (default when using smpirun): Link multiple
1144 times against the binary.
1145 - **mmap** (slower, but maybe somewhat more stable):
1146 Runtime automatic switching of the data segments.
1149 This configuration option cannot be set in your platform file. You can only
1150 pass it as an argument to smpirun.
1152 .. _cfg=smpi/privatize-libs:
1154 Automatic privatization of global variables inside external libraries
1155 .....................................................................
1157 **Option** ``smpi/privatize-libs`` **default:** unset
1159 **Linux/BSD only:** When using dlopen (default) privatization,
1160 privatize specific shared libraries with internal global variables, if
1161 they can't be linked statically. For example libgfortran is usually
1162 used for Fortran I/O and indexes in files can be mixed up.
1164 Multiple libraries can be given, semicolon separated.
1166 This configuration option can only use either full paths to libraries,
1167 or full names. Check with ldd the name of the library you want to
1170 .. code-block:: shell
1174 libgfortran.so.3 => /usr/lib/x86_64-linux-gnu/libgfortran.so.3 (0x00007fbb4d91b000)
1177 Then you can use ``--cfg=smpi/privatize-libs:libgfortran.so.3``
1178 or ``--cfg=smpi/privatize-libs:/usr/lib/x86_64-linux-gnu/libgfortran.so.3``,
1179 but not ``libgfortran`` nor ``libgfortran.so``.
1181 .. _cfg=smpi/send-is-detached-thresh:
1183 Simulating MPI detached send
1184 ............................
1186 **Option** ``smpi/send-is-detached-thresh`` **default:** 65536
1188 This threshold specifies the size in bytes under which the send will
1189 return immediately. This is different from the threshold detailed in
1190 :ref:`cfg=smpi/async-small-thresh` because the message is not
1191 effectively sent when the send is posted. SMPI still waits for the
1192 correspondant receive to be posted to perform the communication
1195 .. _cfg=smpi/coll-selector:
1197 Simulating MPI collective algorithms
1198 ....................................
1200 **Option** ``smpi/coll-selector`` **Possible values:** naive (default), ompi, mpich
1202 SMPI implements more than 100 different algorithms for MPI collective
1203 communication, to accurately simulate the behavior of most of the
1204 existing MPI libraries. The ``smpi/coll-selector`` item can be used to
1205 use the decision logic of either OpenMPI or MPICH libraries (by
1206 default SMPI uses naive version of collective operations).
1208 Each collective operation can be manually selected with a
1209 ``smpi/collective_name:algo_name``. Available algorithms are listed in
1210 :ref:`SMPI_use_colls`.
1212 .. TODO:: All available collective algorithms will be made available
1213 via the ``smpirun --help-coll`` command.
1215 .. _cfg=smpi/iprobe:
1217 Inject constant times for MPI_Iprobe
1218 ....................................
1220 **Option** ``smpi/iprobe`` **default:** 0.0001
1222 The behavior and motivation for this configuration option is identical
1223 with :ref:`smpi/test <cfg=smpi/test>`, but for the function
1226 .. _cfg=smpi/iprobe-cpu-usage:
1228 Reduce speed for iprobe calls
1229 .............................
1231 **Option** ``smpi/iprobe-cpu-usage`` **default:** 1 (no change)
1233 MPI_Iprobe calls can be heavily used in applications. To account
1234 correctly for the energy cores spend probing, it is necessary to
1235 reduce the load that these calls cause inside SimGrid.
1237 For instance, we measured a max power consumption of 220 W for a
1238 particular application but only 180 W while this application was
1239 probing. Hence, the correct factor that should be passed to this
1240 option would be 180/220 = 0.81.
1244 Inject constant times for MPI_Init
1245 ..................................
1247 **Option** ``smpi/init`` **default:** 0
1249 The behavior and motivation for this configuration option is identical
1250 with :ref:`smpi/test <cfg=smpi/test>`, but for the function ``MPI_Init()``.
1254 Inject constant times for MPI_Isend()
1255 .....................................
1257 **Option** ``smpi/ois``
1259 The behavior and motivation for this configuration option is identical
1260 with :ref:`smpi/os <cfg=smpi/os>`, but for the function ``MPI_Isend()``.
1264 Inject constant times for MPI_send()
1265 ....................................
1267 **Option** ``smpi/os``
1269 In several network models such as LogP, send (MPI_Send, MPI_Isend) and
1270 receive (MPI_Recv) operations incur costs (i.e., they consume CPU
1271 time). SMPI can factor these costs in as well, but the user has to
1272 configure SMPI accordingly as these values may vary by machine. This
1273 can be done by using ``smpi/os`` for MPI_Send operations; for MPI_Isend
1274 and MPI_Recv, use ``smpi/ois`` and ``smpi/or``, respectively. These work
1275 exactly as ``smpi/ois``.
1277 This item can consist of multiple sections; each section takes three
1278 values, for example ``1:3:2;10:5:1``. The sections are divided by ";"
1279 so this example contains two sections. Furthermore, each section
1280 consists of three values.
1282 1. The first value denotes the minimum size for this section to take effect;
1283 read it as "if message size is greater than this value (and other section has a larger
1284 first value that is also smaller than the message size), use this".
1285 In the first section above, this value is "1".
1287 2. The second value is the startup time; this is a constant value that will always
1288 be charged, no matter what the size of the message. In the first section above,
1291 3. The third value is the `per-byte` cost. That is, it is charged for every
1292 byte of the message (incurring cost messageSize*cost_per_byte)
1293 and hence accounts also for larger messages. In the first
1294 section of the example above, this value is "2".
1296 Now, SMPI always checks which section it should take for a given
1297 message; that is, if a message of size 11 is sent with the
1298 configuration of the example above, only the second section will be
1299 used, not the first, as the first value of the second section is
1300 closer to the message size. Hence, when ``smpi/os=1:3:2;10:5:1``, a
1301 message of size 11 incurs the following cost inside MPI_Send:
1302 ``5+11*1`` because 5 is the startup cost and 1 is the cost per byte.
1304 Note that the order of sections can be arbitrary; they will be ordered internally.
1308 Inject constant times for MPI_Recv()
1309 ....................................
1311 **Option** ``smpi/or``
1313 The behavior and motivation for this configuration option is identical
1314 with :ref:`smpi/os <cfg=smpi/os>`, but for the function ``MPI_Recv()``.
1317 .. _cfg=smpi/grow-injected-times:
1319 Inject constant times for MPI_Test
1320 ..................................
1322 **Option** ``smpi/test`` **default:** 0.0001
1324 By setting this option, you can control the amount of time a process
1325 sleeps when MPI_Test() is called; this is important, because SimGrid
1326 normally only advances the time while communication is happening and
1327 thus, MPI_Test will not add to the time, resulting in a deadlock if
1328 used as a break-condition as in the following example:
1333 MPI_Test(request, flag, status);
1337 To speed up execution, we use a counter to keep track on how often we
1338 already checked if the handle is now valid or not. Hence, we actually
1339 use counter*SLEEP_TIME, that is, the time MPI_Test() causes the
1340 process to sleep increases linearly with the number of previously
1341 failed tests. This behavior can be disabled by setting
1342 ``smpi/grow-injected-times`` to **no**. This will also disable this
1343 behavior for MPI_Iprobe.
1345 .. _cfg=smpi/shared-malloc:
1346 .. _cfg=smpi/shared-malloc-hugepage:
1351 **Option** ``smpi/shared-malloc`` **Possible values:** global (default), local
1353 If your simulation consumes too much memory, you may want to modify
1354 your code so that the working areas are shared by all MPI ranks. For
1355 example, in a bloc-cyclic matrix multiplication, you will only
1356 allocate one set of blocs, and every processes will share them.
1357 Naturally, this will lead to very wrong results, but this will save a
1358 lot of memory so this is still desirable for some studies. For more on
1359 the motivation for that feature, please refer to the `relevant section
1360 <https://simgrid.github.io/SMPI_CourseWare/topic_understanding_performance/matrixmultiplication>`_
1361 of the SMPI CourseWare (see Activity #2.2 of the pointed
1362 assignment). In practice, change the call to malloc() and free() into
1363 SMPI_SHARED_MALLOC() and SMPI_SHARED_FREE().
1365 SMPI provides two algorithms for this feature. The first one, called
1366 ``local``, allocates one bloc per call to SMPI_SHARED_MALLOC() in your
1367 code (each call location gets its own bloc) and this bloc is shared
1368 amongst all MPI ranks. This is implemented with the shm_* functions
1369 to create a new POSIX shared memory object (kept in RAM, in /dev/shm)
1370 for each shared bloc.
1372 With the ``global`` algorithm, each call to SMPI_SHARED_MALLOC()
1373 returns a new address, but it only points to a shadow bloc: its memory
1374 area is mapped on a 1MiB file on disk. If the returned bloc is of size
1375 N MiB, then the same file is mapped N times to cover the whole bloc.
1376 At the end, no matter how many SMPI_SHARED_MALLOC you do, this will
1377 only consume 1 MiB in memory.
1379 You can disable this behavior and come back to regular mallocs (for
1380 example for debugging purposes) using @c "no" as a value.
1382 If you want to keep private some parts of the buffer, for instance if these
1383 parts are used by the application logic and should not be corrupted, you
1384 can use SMPI_PARTIAL_SHARED_MALLOC(size, offsets, offsets_count). Example:
1388 mem = SMPI_PARTIAL_SHARED_MALLOC(500, {27,42 , 100,200}, 2);
1390 This will allocate 500 bytes to mem, such that mem[27..41] and
1391 mem[100..199] are shared while other area remain private.
1393 Then, it can be deallocated by calling SMPI_SHARED_FREE(mem).
1395 When smpi/shared-malloc:global is used, the memory consumption problem
1396 is solved, but it may induce too much load on the kernel's pages table.
1397 In this case, you should use huge pages so that we create only one
1398 entry per Mb of malloced data instead of one entry per 4k.
1399 To activate this, you must mount a hugetlbfs on your system and allocate
1400 at least one huge page:
1402 .. code-block:: shell
1405 sudo mount none /home/huge -t hugetlbfs -o rw,mode=0777
1406 sudo sh -c 'echo 1 > /proc/sys/vm/nr_hugepages' # echo more if you need more
1408 Then, you can pass the option
1409 ``--cfg=smpi/shared-malloc-hugepage:/home/huge`` to smpirun to
1410 actually activate the huge page support in shared mallocs.
1414 Inject constant times for MPI_Wtime, gettimeofday and clock_gettime
1415 ...................................................................
1417 **Option** ``smpi/wtime`` **default:** 10 ns
1419 This option controls the amount of (simulated) time spent in calls to
1420 MPI_Wtime(), gettimeofday() and clock_gettime(). If you set this value
1421 to 0, the simulated clock is not advanced in these calls, which leads
1422 to issue if your application contains such a loop:
1426 while(MPI_Wtime() < some_time_bound) {
1427 /* some tests, with no communication nor computation */
1430 When the option smpi/wtime is set to 0, the time advances only on
1431 communications and computations, so the previous code results in an
1432 infinite loop: the current [simulated] time will never reach
1433 ``some_time_bound``. This infinite loop is avoided when that option
1434 is set to a small amount, as it is by default since SimGrid v3.21.
1436 Note that if your application does not contain any loop depending on
1437 the current time only, then setting this option to a non-zero value
1438 will slow down your simulations by a tiny bit: the simulation loop has
1439 to be broken and reset each time your code ask for the current time.
1440 If the simulation speed really matters to you, you can avoid this
1441 extra delay by setting smpi/wtime to 0.
1443 Other Configurations
1444 --------------------
1446 .. _cfg=debug/clean-atexit:
1448 Cleanup at Termination
1449 ......................
1451 **Option** ``debug/clean-atexit`` **default:** on
1453 If your code is segfaulting during its finalization, it may help to
1454 disable this option to request SimGrid to not attempt any cleanups at
1455 the end of the simulation. Since the Unix process is ending anyway,
1456 the operating system will wipe it all.
1463 **Option** ``path`` **default:** . (current dir)
1465 It is possible to specify a list of directories to search into for the
1466 trace files (see :ref:`pf_trace`) by using this configuration
1467 item. To add several directory to the path, set the configuration
1468 item several times, as in ``--cfg=path:toto --cfg=path:tutu``
1470 .. _cfg=debug/breakpoint:
1475 **Option** ``debug/breakpoint`` **default:** unset
1477 This configuration option sets a breakpoint: when the simulated clock
1478 reaches the given time, a SIGTRAP is raised. This can be used to stop
1479 the execution and get a backtrace with a debugger.
1481 It is also possible to set the breakpoint from inside the debugger, by
1482 writing in global variable simgrid::simix::breakpoint. For example,
1485 .. code-block:: shell
1487 set variable simgrid::simix::breakpoint = 3.1416
1489 .. _cfg=debug/verbose-exit:
1494 **Option** ``debug/verbose-exit`` **default:** on
1496 By default, when Ctrl-C is pressed, the status of all existing actors
1497 is displayed before exiting the simulation. This is very useful to
1498 debug your code, but it can reveal troublesome if you have many
1499 actors. Set this configuration item to **off** to disable this
1502 .. _cfg=exception/cutpath:
1504 Truncate local path from exception backtrace
1505 ............................................
1507 **Option** ``exception/cutpath`` **default:** off
1509 This configuration option is used to remove the path from the
1510 backtrace shown when an exception is thrown. This is mainly useful for
1511 the tests: the full file path makes the tests not reproducible because
1512 the path of source files depend of the build settings. That would
1513 break most of our tests as we keep comparing output.
1515 Logging Configuration
1516 ---------------------
1518 It can be done by using XBT. Go to :ref:`XBT_log` for more details.