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 occure 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.
84 - **clean-atexit:** :ref:`cfg=clean-atexit`
86 - **contexts/factory:** :ref:`cfg=contexts/factory`
87 - **contexts/guard-size:** :ref:`cfg=contexts/guard-size`
88 - **contexts/nthreads:** :ref:`cfg=contexts/nthreads`
89 - **contexts/parallel-threshold:** :ref:`cfg=contexts/parallel-threshold`
90 - **contexts/stack-size:** :ref:`cfg=contexts/stack-size`
91 - **contexts/synchro:** :ref:`cfg=contexts/synchro`
93 - **cpu/maxmin-selective-update:** :ref:`Cpu Optimization Level <options_model_optim>`
94 - **cpu/model:** :ref:`options_model_select`
95 - **cpu/optim:** :ref:`Cpu Optimization Level <options_model_optim>`
97 - **exception/cutpath:** :ref:`cfg=exception/cutpath`
99 - **host/model:** :ref:`options_model_select`
101 - **maxmin/precision:** :ref:`cfg=maxmin/precision`
102 - **maxmin/concurrency-limit:** :ref:`cfg=maxmin/concurrency-limit`
104 - **msg/debug-multiple-use:** :ref:`cfg=msg/debug-multiple-use`
106 - **model-check:** :ref:`options_modelchecking`
107 - **model-check/checkpoint:** :ref:`cfg=model-check/checkpoint`
108 - **model-check/communications-determinism:** :ref:`cfg=model-check/communications-determinism`
109 - **model-check/dot-output:** :ref:`cfg=model-check/dot-output`
110 - **model-check/hash:** :ref:`cfg=model-checker/hash`
111 - **model-check/max-depth:** :ref:`cfg=model-check/max-depth`
112 - **model-check/property:** :ref:`cfg=model-check/property`
113 - **model-check/reduction:** :ref:`cfg=model-check/reduction`
114 - **model-check/replay:** :ref:`cfg=model-check/replay`
115 - **model-check/send-determinism:** :ref:`cfg=model-check/send-determinism`
116 - **model-check/sparse-checkpoint:** :ref:`cfg=model-check/sparse-checkpoint`
117 - **model-check/termination:** :ref:`cfg=model-check/termination`
118 - **model-check/timeout:** :ref:`cfg=model-check/timeout`
119 - **model-check/visited:** :ref:`cfg=model-check/visited`
121 - **network/bandwidth-factor:** :ref:`cfg=network/bandwidth-factor`
122 - **network/crosstraffic:** :ref:`cfg=network/crosstraffic`
123 - **network/latency-factor:** :ref:`cfg=network/latency-factor`
124 - **network/maxmin-selective-update:** :ref:`Network Optimization Level <options_model_optim>`
125 - **network/model:** :ref:`options_model_select`
126 - **network/optim:** :ref:`Network Optimization Level <options_model_optim>`
127 - **network/TCP-gamma:** :ref:`cfg=network/TCP-gamma`
128 - **network/weight-S:** :ref:`cfg=network/weight-S`
130 - **ns3/TcpModel:** :ref:`options_pls`
131 - **path:** :ref:`cfg=path`
132 - **plugin:** :ref:`cfg=plugin`
134 - **simix/breakpoint:** :ref:`cfg=simix/breakpoint`
136 - **storage/max_file_descriptors:** :ref:`cfg=storage/max_file_descriptors`
138 - **surf/precision:** :ref:`cfg=surf/precision`
140 - **For collective operations of SMPI,** please refer to Section :ref:`cfg=smpi/coll-selector`
141 - **smpi/async-small-thresh:** :ref:`cfg=smpi/async-small-thresh`
142 - **smpi/bw-factor:** :ref:`cfg=smpi/bw-factor`
143 - **smpi/coll-selector:** :ref:`cfg=smpi/coll-selector`
144 - **smpi/comp-adjustment-file:** :ref:`cfg=smpi/comp-adjustment-file`
145 - **smpi/cpu-threshold:** :ref:`cfg=smpi/cpu-threshold`
146 - **smpi/display-timing:** :ref:`cfg=smpi/display-timing`
147 - **smpi/grow-injected-times:** :ref:`cfg=smpi/grow-injected-times`
148 - **smpi/host-speed:** :ref:`cfg=smpi/host-speed`
149 - **smpi/IB-penalty-factors:** :ref:`cfg=smpi/IB-penalty-factors`
150 - **smpi/iprobe:** :ref:`cfg=smpi/iprobe`
151 - **smpi/iprobe-cpu-usage:** :ref:`cfg=smpi/iprobe-cpu-usage`
152 - **smpi/init:** :ref:`cfg=smpi/init`
153 - **smpi/keep-temps:** :ref:`cfg=smpi/keep-temps`
154 - **smpi/lat-factor:** :ref:`cfg=smpi/lat-factor`
155 - **smpi/ois:** :ref:`cfg=smpi/ois`
156 - **smpi/or:** :ref:`cfg=smpi/or`
157 - **smpi/os:** :ref:`cfg=smpi/os`
158 - **smpi/papi-events:** :ref:`cfg=smpi/papi-events`
159 - **smpi/privatization:** :ref:`cfg=smpi/privatization`
160 - **smpi/privatize-libs:** :ref:`cfg=smpi/privatize-libs`
161 - **smpi/send-is-detached-thresh:** :ref:`cfg=smpi/send-is-detached-thresh`
162 - **smpi/shared-malloc:** :ref:`cfg=smpi/shared-malloc`
163 - **smpi/shared-malloc-hugepage:** :ref:`cfg=smpi/shared-malloc-hugepage`
164 - **smpi/simulate-computation:** :ref:`cfg=smpi/simulate-computation`
165 - **smpi/test:** :ref:`cfg=smpi/test`
166 - **smpi/wtime:** :ref:`cfg=smpi/wtime`
168 - **Tracing configuration options** can be found in Section :ref:`tracing_tracing_options`
170 - **storage/model:** :ref:`options_model_select`
171 - **verbose-exit:** :ref:`cfg=verbose-exit`
173 - **vm/model:** :ref:`options_model_select`
177 Configuring the Platform Models
178 -------------------------------
180 .. _options_model_select:
182 Choosing the Platform Models
183 ............................
185 SimGrid comes with several network, CPU and storage models built in,
186 and you can change the used model at runtime by changing the passed
187 configuration. The three main configuration items are given below.
188 For each of these items, passing the special ``help`` value gives you
189 a short description of all possible values (for example,
190 ``--cfg=network/model:help`` will present all provided network
191 models). Also, ``--help-models`` should provide information about all
192 models for all existing resources.
194 - ``network/model``: specify the used network model. Possible values:
196 - **LV08 (default one):** Realistic network analytic model
197 (slow-start modeled by multiplying latency by 13.01, bandwidth by
198 .97; bottleneck sharing uses a payload of S=20537 for evaluating
199 RTT). Described in `Accuracy Study and Improvement of Network
200 Simulation in the SimGrid Framework
201 <http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf>`_.
202 - **Constant:** Simplistic network model where all communication
203 take a constant time (one second). This model provides the lowest
204 realism, but is (marginally) faster.
205 - **SMPI:** Realistic network model specifically tailored for HPC
206 settings (accurate modeling of slow start with correction factors on
207 three intervals: < 1KiB, < 64 KiB, >= 64 KiB). This model can be
208 :ref:`further configured <options_model_network>`.
209 - **IB:** Realistic network model specifically tailored for HPC
210 settings with InfiniBand networks (accurate modeling contention
211 behavior, based on the model explained in `this PhD work
212 <http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf>`_.
213 This model can be :ref:`further configured <options_model_network>`.
214 - **CM02:** Legacy network analytic model. Very similar to LV08, but
215 without corrective factors. The timings of small messages are thus
216 poorly modeled. This model is described in `A Network Model for
217 Simulation of Grid Application
218 <ftp://ftp.ens-lyon.fr/pub/LIP/Rapports/RR/RR2002/RR2002-40.ps.gz>`_.
219 - **Reno/Reno2/Vegas:** Models from Steven H. Low using lagrange_solve instead of
220 lmm_solve (experts only; check the code for more info).
221 - **NS3** (only available if you compiled SimGrid accordingly):
222 Use the packet-level network
223 simulators as network models (see :ref:`pls_ns3`).
224 This model can be :ref:`further configured <options_pls>`.
226 - ``cpu/model``: specify the used CPU model. We have only one model
229 - **Cas01:** Simplistic CPU model (time=size/power)
231 - ``host/model``: The host concept is the aggregation of a CPU with a
232 network card. Three models exists, but actually, only 2 of them are
233 interesting. The "compound" one is simply due to the way our
234 internal code is organized, and can easily be ignored. So at the
235 end, you have two host models: The default one allows to aggregate
236 an existing CPU model with an existing network model, but does not
237 allow parallel tasks because these beasts need some collaboration
238 between the network and CPU model. That is why, ptask_07 is used by
239 default when using SimDag.
241 - **default:** Default host model. Currently, CPU:Cas01 and
242 network:LV08 (with cross traffic enabled)
243 - **compound:** Host model that is automatically chosen if
244 you change the network and CPU models
245 - **ptask_L07:** Host model somehow similar to Cas01+CM02 but
246 allowing "parallel tasks", that are intended to model the moldable
247 tasks of the grid scheduling literature.
249 - ``storage/model``: specify the used storage model. Only one model is
251 - ``vm/model``: specify the model for virtual machines. Only one model
254 .. todo: make 'compound' the default host model.
256 .. _options_model_optim:
261 The network and CPU models that are based on lmm_solve (that
262 is, all our analytical models) accept specific optimization
265 - items ``network/optim`` and ``cpu/optim`` (both default to 'Lazy'):
267 - **Lazy:** Lazy action management (partial invalidation in lmm +
268 heap in action remaining).
269 - **TI:** Trace integration. Highly optimized mode when using
270 availability traces (only available for the Cas01 CPU model for
272 - **Full:** Full update of remaining and variables. Slow but may be
273 useful when debugging.
275 - items ``network/maxmin-selective-update`` and
276 ``cpu/maxmin-selective-update``: configure whether the underlying
277 should be lazily updated or not. It should have no impact on the
278 computed timings, but should speed up the computation. |br| It is
279 still possible to disable this feature because it can reveal
280 counter-productive in very specific scenarios where the
281 interaction level is high. In particular, if all your
282 communication share a given backbone link, you should disable it:
283 without it, a simple regular loop is used to update each
284 communication. With it, each of them is still updated (because of
285 the dependency induced by the backbone), but through a complicated
286 and slow pattern that follows the actual dependencies.
288 .. _cfg=maxmin/precision:
289 .. _cfg=surf/precision:
294 **Option** ``maxmin/precision`` **Default:** 0.00001 (in flops or bytes) |br|
295 **Option** ``surf/precision`` **Default:** 0.00001 (in seconds)
297 The analytical models handle a lot of floating point values. It is
298 possible to change the epsilon used to update and compare them through
299 this configuration item. Changing it may speedup the simulation by
300 discarding very small actions, at the price of a reduced numerical
301 precision. You can modify separately the precision used to manipulate
302 timings (in seconds) and the one used to manipulate amounts of work
305 .. _cfg=maxmin/concurrency-limit:
310 **Option** ``maxmin/concurrency-limit`` **Default:** -1 (no limit)
312 The maximum number of variables per resource can be tuned through this
313 option. You can have as many simultaneous actions per resources as you
314 want. If your simulation presents a very high level of concurrency, it
315 may help to use e.g. 100 as a value here. It means that at most 100
316 actions can consume a resource at a given time. The extraneous actions
317 are queued and wait until the amount of concurrency of the considered
318 resource lowers under the given boundary.
320 Such limitations help both to the simulation speed and simulation accuracy
321 on highly constrained scenarios, but the simulation speed suffers of this
322 setting on regular (less constrained) scenarios so it is off by default.
324 .. _options_model_network:
326 Configuring the Network Model
327 .............................
329 .. _cfg=network/TCP-gamma:
331 Maximal TCP Window Size
332 ^^^^^^^^^^^^^^^^^^^^^^^
334 **Option** ``network/TCP-gamma`` **Default:** 4194304
336 The analytical models need to know the maximal TCP window size to take
337 the TCP congestion mechanism into account. On Linux, this value can
338 be retrieved using the following commands. Both give a set of values,
339 and you should use the last one, which is the maximal size.
341 .. code-block:: shell
343 cat /proc/sys/net/ipv4/tcp_rmem # gives the sender window
344 cat /proc/sys/net/ipv4/tcp_wmem # gives the receiver window
346 .. _cfg=smpi/IB-penalty-factors:
347 .. _cfg=network/bandwidth-factor:
348 .. _cfg=network/latency-factor:
349 .. _cfg=network/weight-S:
351 Correcting Important Network Parameters
352 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
354 SimGrid can take network irregularities such as a slow startup or
355 changing behavior depending on the message size into account. You
356 should not change these values unless you really know what you're
357 doing. The corresponding values were computed through data fitting
358 one the timings of packet-level simulators, as described in `Accuracy
359 Study and Improvement of Network Simulation in the SimGrid Framework
360 <http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf>`_.
363 If you are using the SMPI model, these correction coefficients are
364 themselves corrected by constant values depending on the size of the
365 exchange. By default SMPI uses factors computed on the Stampede
366 Supercomputer at TACC, with optimal deployment of processes on
367 nodes. Again, only hardcore experts should bother about this fact.
369 InfiniBand network behavior can be modeled through 3 parameters
370 ``smpi/IB-penalty-factors:"βe;βs;γs"``, as explained in `this PhD
372 <http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf>`_.
374 .. todo:: This section should be rewritten, and actually explain the
375 options network/bandwidth-factor, network/latency-factor,
378 .. _cfg=network/crosstraffic:
380 Simulating Cross-Traffic
381 ^^^^^^^^^^^^^^^^^^^^^^^^
383 Since SimGrid v3.7, cross-traffic effects can be taken into account in
384 analytical simulations. It means that ongoing and incoming
385 communication flows are treated independently. In addition, the LV08
386 model adds 0.05 of usage on the opposite direction for each new
387 created flow. This can be useful to simulate some important TCP
388 phenomena such as ack compression.
390 For that to work, your platform must have two links for each
391 pair of interconnected hosts. An example of usable platform is
392 available in ``examples/platforms/crosstraffic.xml``.
394 This is activated through the ``network/crosstraffic`` item, that
395 can be set to 0 (disable this feature) or 1 (enable it).
397 Note that with the default host model this option is activated by default.
399 .. _cfg=smpi/async-small-thresh:
401 Simulating Asyncronous Send
402 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
404 (this configuration item is experimental and may change or disapear)
406 It is possible to specify that messages below a certain size will be
407 sent as soon as the call to MPI_Send is issued, without waiting for
408 the correspondant receive. This threshold can be configured through
409 the ``smpi/async-small-thresh`` item. The default value is 0. This
410 behavior can also be manually set for mailboxes, by setting the
411 receiving mode of the mailbox with a call to
412 :cpp:func:`MSG_mailbox_set_async`. After this, all messages sent to
413 this mailbox will have this behavior regardless of the message size.
415 This value needs to be smaller than or equals to the threshold set at
416 @ref options_model_smpi_detached , because asynchronous messages are
417 meant to be detached as well.
424 **Option** ``ns3/TcpModel`` **Default:** "default" (NS3 default)
426 When using NS3, there is an extra item ``ns3/TcpModel``, corresponding
427 to the ``ns3::TcpL4Protocol::SocketType`` configuration item in
428 NS3. The only valid values (enforced on the SimGrid side) are
429 'default' (no change to the NS3 configuration), 'NewReno' or 'Reno' or
432 Configuring the Storage model
433 .............................
435 .. _cfg=storage/max_file_descriptors:
437 File Descriptor Cound per Host
438 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
440 **Option** ``storage/max_file_descriptors`` **Default:** 1024
442 Each host maintains a fixed-size array of its file descriptors. You
443 can change its size through this item to either enlarge it if your
444 application requires it or to reduce it to save memory space.
451 SimGrid plugins allow to extend the framework without changing its
452 source code directly. Read the source code of the existing plugins to
453 learn how to do so (in ``src/plugins``), and ask your questions to the
454 usual channels (Stack Overflow, Mailing list, IRC). The basic idea is
455 that plugins usually register callbacks to some signals of interest.
456 If they need to store some information about a given object (Link, CPU
457 or Actor), they do so through the use of a dedicated object extension.
459 Some of the existing plugins can be activated from the command line,
460 meaning that you can activate them from the command line without any
461 modification to your simulation code. For example, you can activate
462 the host energy plugin by adding ``--cfg=plugin:host_energy`` to your
465 Here is the full list of plugins that can be activated this way:
467 - **host_energy:** keeps track of the energy dissipated by
468 computations. More details in @ref plugin_energy.
469 - **link_energy:** keeps track of the energy dissipated by
470 communications. More details in @ref SURF_plugin_energy.
471 - **host_load:** keeps track of the computational load.
472 More details in @ref plugin_load.
474 .. _options_modelchecking:
476 Configuring the Model-Checking
477 ------------------------------
479 To enable the SimGrid model-checking support the program should
480 be executed using the simgrid-mc wrapper:
482 .. code-block:: shell
484 simgrid-mc ./my_program
486 Safety properties are expressed as assertions using the function
487 :cpp:func:`void MC_assert(int prop)`.
489 .. _cfg=model-check/property:
491 Specifying a liveness property
492 ..............................
494 **Option** ``model-check/property`` **Default:** unset
496 If you want to specify liveness properties, you have to pass them on
497 the command line, specifying the name of the file containing the
498 property, as formatted by the ltl2ba program.
501 .. code-block:: shell
503 simgrid-mc ./my_program --cfg=model-check/property:<filename>
505 .. _cfg=model-check/checkpoint:
507 Going for Stateful Verification
508 ...............................
510 By default, the system is backtracked to its initial state to explore
511 another path instead of backtracking to the exact step before the fork
512 that we want to explore (this is called stateless verification). This
513 is done this way because saving intermediate states can rapidly
514 exhaust the available memory. If you want, you can change the value of
515 the ``model-check/checkpoint`` item. For example,
516 ``--cfg=model-check/checkpoint:1`` asks to take a checkpoint every
517 step. Beware, this will certainly explode your memory. Larger values
518 are probably better, make sure to experiment a bit to find the right
519 setting for your specific system.
521 .. _cfg=model-check/reduction:
523 Specifying the kind of reduction
524 ................................
526 The main issue when using the model-checking is the state space
527 explosion. To counter that problem, you can chose a exploration
528 reduction techniques with
529 ``--cfg=model-check/reduction:<technique>``. For now, this
530 configuration variable can take 2 values:
532 - **none:** Do not apply any kind of reduction (mandatory for now for
534 - **dpor:** Apply Dynamic Partial Ordering Reduction. Only valid if
535 you verify local safety properties (default value for safety
538 There is unfortunately no silver bullet here, and the most efficient
539 reduction techniques cannot be applied to any properties. In
540 particular, the DPOR method cannot be applied on liveness properties
541 since our implementation of DPOR may break some cycles, while cycles
542 are very important to the soundness of the exploration for liveness
545 .. _cfg=model-check/visited:
547 Size of Cycle Detection Set
548 ...........................
550 In order to detect cycles, the model-checker needs to check if a new
551 explored state is in fact the same state than a previous one. For
552 that, the model-checker can take a snapshot of each visited state:
553 this snapshot is then used to compare it with subsequent states in the
556 The ``model-check/visited`` item is the maximum number of states which
557 are stored in memory. If the maximum number of snapshotted state is
558 reached, some states will be removed from the memory and some cycles
559 might be missed. Small values can lead to incorrect verifications, but
560 large value can exhaust your memory, so choose carefully.
562 By default, no state is snapshotted and cycles cannot be detected.
564 .. _cfg=model-check/termination:
566 Non-Termination Detection
567 .........................
569 The ``model-check/termination`` configuration item can be used to
570 report if a non-termination execution path has been found. This is a
571 path with a cycle which means that the program might never terminate.
573 This only works in safety mode, not in liveness mode.
575 This options is disabled by default.
577 .. _cfg=model-check/dot-output:
582 If set, the ``model-check/dot-output`` configuration item is the name
583 of a file in which to write a dot file of the path leading the found
584 property (safety or liveness violation) as well as the cycle for
585 liveness properties. This dot file can then fed to the graphviz dot
586 tool to generate an corresponding graphical representation.
588 .. _cfg=model-check/max-depth:
590 Exploration Depth Limit
591 .......................
593 The ``model-checker/max-depth`` can set the maximum depth of the
594 exploration graph of the model-checker. If this limit is reached, a
595 logging message is sent and the results might not be exact.
597 By default, there is not depth limit.
599 .. _cfg=model-check/timeout:
604 By default, the model-checker does not handle timeout conditions: the `wait`
605 operations never time out. With the ``model-check/timeout`` configuration item
606 set to **yes**, the model-checker will explore timeouts of `wait` operations.
608 .. _cfg=model-check/communications-determinism:
609 .. _cfg=model-check/send-determinism:
611 Communication Determinism
612 .........................
614 The ``model-check/communications-determinism`` and
615 ``model-check/send-determinism`` items can be used to select the
616 communication determinism mode of the model-checker which checks
617 determinism properties of the communications of an application.
619 .. _cfg=model-check/sparse-checkpoint:
621 Incremental Checkpoints
622 .......................
624 When the model-checker is configured to take a snapshot of each
625 explored state (with the ``model-checker/visited`` item), the memory
626 consumption can rapidly reach GiB ou Tib of memory. However, for many
627 workloads, the memory does not change much between different snapshots
628 and taking a complete copy of each snapshot is a waste of memory.
630 The ``model-check/sparse-checkpoint`` option item can be set to
631 **yes** to avoid making a complete copy of each snapshot. Instead,
632 each snapshot will be decomposed in blocks which will be stored
633 separately. If multiple snapshots share the same block (or if the
634 same block is used in the same snapshot), the same copy of the block
635 will be shared leading to a reduction of the memory footprint.
637 For many applications, this option considerably reduces the memory
638 consumption. In somes cases, the model-checker might be slightly
639 slower because of the time taken to manage the metadata about the
640 blocks. In other cases however, this snapshotting strategy will be
641 much faster by reducing the cache consumption. When the memory
642 consumption is important, by avoiding to hit the swap or reducing the
643 swap usage, this option might be much faster than the basic
644 snapshotting strategy.
646 This option is currently disabled by default.
648 Verification Performance Considerations
649 .......................................
651 The size of the stacks can have a huge impact on the memory
652 consumption when using model-checking. By default, each snapshot will
653 save a copy of the whole stacks and not only of the part which is
654 really meaningful: you should expect the contribution of the memory
655 consumption of the snapshots to be @f$ @mbox{number of processes}
656 @times @mbox{stack size} @times @mbox{number of states} @f$.
658 The ``model-check/sparse-checkpoint`` can be used to reduce the memory
659 consumption by trying to share memory between the different snapshots.
661 When compiled against the model checker, the stacks are not
662 protected with guards: if the stack size is too small for your
663 application, the stack will silently overflow on other parts of the
664 memory (see :ref:`contexts/guard-size <cfg=contexts/guard-size>`).
666 .. _cfg=model-checker/hash:
671 Usually most of the time of the model-checker is spent comparing states. This
672 process is complicated and consumes a lot of bandwidth and cache.
673 In order to speedup the state comparison, the experimental ``model-checker/hash``
674 configuration item enables the computation of a hash summarizing as much
675 information of the state as possible into a single value. This hash can be used
676 to avoid most of the comparisons: the costly comparison is then only used when
677 the hashes are identical.
679 Currently most of the state is not included in the hash because the
680 implementation was found to be buggy and this options is not as useful as
681 it could be. For this reason, it is currently disabled by default.
683 .. _cfg=model-check/replay:
685 Replaying buggy execution paths out of the model-checker
686 ........................................................
688 Debugging the problems reported by the model-checker is challenging: First, the
689 application under verification cannot be debugged with gdb because the
690 model-checker already traces it. Then, the model-checker may explore several
691 execution paths before encountering the issue, making it very difficult to
692 understand the outputs. Fortunately, SimGrid provides the execution path leading
693 to any reported issue so that you can replay this path out of the model checker,
694 enabling the usage of classical debugging tools.
696 When the model-checker finds an interesting path in the application
697 execution graph (where a safety or liveness property is violated), it
698 generates an identifier for this path. Here is an example of output:
700 .. code-block:: shell
702 [ 0.000000] (0:@) Check a safety property
703 [ 0.000000] (0:@) **************************
704 [ 0.000000] (0:@) *** PROPERTY NOT VALID ***
705 [ 0.000000] (0:@) **************************
706 [ 0.000000] (0:@) Counter-example execution trace:
707 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(3)
708 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(4)
709 [ 0.000000] (0:@) Path = 1/3;1/4
710 [ 0.000000] (0:@) Expanded states = 27
711 [ 0.000000] (0:@) Visited states = 68
712 [ 0.000000] (0:@) Executed transitions = 46
714 The interesting line is ``Path = 1/3;1/4``, which means that you should use
715 `--cfg=model-check/replay:1/3;1/4`` to replay your application on the buggy
716 execution path. The other options should be the same (but the model-checker
717 should be disabled). Note that format and meaning of the path may change between
720 Configuring the User Code Virtualization
721 ----------------------------------------
723 .. _cfg=contexts/factory:
725 Selecting the Virtualization Factory
726 ....................................
728 **Option** contexts/factory **Default:** "raw"
730 In SimGrid, the user code is virtualized in a specific mechanism that
731 allows the simulation kernel to control its execution: when a user
732 process requires a blocking action (such as sending a message), it is
733 interrupted, and only gets released when the simulated clock reaches
734 the point where the blocking operation is done. This is explained
735 graphically in the `relevant tutorial, available online
736 <https://simgrid.org/tutorials/simgrid-simix-101.pdf>`_.
738 In SimGrid, the containers in which user processes are virtualized are
739 called contexts. Several context factory are provided, and you can
740 select the one you want to use with the ``contexts/factory``
741 configuration item. Some of the following may not exist on your
742 machine because of portability issues. In any case, the default one
743 should be the most effcient one (please report bugs if the
744 auto-detection fails for you). They are approximately sorted here from
745 the slowest to the most efficient:
747 - **thread:** very slow factory using full featured threads (either
748 pthreads or windows native threads). They are slow but very
749 standard. Some debuggers or profilers only work with this factory.
750 - **java:** Java applications are virtualized onto java threads (that
751 are regular pthreads registered to the JVM)
752 - **ucontext:** fast factory using System V contexts (Linux and FreeBSD only)
753 - **boost:** This uses the `context
754 implementation <http://www.boost.org/doc/libs/1_59_0/libs/context/doc/html/index.html>`_
755 of the boost library for a performance that is comparable to our
757 |br| Install the relevant library (e.g. with the
758 libboost-contexts-dev package on Debian/Ubuntu) and recompile
760 - **raw:** amazingly fast factory using a context switching mechanism
761 of our own, directly implemented in assembly (only available for x86
762 and amd64 platforms for now) and without any unneeded system call.
764 The main reason to change this setting is when the debugging tools get
765 fooled by the optimized context factories. Threads are the most
766 debugging-friendly contextes, as they allow to set breakpoints
767 anywhere with gdb and visualize backtraces for all processes, in order
768 to debug concurrency issues. Valgrind is also more comfortable with
769 threads, but it should be usable with all factories (Exception: the
770 callgrind tool really dislikes raw and ucontext factories).
772 .. _cfg=contexts/stack-size:
774 Adapting the Stack Size
775 .......................
777 **Option** ``contexts/stack-size`` **Default:** 8192 KiB
779 Each virtualized used process is executed using a specific system
780 stack. The size of this stack has a huge impact on the simulation
781 scalability, but its default value is rather large. This is because
782 the error messages that you get when the stack size is too small are
783 rather disturbing: this leads to stack overflow (overwriting other
784 stacks), leading to segfaults with corrupted stack traces.
786 If you want to push the scalability limits of your code, you might
787 want to reduce the ``contexts/stack-size`` item. Its default value is
788 8192 (in KiB), while our Chord simulation works with stacks as small
789 as 16 KiB, for example. This *setting is ignored* when using the
790 thread factory. Instead, you should compile SimGrid and your
791 application with ``-fsplit-stack``. Note that this compilation flag is
792 not compatible with the model-checker right now.
794 The operating system should only allocate memory for the pages of the
795 stack which are actually used and you might not need to use this in
796 most cases. However, this setting is very important when using the
797 model checker (see :ref:`options_mc_perf`).
799 .. _cfg=contexts/guard-size:
801 Disabling Stack Guard Pages
802 ...........................
804 **Option** ``contexts/guard-size`` **Default** 1 page in most case (0 pages on Windows or with MC)
806 Unless you use the threads context factory (see
807 :ref:`cfg=contexts/factory`), a stack guard page is usually used
808 which prevents the stack of a given actor from overflowing on another
809 stack. But the performance impact may become prohibitive when the
810 amount of actors increases. The option ``contexts/guard-size`` is the
811 number of stack guard pages used. By setting it to 0, no guard pages
812 will be used: in this case, you should avoid using small stacks (with
813 :ref:`contexts/stack-size <cfg=contexts/stack-size>`) as the stack
814 will silently overflow on other parts of the memory.
816 When no stack guard page is created, stacks may then silently overflow
817 on other parts of the memory if their size is too small for the
820 .. _cfg=contexts/nthreads:
821 .. _cfg=contexts/parallel-threshold:
822 .. _cfg=contexts/synchro:
824 Running User Code in Parallel
825 .............................
827 Parallel execution of the user code is only considered stable in
828 SimGrid v3.7 and higher, and mostly for MSG simulations. SMPI
829 simulations may well fail in parallel mode. It is described in
830 `INRIA RR-7653 <http://hal.inria.fr/inria-00602216/>`_.
832 If you are using the **ucontext** or **raw** context factories, you can
833 request to execute the user code in parallel. Several threads are
834 launched, each of them handling as much user contexts at each run. To
835 actiave this, set the ``contexts/nthreads`` item to the amount of
836 cores that you have in your computer (or lower than 1 to have
837 the amount of cores auto-detected).
839 Even if you asked several worker threads using the previous option,
840 you can request to start the parallel execution (and pay the
841 associated synchronization costs) only if the potential parallelism is
842 large enough. For that, set the ``contexts/parallel-threshold``
843 item to the minimal amount of user contexts needed to start the
844 parallel execution. In any given simulation round, if that amount is
845 not reached, the contexts will be run sequentially directly by the
846 main thread (thus saving the synchronization costs). Note that this
847 option is mainly useful when the grain of the user code is very fine,
848 because our synchronization is now very efficient.
850 When parallel execution is activated, you can choose the
851 synchronization schema used with the ``contexts/synchro`` item,
852 which value is either:
854 - **futex:** ultra optimized synchronisation schema, based on futexes
855 (fast user-mode mutexes), and thus only available on Linux systems.
856 This is the default mode when available.
857 - **posix:** slow but portable synchronisation using only POSIX
859 - **busy_wait:** not really a synchronisation: the worker threads
860 constantly request new contexts to execute. It should be the most
861 efficient synchronisation schema, but it loads all the cores of
862 your machine for no good reason. You probably prefer the other less
865 Configuring the Tracing
866 -----------------------
868 The :ref:`tracing subsystem <outcomes_vizu>` can be configured in
869 several different ways depending on the nature of the simulator (MSG,
870 SimDag, SMPI) and the kind of traces that need to be obtained. See the
871 :ref:`Tracing Configuration Options subsection
872 <tracing_tracing_options>` to get a detailed description of each
873 configuration option.
875 We detail here a simple way to get the traces working for you, even if
876 you never used the tracing API.
879 - Any SimGrid-based simulator (MSG, SimDag, SMPI, ...) and raw traces:
881 .. code-block:: shell
883 --cfg=tracing:yes --cfg=tracing/uncategorized:yes --cfg=triva/uncategorized:uncat.plist
885 The first parameter activates the tracing subsystem, the second
886 tells it to trace host and link utilization (without any
887 categorization) and the third creates a graph configuration file to
888 configure Triva when analysing the resulting trace file.
890 - MSG or SimDag-based simulator and categorized traces (you need to
891 declare categories and classify your tasks according to them)
893 .. code-block:: shell
895 --cfg=tracing:yes --cfg=tracing/categorized:yes --cfg=triva/categorized:cat.plist
897 The first parameter activates the tracing subsystem, the second
898 tells it to trace host and link categorized utilization and the
899 third creates a graph configuration file to configure Triva when
900 analysing the resulting trace file.
902 - SMPI simulator and traces for a space/time view:
904 .. code-block:: shell
908 The `-trace` parameter for the smpirun script runs the simulation
909 with ``--cfg=tracing:yes --cfg=tracing/smpi:yes``. Check the
910 smpirun's `-help` parameter for additional tracing options.
912 Sometimes you might want to put additional information on the trace to
913 correctly identify them later, or to provide data that can be used to
914 reproduce an experiment. You have two ways to do that:
916 - Add a string on top of the trace file as comment:
918 .. code-block:: shell
920 --cfg=tracing/comment:my_simulation_identifier
922 - Add the contents of a textual file on top of the trace file as comment:
924 .. code-block:: shell
926 --cfg=tracing/comment-file:my_file_with_additional_information.txt
928 Please, use these two parameters (for comments) to make reproducible
929 simulations. For additional details about this and all tracing
930 options, check See the :ref:`tracing_tracing_options`.
935 .. _cfg=msg/debug-multiple-use:
940 **Option** ``msg/debug-multiple-use`` **Default:** off
942 Sometimes your application may try to send a task that is still being
943 executed somewhere else, making it impossible to send this task. However,
944 for debugging purposes, one may want to know what the other host is/was
945 doing. This option shows a backtrace of the other process.
950 The SMPI interface provides several specific configuration items.
951 These are uneasy to see since the code is usually launched through the
952 ``smiprun`` script directly.
954 .. _cfg=smpi/host-speed:
955 .. _cfg=smpi/cpu-threshold:
956 .. _cfg=smpi/simulate-computation:
958 Automatic Benchmarking of SMPI Code
959 ...................................
961 In SMPI, the sequential code is automatically benchmarked, and these
962 computations are automatically reported to the simulator. That is to
963 say that if you have a large computation between a ``MPI_Recv()`` and
964 a ``MPI_Send()``, SMPI will automatically benchmark the duration of
965 this code, and create an execution task within the simulator to take
966 this into account. For that, the actual duration is measured on the
967 host machine and then scaled to the power of the corresponding
968 simulated machine. The variable ``smpi/host-speed`` allows to specify
969 the computational speed of the host machine (in flop/s) to use when
970 scaling the execution times. It defaults to 20000, but you really want
971 to update it to get accurate simulation results.
973 When the code is constituted of numerous consecutive MPI calls, the
974 previous mechanism feeds the simulation kernel with numerous tiny
975 computations. The ``smpi/cpu-threshold`` item becomes handy when this
976 impacts badly the simulation performance. It specifies a threshold (in
977 seconds) below which the execution chunks are not reported to the
978 simulation kernel (default value: 1e-6).
980 .. note:: The option ``smpi/cpu-threshold`` ignores any computation
981 time spent below this threshold. SMPI does not consider the
982 `amount` of these computations; there is no offset for this. Hence,
983 a value that is too small, may lead to unreliable simulation
986 In some cases, however, one may wish to disable simulation of
987 application computation. This is the case when SMPI is used not to
988 simulate an MPI applications, but instead an MPI code that performs
989 "live replay" of another MPI app (e.g., ScalaTrace's replay tool,
990 various on-line simulators that run an app at scale). In this case the
991 computation of the replay/simulation logic should not be simulated by
992 SMPI. Instead, the replay tool or on-line simulator will issue
993 "computation events", which correspond to the actual MPI simulation
994 being replayed/simulated. At the moment, these computation events can
995 be simulated using SMPI by calling internal smpi_execute*() functions.
997 To disable the benchmarking/simulation of computation in the simulated
998 application, the variable ``smpi/simulate-computation`` should be set
999 to no. This option just ignores the timings in your simulation; it
1000 still executes the computations itself. If you want to stop SMPI from
1001 doing that, you should check the SMPI_SAMPLE macros, documented in
1002 Section :ref:`SMPI_adapting_speed`.
1004 +------------------------------------+-------------------------+-----------------------------+
1005 | Solution | Computations executed? | Computations simulated? |
1006 +====================================+=========================+=============================+
1007 | --cfg=smpi/simulate-computation:no | Yes | Never |
1008 +------------------------------------+-------------------------+-----------------------------+
1009 | --cfg=smpi/cpu-threshold:42 | Yes, in all cases | If it lasts over 42 seconds |
1010 +------------------------------------+-------------------------+-----------------------------+
1011 | SMPI_SAMPLE() macro | Only once per loop nest | Always |
1012 +------------------------------------+-------------------------+-----------------------------+
1014 .. _cfg=smpi/comp-adjustment-file:
1016 Slow-down or speed-up parts of your code
1017 ........................................
1019 **Option** ``smpi/comp-adjustment-file:`` **Default:** unset
1021 This option allows you to pass a file that contains two columns: The
1022 first column defines the section that will be subject to a speedup;
1023 the second column is the speedup. For instance:
1025 .. code-block:: shell
1027 "start:stop","ratio"
1028 "exchange_1.f:30:exchange_1.f:130",1.18244559422142
1030 The first line is the header - you must include it. The following
1031 line means that the code between two consecutive MPI calls on line 30
1032 in exchange_1.f and line 130 in exchange_1.f should receive a speedup
1033 of 1.18244559422142. The value for the second column is therefore a
1034 speedup, if it is larger than 1 and a slow-down if it is smaller
1035 than 1. Nothing will be changed if it is equal to 1.
1037 Of course, you can set any arbitrary filenames you want (so the start
1038 and end don't have to be in the same file), but be aware that this
1039 mechanism only supports `consecutive calls!`
1041 Please note that you must pass the ``-trace-call-location`` flag to
1042 smpicc or smpiff, respectively. This flag activates some internal
1043 macro definitions that help with obtaining the call location.
1045 .. _cfg=smpi/bw-factor:
1050 **Option** ``smpi/bw-factor``
1051 |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
1053 The possible throughput of network links is often dependent on the
1054 message sizes, as protocols may adapt to different message sizes. With
1055 this option, a series of message sizes and factors are given, helping
1056 the simulation to be more realistic. For instance, the current default
1057 value means that messages with size 65472 and more will get a total of
1058 MAX_BANDWIDTH*0.940694, messages of size 15424 to 65471 will get
1059 MAX_BANDWIDTH*0.697866 and so on (where MAX_BANDWIDTH denotes the
1060 bandwidth of the link).
1062 An experimental script to compute these factors is available online. See
1063 http://simgrid.gforge.inria.fr/contrib/smpi-calibration-doc.html
1064 http://simgrid.gforge.inria.fr/contrib/smpi-saturation-doc.html
1066 .. _cfg=smpi/display-timing:
1068 Reporting Simulation Time
1069 .........................
1071 **Option** ``smpi/display-timing`` **Default:** 0 (false)
1073 Most of the time, you run MPI code with SMPI to compute the time it
1074 would take to run it on a platform. But since the code is run through
1075 the ``smpirun`` script, you don't have any control on the launcher
1076 code, making it difficult to report the simulated time when the
1077 simulation ends. If you enable the ``smpi/display-timing`` item,
1078 ``smpirun`` will display this information when the simulation
1081 .. _cfg=smpi/keep-temps:
1083 Keeping temporary files after simulation
1084 ........................................
1086 **Option** ``smpi/keep-temps`` **default:** 0 (false)
1088 SMPI usually generates a lot of temporary files that are cleaned after
1089 use. This option request to preserve them, for example to debug or
1090 profile your code. Indeed, the binary files are removed very early
1091 under the dlopen privatization schema, which tend to fool the
1094 .. _cfg=smpi/lat-factor:
1099 **Option** ``smpi/lat-factor`` |br|
1100 **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
1102 The motivation and syntax for this option is identical to the motivation/syntax
1103 of :ref:`cfg=smpi/bw-factor`.
1105 There is an important difference, though: While smpi/bw-factor `reduces` the
1106 actual bandwidth (i.e., values between 0 and 1 are valid), latency factors
1107 increase the latency, i.e., values larger than or equal to 1 are valid here.
1109 .. _cfg=smpi/papi-events:
1111 Trace hardware counters with PAPI
1112 .................................
1114 **Option** ``smpi/papi-events`` **default:** unset
1116 When the PAPI support was compiled in SimGrid, this option takes the
1117 names of PAPI counters and adds their respective values to the trace
1118 files (See Section :ref:`tracing_tracing_options`).
1122 This feature currently requires superuser privileges, as registers
1123 are queried. Only use this feature with code you trust! Call
1124 smpirun for instance via ``smpirun -wrapper "sudo "
1125 <your-parameters>`` or run ``sudo sh -c "echo 0 >
1126 /proc/sys/kernel/perf_event_paranoid"`` In the later case, sudo
1127 will not be required.
1129 It is planned to make this feature available on a per-process (or per-thread?) basis.
1130 The first draft, however, just implements a "global" (i.e., for all processes) set
1131 of counters, the "default" set.
1133 .. code-block:: shell
1135 --cfg=smpi/papi-events:"default:PAPI_L3_LDM:PAPI_L2_LDM"
1137 .. _cfg=smpi/privatization:
1139 Automatic Privatization of Global Variables
1140 ...........................................
1142 **Option** ``smpi/privatization`` **default:** "dlopen" (when using smpirun)
1144 MPI executables are usually meant to be executed in separated
1145 processes, but SMPI is executed in only one process. Global variables
1146 from executables will be placed in the same memory zone and shared
1147 between processes, causing intricate bugs. Several options are
1148 possible to avoid this, as described in the main `SMPI publication
1149 <https://hal.inria.fr/hal-01415484>`_ and in the :ref:`SMPI
1150 documentation <SMPI_what_globals>`. SimGrid provides two ways of
1151 automatically privatizing the globals, and this option allows to
1152 choose between them.
1154 - **no** (default when not using smpirun): Do not automatically
1155 privatize variables. Pass ``-no-privatize`` to smpirun to disable
1157 - **dlopen** or **yes** (default when using smpirun): Link multiple
1158 times against the binary.
1159 - **mmap** (slower, but maybe somewhat more stable):
1160 Runtime automatic switching of the data segments.
1163 This configuration option cannot be set in your platform file. You can only
1164 pass it as an argument to smpirun.
1166 .. _cfg=smpi/privatize-libs:
1168 Automatic privatization of global variables inside external libraries
1169 .....................................................................
1171 **Option** ``smpi/privatize-libs`` **default:** unset
1173 **Linux/BSD only:** When using dlopen (default) privatization,
1174 privatize specific shared libraries with internal global variables, if
1175 they can't be linked statically. For example libgfortran is usually
1176 used for Fortran I/O and indexes in files can be mixed up.
1178 Multiple libraries can be given, semicolon separated.
1180 This configuration option can only use either full paths to libraries,
1181 or full names. Check with ldd the name of the library you want to
1184 .. code-block:: shell
1188 libgfortran.so.3 => /usr/lib/x86_64-linux-gnu/libgfortran.so.3 (0x00007fbb4d91b000)
1191 Then you can use ``--cfg=smpi/privatize-libs:libgfortran.so.3``
1192 or ``--cfg=smpi/privatize-libs:/usr/lib/x86_64-linux-gnu/libgfortran.so.3``,
1193 but not ``libgfortran`` nor ``libgfortran.so``.
1195 .. _cfg=smpi/send-is-detached-thresh:
1197 Simulating MPI detached send
1198 ............................
1200 **Option** ``smpi/send-is-detached-thresh`` **default:** 65536
1202 This threshold specifies the size in bytes under which the send will
1203 return immediately. This is different from the threshold detailed in
1204 :ref:`options_model_network_asyncsend` because the message is not
1205 effectively sent when the send is posted. SMPI still waits for the
1206 correspondant receive to be posted to perform the communication
1209 .. _cfg=smpi/coll-selector:
1211 Simulating MPI collective algorithms
1212 ....................................
1214 **Option** ``smpi/coll-selector`` **Possible values:** naive (default), ompi, mpich
1216 SMPI implements more than 100 different algorithms for MPI collective
1217 communication, to accurately simulate the behavior of most of the
1218 existing MPI libraries. The ``smpi/coll-selector`` item can be used to
1219 use the decision logic of either OpenMPI or MPICH libraries (by
1220 default SMPI uses naive version of collective operations).
1222 Each collective operation can be manually selected with a
1223 ``smpi/collective_name:algo_name``. Available algorithms are listed in
1224 :ref:`SMPI_use_colls`.
1226 .. TODO:: All available collective algorithms will be made available
1227 via the ``smpirun --help-coll`` command.
1229 .. _cfg=smpi/iprobe:
1231 Inject constant times for MPI_Iprobe
1232 ....................................
1234 **Option** ``smpi/iprobe`` **default:** 0.0001
1236 The behavior and motivation for this configuration option is identical
1237 with :ref:`smpi/test <cfg=smpi/test>`, but for the function
1240 .. _cfg=smpi/iprobe-cpu-usage:
1242 Reduce speed for iprobe calls
1243 .............................
1245 **Option** ``smpi/iprobe-cpu-usage`` **default:** 1 (no change)
1247 MPI_Iprobe calls can be heavily used in applications. To account
1248 correctly for the energy cores spend probing, it is necessary to
1249 reduce the load that these calls cause inside SimGrid.
1251 For instance, we measured a max power consumption of 220 W for a
1252 particular application but only 180 W while this application was
1253 probing. Hence, the correct factor that should be passed to this
1254 option would be 180/220 = 0.81.
1258 Inject constant times for MPI_Init
1259 ..................................
1261 **Option** ``smpi/init`` **default:** 0
1263 The behavior and motivation for this configuration option is identical
1264 with :ref:`smpi/test <cfg=smpi/test>`, but for the function ``MPI_Init()``.
1268 Inject constant times for MPI_Isend()
1269 .....................................
1271 **Option** ``smpi/ois``
1273 The behavior and motivation for this configuration option is identical
1274 with :ref:`smpi/os <cfg=smpi/os>`, but for the function ``MPI_Isend()``.
1278 Inject constant times for MPI_send()
1279 ....................................
1281 **Option** ``smpi/os``
1283 In several network models such as LogP, send (MPI_Send, MPI_Isend) and
1284 receive (MPI_Recv) operations incur costs (i.e., they consume CPU
1285 time). SMPI can factor these costs in as well, but the user has to
1286 configure SMPI accordingly as these values may vary by machine. This
1287 can be done by using ``smpi/os`` for MPI_Send operations; for MPI_Isend
1288 and MPI_Recv, use ``smpi/ois`` and ``smpi/or``, respectively. These work
1289 exactly as ``smpi/ois``.
1291 This item can consist of multiple sections; each section takes three
1292 values, for example ``1:3:2;10:5:1``. The sections are divided by ";"
1293 so this example contains two sections. Furthermore, each section
1294 consists of three values.
1296 1. The first value denotes the minimum size for this section to take effect;
1297 read it as "if message size is greater than this value (and other section has a larger
1298 first value that is also smaller than the message size), use this".
1299 In the first section above, this value is "1".
1301 2. The second value is the startup time; this is a constant value that will always
1302 be charged, no matter what the size of the message. In the first section above,
1305 3. The third value is the `per-byte` cost. That is, it is charged for every
1306 byte of the message (incurring cost messageSize*cost_per_byte)
1307 and hence accounts also for larger messages. In the first
1308 section of the example above, this value is "2".
1310 Now, SMPI always checks which section it should take for a given
1311 message; that is, if a message of size 11 is sent with the
1312 configuration of the example above, only the second section will be
1313 used, not the first, as the first value of the second section is
1314 closer to the message size. Hence, when ``smpi/os=1:3:2;10:5:1``, a
1315 message of size 11 incurs the following cost inside MPI_Send:
1316 ``5+11*1`` because 5 is the startup cost and 1 is the cost per byte.
1318 Note that the order of sections can be arbitrary; they will be ordered internally.
1322 Inject constant times for MPI_Recv()
1323 ....................................
1325 **Option** ``smpi/or``
1327 The behavior and motivation for this configuration option is identical
1328 with :ref:`smpi/os <cfg=smpi/os>`, but for the function ``MPI_Recv()``.
1331 .. _cfg=smpi/grow-injected-times:
1333 Inject constant times for MPI_Test
1334 ..................................
1336 **Option** ``smpi/test`` **default:** 0.0001
1338 By setting this option, you can control the amount of time a process
1339 sleeps when MPI_Test() is called; this is important, because SimGrid
1340 normally only advances the time while communication is happening and
1341 thus, MPI_Test will not add to the time, resulting in a deadlock if
1342 used as a break-condition as in the following example:
1347 MPI_Test(request, flag, status);
1351 To speed up execution, we use a counter to keep track on how often we
1352 already checked if the handle is now valid or not. Hence, we actually
1353 use counter*SLEEP_TIME, that is, the time MPI_Test() causes the
1354 process to sleep increases linearly with the number of previously
1355 failed tests. This behavior can be disabled by setting
1356 ``smpi/grow-injected-times`` to **no**. This will also disable this
1357 behavior for MPI_Iprobe.
1359 .. _cfg=smpi/shared-malloc:
1360 .. _cfg=smpi/shared-malloc-hugepage:
1365 **Option** ``smpi/shared-malloc`` **Possible values:** global (default), local
1367 If your simulation consumes too much memory, you may want to modify
1368 your code so that the working areas are shared by all MPI ranks. For
1369 example, in a bloc-cyclic matrix multiplication, you will only
1370 allocate one set of blocs, and every processes will share them.
1371 Naturally, this will lead to very wrong results, but this will save a
1372 lot of memory so this is still desirable for some studies. For more on
1373 the motivation for that feature, please refer to the `relevant section
1374 <https://simgrid.github.io/SMPI_CourseWare/topic_understanding_performance/matrixmultiplication>`_
1375 of the SMPI CourseWare (see Activity #2.2 of the pointed
1376 assignment). In practice, change the call to malloc() and free() into
1377 SMPI_SHARED_MALLOC() and SMPI_SHARED_FREE().
1379 SMPI provides two algorithms for this feature. The first one, called
1380 ``local``, allocates one bloc per call to SMPI_SHARED_MALLOC() in your
1381 code (each call location gets its own bloc) and this bloc is shared
1382 amongst all MPI ranks. This is implemented with the shm_* functions
1383 to create a new POSIX shared memory object (kept in RAM, in /dev/shm)
1384 for each shared bloc.
1386 With the ``global`` algorithm, each call to SMPI_SHARED_MALLOC()
1387 returns a new adress, but it only points to a shadow bloc: its memory
1388 area is mapped on a 1MiB file on disk. If the returned bloc is of size
1389 N MiB, then the same file is mapped N times to cover the whole bloc.
1390 At the end, no matter how many SMPI_SHARED_MALLOC you do, this will
1391 only consume 1 MiB in memory.
1393 You can disable this behavior and come back to regular mallocs (for
1394 example for debugging purposes) using @c "no" as a value.
1396 If you want to keep private some parts of the buffer, for instance if these
1397 parts are used by the application logic and should not be corrupted, you
1398 can use SMPI_PARTIAL_SHARED_MALLOC(size, offsets, offsets_count). Example:
1402 mem = SMPI_PARTIAL_SHARED_MALLOC(500, {27,42 , 100,200}, 2);
1404 This will allocate 500 bytes to mem, such that mem[27..41] and
1405 mem[100..199] are shared while other area remain private.
1407 Then, it can be deallocated by calling SMPI_SHARED_FREE(mem).
1409 When smpi/shared-malloc:global is used, the memory consumption problem
1410 is solved, but it may induce too much load on the kernel's pages table.
1411 In this case, you should use huge pages so that we create only one
1412 entry per Mb of malloced data instead of one entry per 4k.
1413 To activate this, you must mount a hugetlbfs on your system and allocate
1414 at least one huge page:
1416 .. code-block:: shell
1419 sudo mount none /home/huge -t hugetlbfs -o rw,mode=0777
1420 sudo sh -c 'echo 1 > /proc/sys/vm/nr_hugepages' # echo more if you need more
1422 Then, you can pass the option
1423 ``--cfg=smpi/shared-malloc-hugepage:/home/huge`` to smpirun to
1424 actually activate the huge page support in shared mallocs.
1428 Inject constant times for MPI_Wtime, gettimeofday and clock_gettime
1429 ...................................................................
1431 **Option** ``smpi/wtime`` **default:** 10 ns
1433 This option controls the amount of (simulated) time spent in calls to
1434 MPI_Wtime(), gettimeofday() and clock_gettime(). If you set this value
1435 to 0, the simulated clock is not advanced in these calls, which leads
1436 to issue if your application contains such a loop:
1440 while(MPI_Wtime() < some_time_bound) {
1441 /* some tests, with no communication nor computation */
1444 When the option smpi/wtime is set to 0, the time advances only on
1445 communications and computations, so the previous code results in an
1446 infinite loop: the current [simulated] time will never reach
1447 ``some_time_bound``. This infinite loop is avoided when that option
1448 is set to a small amount, as it is by default since SimGrid v3.21.
1450 Note that if your application does not contain any loop depending on
1451 the current time only, then setting this option to a non-zero value
1452 will slow down your simulations by a tiny bit: the simulation loop has
1453 to be broken and reset each time your code ask for the current time.
1454 If the simulation speed really matters to you, you can avoid this
1455 extra delay by setting smpi/wtime to 0.
1457 Other Configurations
1458 --------------------
1460 .. _cfg=clean-atexit:
1462 Cleanup at Termination
1463 ......................
1465 **Option** ``clean-atexit`` **default:** on
1467 If your code is segfaulting during its finalization, it may help to
1468 disable this option to request SimGrid to not attempt any cleanups at
1469 the end of the simulation. Since the Unix process is ending anyway,
1470 the operating system will wipe it all.
1477 **Option** ``path`` **default:** . (current dir)
1479 It is possible to specify a list of directories to search into for the
1480 trace files (see :ref:`pf_trace`) by using this configuration
1481 item. To add several directory to the path, set the configuration
1482 item several times, as in ``--cfg=path:toto --cfg=path:tutu``
1484 .. _cfg=simix/breakpoint:
1489 **Option** ``simix/breakpoint`` **default:** unset
1491 This configuration option sets a breakpoint: when the simulated clock
1492 reaches the given time, a SIGTRAP is raised. This can be used to stop
1493 the execution and get a backtrace with a debugger.
1495 It is also possible to set the breakpoint from inside the debugger, by
1496 writing in global variable simgrid::simix::breakpoint. For example,
1499 .. code-block:: shell
1501 set variable simgrid::simix::breakpoint = 3.1416
1503 .. _cfg=verbose-exit:
1508 **Option** ``verbose-exit`` **default:** on
1510 By default, when Ctrl-C is pressed, the status of all existing actors
1511 is displayed before exiting the simulation. This is very useful to
1512 debug your code, but it can reveal troublesome if you have many
1513 actors. Set this configuration item to **off** to disable this
1516 .. _cfg=exception/cutpath:
1518 Truncate local path from exception backtrace
1519 ............................................
1521 **Option** ``exception/cutpath`` **default:** off
1523 This configuration option is used to remove the path from the
1524 backtrace shown when an exception is thrown. This is mainly useful for
1525 the tests: the full file path makes the tests not reproducible because
1526 the path of source files depend of the build settings. That would
1527 break most of our tests as we keep comparing output.
1529 Logging Configuration
1530 ---------------------
1532 It can be done by using XBT. Go to :ref:`XBT_log` for more details.