1 .. _usecase_modelchecking:
3 Formal Verification and Model-checking
4 ======================================
6 SimGrid can not only predict the performance of your application, but also assess its correctness through formal methods. Mc SimGrid is
7 a full-featured model-checker that is embedded in the SimGrid framework. It can be used to formally verify safety and liveness
8 properties on codes running on top of SimGrid, be it :ref:`simple algorithms <usecase_simalgo>` or :ref:`full MPI applications
11 Primer on formal methods
12 ------------------------
14 Formal methods are techniques leveraging mathematics to test and assess systems. They are routinely used to assess computer hardware,
15 transportation systems or any other complex engineering process. Among these methods, model-checking is a technique to automatically
16 prove that a given model verifies a given property by systematically checking all states of the model. The property and model are
17 written in a mathematical language and fed to an automated tool called model checker. When the model does not verify the property, the
18 model checker gives a counter-example that can be used to refine and improve the model. Conversely, if no counter-example can be found
19 after an exhaustive exploration of the model, we know that the property holds for the model. It may also happen that the model is too
20 large to be exhaustively explored, in which case the model-checker is not conclusive. Model checkers rely on so-called reduction
21 techniques (based on symmetries and equivalence) to efficiently explore the system state.
23 Dynamic verification applies similar ideas to programs, without requiring a mathematical model of the system. Instead, the program
24 itself is used as a model to verify against a property. Along these lines, Mc SimGrid is a stateful model checker: it does not leverage
25 static analysis nor symbolic execution. Instead, the program is simply executed through all possible outcomes. On indecision points, a
26 system checkpoint is taken, the first branch is executed exhaustively, and then the system is roll back to that point to explore the
29 Mc SimGrid targets distributed applications that interact through message passing or through synchronization mechanisms (mutex,
30 barrier, etc). Since it does not explicitly observe memory accesses, Mc SimGrid cannot automatically detect race conditions in
31 multithreaded programs. It can however be used to detect misuses of the synchronization functions, such as the ones resulting in
34 Mc SimGrid can be used to verify classical `safety and liveness properties <https://en.wikipedia.org/wiki/Linear_time_property>`_, but
35 also `communication determinism <https://hal.inria.fr/hal-01953167/document>`_, a property that allows more efficient solutions toward
36 fault-tolerance. It can alleviate the state space explosion problem through `Dynamic Partial Ordering Reduction (DPOR)
37 <https://en.wikipedia.org/wiki/Partial_order_reduction>`_ and `state equality <https://hal.inria.fr/hal-01900120/document>`_. Note that
38 Mc SimGrid is currently less mature than other parts of the framework, but it improves every month. Please report any question and
39 issue so that we can further improve it.
44 It is included in the SimGrid source code, but it is not compiled in by default as it induces a small performance overhead to the
45 simulations. It is also not activated in the Debian package, nor in the Python binary distributions. If you just plan to
46 experiment with Mc SimGrid, the easiest is to get the corresponding docker image. On the long term, you probably want to install it on
47 your machine: it works out of the box on Linux, Windows (with WSL2) and FreeBSD. Simply request it from cmake (``cmake
48 -Denable_model-checking .``) and then compile SimGrid :ref:`as usual <install_src>`. Unfortunately, Mc SimGrid does not work natively
49 on Mac OS X yet, so mac users should stick to the docker method for now.
51 .. code-block:: console
53 $ docker image pull simgrid/tuto-mc
54 $ mkdir ~/tuto-mcsimgrid # or chose another directory to share between your computer and the docker container
55 $ docker run --user $UID:$GID -it --rm --name mcsimgrid --volume ~/tuto-mcsimgrid:/source/tutorial simgrid/tuto-mc bash
57 In the container, you have access to the following directories of interest:
59 - ``/source/tutorial``: A view to the ``~/tuto-mcsimgrid`` directory on your disk, out of the container.
60 Edit the files you want from your computer and save them in ``~/tuto-mcsimgrid``;
61 Compile and use them immediately within the container in ``/source/tutorial``.
62 - ``/source/tuto-mc.git``: Files provided with this tutorial.
63 - ``/source/simgrid.git``: Source code of SimGrid, pre-configured in MC mode. The framework is also installed in ``/usr``
64 so the source code is only provided for your information.
66 Lab1: non-deterministic receive
67 -------------------------------
72 Let's go with a first example of a bugged program. Once in the container, copy all files from the tutorial into the directory shared
73 between your host computer and the container.
75 .. code-block:: console
77 # From within the container
78 $ cp -r /source/tuto-mc.git/* /source/tutorial/
79 $ cd /source/tutorial/
81 Several files should have appeared in the ``~/tuto-mcsimgrid`` directory of your computer.
82 This tutorial uses `ndet-receive-s4u.cpp <https://framagit.org/simgrid/tutorial-model-checking/-/blob/main/ndet-receive-s4u.cpp>`_,
83 that uses the :ref:`S4U interface <S4U_doc>` of SimGrid, but we provide a
84 `MPI version <https://framagit.org/simgrid/tutorial-model-checking/-/blob/main/ndet-receive-mpi.cpp>`_
85 if you prefer (see below for details on using the MPI version).
88 :header: Code of ``ndet-receive-s4u.cpp``: click here to open
90 You can also `view it online <https://framagit.org/simgrid/tutorial-model-checking/-/blob/main/ndet-receive-s4u.cpp>`_
92 .. literalinclude:: tuto_mc/ndet-receive-s4u.cpp
96 The provided code is rather simple: Three ``client`` are launched with an integer from ``1, 2, 3`` as a parameter. These actors simply
97 send their parameter to a given mailbox. A ``server`` receives 3 messages and assumes that the last received message is the number ``3``.
98 If you compile and run it, it simply works:
100 .. code-block:: console
104 $ ./ndet-receive-s4u small_platform.xml
105 [Jupiter:client:(2) 0.000000] [example/INFO] Sending 1
106 [Bourassa:client:(3) 0.000000] [example/INFO] Sending 2
107 [Ginette:client:(4) 0.000000] [example/INFO] Sending 3
108 [Jupiter:client:(2) 0.020516] [example/INFO] Sent!
109 [Bourassa:client:(3) 0.047027] [example/INFO] Sent!
110 [Ginette:client:(4) 0.064651] [example/INFO] Sent!
111 [Tremblay:server:(1) 0.064651] [example/INFO] OK
113 Running and understanding Mc SimGrid
114 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
116 If you think of it, that's weird that this code works: all the messages are sent at the exact same time (t=0), so there is no reason for
117 the message ``3`` to arrive last. Depending on the link speed, any order should be possible. To trigger the bug, you could fiddle with the
118 source code and/or the platform file, but this is not a method. Time to start Mc SimGrid, the SimGrid model checker, to exhaustively test
119 all message orders. For that, you simply launch your simulation as a parameter to the ``simgrid-mc`` binary as you would do with ``valgrind``:
121 .. code-block:: console
123 $ simgrid-mc ./ndet-receive-s4u small_platform.xml
124 (some output ignored)
125 [Tremblay:server:(1) 0.000000] (...) Assertion value_got == 3 failed
126 (more output ignored)
128 If it fails with the error ``[root/CRITICAL] Could not wait for the model-checker.``, you need to explicitly add the PTRACE capability to
129 your docker. Restart your docker with the additional parameter ``--cap-add SYS_PTRACE``.
131 At the end, it works: Mc SimGrid successfully triggers the bug. But the produced output is somewhat long and hairy. Don't worry, we will
132 now read it together. It can be split in several parts:
134 - First, you have some information coming from the application.
136 - On top, you see the output of the application, but somewhat stuttering. This is exactly what happens: since Mc SimGrid is exploring
137 all possible outcome of the code, the execution is sometimes rewind to explore another possible branch (here: another possible
138 message ordering). Note also that all times are always 0 in the model checker, since the time is abstracted away in this mode.
140 .. code-block:: console
142 [0.000000] [mc_safety/INFO] Check a safety property. Reduction is: dpor.
143 [Jupiter:client:(2) 0.000000] [example/INFO] Sending 1
144 [Bourassa:client:(3) 0.000000] [example/INFO] Sending 2
145 [Ginette:client:(4) 0.000000] [example/INFO] Sending 3
146 [Jupiter:client:(2) 0.000000] [example/INFO] Sent!
147 [Bourassa:client:(3) 0.000000] [example/INFO] Sent!
148 [Tremblay:server:(1) 0.000000] [example/INFO] OK
149 [Ginette:client:(4) 0.000000] [example/INFO] Sent!
150 [Jupiter:client:(2) 0.000000] [example/INFO] Sent!
151 [Bourassa:client:(3) 0.000000] [example/INFO] Sent!
152 [Jupiter:client:(2) 0.000000] [example/INFO] Sent!
153 [Bourassa:client:(3) 0.000000] [example/INFO] Sent!
154 [Tremblay:server:(1) 0.000000] [example/INFO] OK
155 [Ginette:client:(4) 0.000000] [example/INFO] Sent!
156 [Jupiter:client:(2) 0.000000] [example/INFO] Sent!
157 [Bourassa:client:(3) 0.000000] [example/INFO] Sent!
158 [Jupiter:client:(2) 0.000000] [example/INFO] Sent!
160 - Then you have the error message, along with a backtrace of the application at the point where the assertion fails. Not all the frames of
161 the backtrace are useful, and some are omitted here.
163 .. code-block:: console
165 [Tremblay:server:(1) 0.000000] /source/tutorial/ndet-receive-s4u.cpp:27: [root/CRITICAL] Assertion value_got == 3 failed
166 Backtrace (displayed in actor server):
167 -> 0# xbt_backtrace_display_current at /source/simgrid.git/src/xbt/backtrace.cpp:30
168 -> 1# server() at /source/tutorial/ndet-receive-s4u.cpp:27
170 - After that comes a lot of information from the model-checker.
172 - First, the error message itself. The ``xbt_assert`` in the code result in an ``abort()`` in the application, that is interpreted as an
173 application crash by the model-checker.
175 .. code-block:: console
177 [0.000000] [mc_ModelChecker/INFO] **************************
178 [0.000000] [mc_ModelChecker/INFO] ** CRASH IN THE PROGRAM **
179 [0.000000] [mc_ModelChecker/INFO] **************************
180 [0.000000] [mc_ModelChecker/INFO] From signal: Aborted
181 [0.000000] [mc_ModelChecker/INFO] A core dump was generated by the system.
183 - An execution trace is then given, listing all the actions that led to that faulty execution. This is not easy to read, because the API
184 calls we made (put/get) are split in atomic calls (iSend+Wait/iRecv+Wait), and all executions are interleaved. Also, Mc SimGrid
185 reports the first faulty execution it finds: it may not be the shorter possible one.
187 .. code-block:: console
189 [0.000000] [mc_ModelChecker/INFO] Counter-example execution trace:
190 [0.000000] [mc_ModelChecker/INFO] 1: iRecv(mbox=0)
191 [0.000000] [mc_ModelChecker/INFO] 2: iSend(mbox=0)
192 [0.000000] [mc_ModelChecker/INFO] 1: WaitComm(from 2 to 1, mbox=0, no timeout)
193 [0.000000] [mc_ModelChecker/INFO] 1: iRecv(mbox=0)
194 [0.000000] [mc_ModelChecker/INFO] 2: WaitComm(from 2 to 1, mbox=0, no timeout)
195 [0.000000] [mc_ModelChecker/INFO] 4: iSend(mbox=0)
196 [0.000000] [mc_ModelChecker/INFO] 1: WaitComm(from 4 to 1, mbox=0, no timeout)
197 [0.000000] [mc_ModelChecker/INFO] 1: iRecv(mbox=0)
198 [0.000000] [mc_ModelChecker/INFO] 3: iSend(mbox=0)
199 [0.000000] [mc_ModelChecker/INFO] 1: WaitComm(from 3 to 1, mbox=0, no timeout)
201 - Then, the execution path is given.
203 .. code-block:: console
205 [0.000000] [mc_record/INFO] Path = 1;2;1;1;2;4;1;1;3;1
207 This is the magical string (here: ``1;2;1;1;2;4;1;1;3;1``) that you should pass to your simulator to follow the same execution path
208 without ``simgrid-mc``. This is because ``simgrid-mc`` forbids to use a debugger such as gdb or valgrind on the code during the
209 model-checking. For example, you can trigger the same execution in valgrind as follows:
211 .. code-block:: console
213 $ valgrind ./ndet-receive-s4u small_platform.xml --cfg=model-check/replay:'1;2;1;1;2;4;1;1;3;1'
214 ==402== Memcheck, a memory error detector
215 ==402== Copyright (C) 2002-2017, and GNU GPL'd, by Julian Seward et al.
216 ==402== Using Valgrind-3.16.1 and LibVEX; rerun with -h for copyright info
217 ==402== Command: ./ndet-receive-s4u small_platform.xml --cfg=model-check/replay:1;2;1;1;2;4;1;1;3;1
219 [0.000000] [xbt_cfg/INFO] Configuration change: Set 'model-check/replay' to '1;2;1;1;2;4;1;1;3;1'
220 [0.000000] [mc_record/INFO] path=1;2;1;1;2;4;1;1;3;1
221 [Jupiter:client:(2) 0.000000] [example/INFO] Sending 1
222 [Bourassa:client:(3) 0.000000] [example/INFO] Sending 2
223 [Ginette:client:(4) 0.000000] [example/INFO] Sending 3
224 [Jupiter:client:(2) 0.000000] [example/INFO] Sent!
225 [Tremblay:server:(1) 0.000000] /source/tutorial/ndet-receive-s4u.cpp:27: [root/CRITICAL] Assertion value_got == 3 failed
226 (some output ignored)
228 ==402== Process terminating with default action of signal 6 (SIGABRT): dumping core
229 ==402== at 0x550FCE1: raise (raise.c:51)
230 ==402== by 0x54F9536: abort (abort.c:79)
231 ==402== by 0x10C696: server() (ndet-receive-s4u.cpp:27)
232 (more valgrind output ignored)
234 - Then, Mc SimGrid displays some statistics about the amount of expanded states (the unique states in which your program was at a given
235 point of the exploration), the visited states (the amount of times we visited another state -- the same state may have been visited
236 several times) and the amount of transitions.
238 .. code-block:: console
240 [0.000000] [mc_dfs/INFO] DFS exploration ended. 22 unique states visited; 4 backtracks (56 transition replays, 30 states visited overall)
242 - Finally, the application stack trace is displayed as the model-checker sees it. It should be the same as the one displayed from the
243 application side, unless you found a bug our tools.
245 Using MPI instead of S4U
246 ^^^^^^^^^^^^^^^^^^^^^^^^
248 If you prefer, you can use MPI instead of the SimGrid-specific interface. Inspect the provided ``ndet-receive-mpi.c`` file: that's just a
249 translation of ``ndet-receive-s4u.cpp`` to MPI.
252 :header: Code of ``ndet-receive-mpi.c``: click here to open
254 You can also `view it online <https://framagit.org/simgrid/tutorial-model-checking/-/blob/main/ndet-receive-mpi.c>`_.
256 .. literalinclude:: tuto_mc/ndet-receive-mpi.c
260 You can compile and run it on top of SimGrid as follows.
262 .. code-block:: console
264 $ smpicc ndet-receive-mpi.c -o ndet-receive-mpi
265 $ smpirun -np 4 -platform small_platform.xml ndet-receive-mpi
267 Interestingly enough, the bug is triggered on my machine even without Mc SimGrid, because the simulator happens to use the execution path
268 leading to it. It may not be the case on your machine, as this depends on the iteration order of an unsorted collection. Instead, we
269 should use Mc SimGrid to exhaustively explore the state space and trigger the bug in all cases.
271 .. code-block:: console
273 $ smpirun -wrapper simgrid-mc -np 4 -platform small_platform.xml ndet-receive-mpi
275 The produced output is then very similar to the one you get with S4U, even if the exact execution path leading to the bug may differs. You
276 can also trigger a given execution path out of the model-checker, for example to explore it with valgrind.
278 .. code-block:: console
280 $ smpirun -wrapper valgrind -np 4 -platform small_platform.xml --cfg=model-check/replay:'1;2;1;1;4;1;1;3;1' ndet-receive-mpi
285 If you want to run such analysis on your own code, out of the provided docker, there is some steps that you should take.
287 - SimGrid should naturally :ref:`be compiled <install_src>` with model-checking support. This requires a full set of dependencies
288 (documented on the :ref:`relevant page <install_src>`) and should not be activated by default as there is a small performance penalty for
289 codes using a SimGrid with MC enabled (even if you don't activate the model-checking at run time).
290 - You should pass some specific flags to the linker when compiling your application: ``-Wl,-znorelro -Wl,-znoseparate-code`` In the
291 docker, the provided CMakeLists.txt provides them for you when compiling the provided code. ``smpicc`` and friends also add this
292 parameter automatically.
293 - If you get error messages complaining about the Dwarf version used, try adding ``-gdwarf-4`` to you CFLAGS and CXXFLAGS.
294 If you find a situation where this flag is needed in ``smpicc``, please report this issue.
295 - Also install ``libboost-stacktrace-dev`` to display nice backtraces from the application side (the one from the model-checking side is
296 available in any case, but it contains less details).
297 - Mc SimGrid uses the ``ptrace`` system call to spy on the verified application. Some versions of Docker forbid the use of this call by
298 default for security reason (it could be used to escape the docker containment with older versions of Linux). If you encounter this
299 issue, you should either update your settings (the security issue was solved in later versions of Linux), or add ``--cap-add
300 SYS_PTRACE`` to the docker parameters, as hinted by the text.
305 This tutorial is not complete yet, as there is nothing on reduction
306 techniques nor on liveness properties. For now, the best source of
307 information on these topics is `this old tutorial
308 <https://simgrid.org/tutorials/simgrid-mc-101.pdf>`_ and `that old
310 <http://people.irisa.fr/Martin.Quinson/blog/2018/0123/McSimGrid-Boston.pdf>`_.