Formal Verification and Model-checking
SimGrid can not only predict the performance of your application, but also assess its correctness through formal methods. Mc SimGrid is a full-featured model-checker that is embedded in the SimGrid framework. It can be used to formally verify safety properties on codes running on top of SimGrid, be it simple algorithms or full MPI applications.
Primer on formal methods
Formal methods are techniques leveraging mathematics to test and assess systems. They are routinely used to assess computer hardware, transportation systems or any other complex engineering process. Among these methods, model-checking is a technique to automatically prove that a given model verifies a given property by systematically checking all states of the model. The property and model are written in a mathematical language and fed to an automated tool called model checker. When the model does not verify the property, the model checker gives a counter-example that can be used to refine and improve the model. Conversely, if no counter-example can be found 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 large to be exhaustively explored, in which case the model-checker is not conclusive. Model checkers rely on so-called reduction techniques (based on symmetries and equivalence) to efficiently explore the system state.
Dynamic verification applies similar ideas to programs, without requiring a mathematical model of the system. Instead, the program 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 static analysis nor symbolic execution. Instead, the program is simply executed through all possible outcomes. On indecision points, a system checkpoint is taken, the first branch is executed exhaustively, and then the system is roll back to that point to explore the other branch.
Mc SimGrid targets distributed applications that interact through message passing or through synchronization mechanisms (mutex, barrier, etc). Since it does not explicitly observe memory accesses, Mc SimGrid cannot automatically detect race conditions in multithreaded programs. It can however be used to detect misuses of the synchronization functions, such as the ones resulting in deadlocks.
Mc SimGrid can be used to verify classical safety properties, but also communication determinism, a property that allows more efficient solutions toward fault-tolerance. It can alleviate the state space explosion problem through Dynamic Partial Ordering Reduction (DPOR) and state equality. Note that Mc SimGrid is currently less mature than other parts of the framework, but it improves every month. Please report any question and issue so that we can further improve it.
Limits
The main limit of Model Checking lies in the huge amount of scenarios to explore. SimGrid tries to explore only non-redundant scenarios thanks to classical reduction techniques (such as DPOR and stateful exploration) but the exploration may well never finish if you don’t carefully adapt your application to this mode.
A classical trap is that the Model Checker can only verify whether your application fits the properties provided, which is useless if you have a bug in your property. Remember also that one way for your application to never violate a given assertion is to not start at all, because of a stupid bug.
Another limit of this mode is that it does not use the performance models of the simulation mode. Time becomes discrete: You can say for example that the application took 42 steps to run, but there is no way to know how much time it took or the number of watts that were dissipated.
Finally, the model checker only explores the interleavings of computations and communications. Other factors such as thread execution interleaving are not considered by the SimGrid model checker.
The model checker may well miss existing issues, as it computes the possible outcomes from a given initial situation. There is no way to prove the correctness of your application in full generality with this tool.
Getting Mc SimGrid
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
simulations. It is also not activated in the Debian package, nor in the Python binary distributions. If you just plan to
experiment with Mc SimGrid, the easiest is to get the corresponding docker image. On the long term, you probably want to install it on
your machine: it works out of the box on Linux, Windows (with WSL2) and FreeBSD. Simply request it from cmake (cmake
-Denable_model-checking .
) and then compile SimGrid as usual. Unfortunately, Mc SimGrid does not work natively
on Mac OS X yet, so mac users should stick to the docker method for now.
$ docker image pull simgrid/tuto-mc
$ mkdir ~/tuto-mcsimgrid # or chose another directory to share between your computer and the docker container
$ docker run --user $UID:$GID -it --rm --name mcsimgrid --volume ~/tuto-mcsimgrid:/source/tutorial simgrid/tuto-mc bash
In the container, you have access to the following directories of interest:
/source/tutorial
: A view to the~/tuto-mcsimgrid
directory on your disk, out of the container. Edit the files you want from your computer and save them in~/tuto-mcsimgrid
; Compile and use them immediately within the container in/source/tutorial
./source/tuto-mc.git
: Files provided with this tutorial./source/simgrid.git
: Source code of SimGrid, pre-configured in MC mode. The framework is also installed in/usr
so the source code is only provided for your information.
Lab1: non-deterministic receive
Motivational example
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 between your host computer and the container.
# From within the container
$ cp -r /source/tuto-mc.git/* /source/tutorial/
$ cd /source/tutorial/
Several files should have appeared in the ~/tuto-mcsimgrid
directory of your computer.
This tutorial uses ndet-receive-s4u.cpp,
that uses the S4U interface of SimGrid, but we provide a
MPI version
if you prefer (see below for details on using the MPI version).
Code of ndet-receive-s4u.cpp
: click here to open
The provided code is rather simple: Three client
are launched with an integer from 1, 2, 3
as a parameter. These actors simply
send their parameter to a given mailbox. A server
receives 3 messages and assumes that the last received message is the number 3
.
If you compile and run it, it simply works:
$ cmake . && make
(output omitted)
$ ./ndet-receive-s4u small_platform.xml
[Jupiter:client:(2) 0.000000] [example/INFO] Sending 1
[Bourassa:client:(3) 0.000000] [example/INFO] Sending 2
[Ginette:client:(4) 0.000000] [example/INFO] Sending 3
[Jupiter:client:(2) 0.020516] [example/INFO] Sent!
[Bourassa:client:(3) 0.047027] [example/INFO] Sent!
[Ginette:client:(4) 0.064651] [example/INFO] Sent!
[Tremblay:server:(1) 0.064651] [example/INFO] OK
Running and understanding Mc SimGrid
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
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
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
all message orders. For that, you simply launch your simulation as a parameter to the simgrid-mc
binary as you would do with valgrind
:
$ simgrid-mc ./ndet-receive-s4u small_platform.xml
(some output ignored)
[Tremblay:server:(1) 0.000000] (...) Assertion value_got == 3 failed
(more output ignored)
If it fails with the error [root/CRITICAL] Could not wait for the model-checker.
, you need to explicitly add the PTRACE capability to
your docker. Restart your docker with the additional parameter --cap-add SYS_PTRACE
.
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 now read it together. It can be split in several parts:
First, you have some information coming from the application.
On top, you see the output of the application, but somewhat stuttering. This is exactly what happens: since Mc SimGrid is exploring all possible outcome of the code, the execution is sometimes rewind to explore another possible branch (here: another possible message ordering). Note also that all times are always 0 in the model checker, since the time is abstracted away in this mode.
[0.000000] [mc_safety/INFO] Check a safety property. Reduction is: dpor. [Jupiter:client:(2) 0.000000] [example/INFO] Sending 1 [Bourassa:client:(3) 0.000000] [example/INFO] Sending 2 [Ginette:client:(4) 0.000000] [example/INFO] Sending 3 [Jupiter:client:(2) 0.000000] [example/INFO] Sent! [Bourassa:client:(3) 0.000000] [example/INFO] Sent! [Tremblay:server:(1) 0.000000] [example/INFO] OK [Ginette:client:(4) 0.000000] [example/INFO] Sent! [Jupiter:client:(2) 0.000000] [example/INFO] Sent! [Bourassa:client:(3) 0.000000] [example/INFO] Sent! [Jupiter:client:(2) 0.000000] [example/INFO] Sent! [Bourassa:client:(3) 0.000000] [example/INFO] Sent! [Tremblay:server:(1) 0.000000] [example/INFO] OK [Ginette:client:(4) 0.000000] [example/INFO] Sent! [Jupiter:client:(2) 0.000000] [example/INFO] Sent! [Bourassa:client:(3) 0.000000] [example/INFO] Sent! [Jupiter:client:(2) 0.000000] [example/INFO] Sent!
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 the backtrace are useful, and some are omitted here.
[Tremblay:server:(1) 0.000000] /source/tutorial/ndet-receive-s4u.cpp:27: [root/CRITICAL] Assertion value_got == 3 failed Backtrace (displayed in actor server): -> 0# xbt_backtrace_display_current at /source/simgrid.git/src/xbt/backtrace.cpp:30 -> 1# server() at /source/tutorial/ndet-receive-s4u.cpp:27
After that comes a lot of information from the model-checker.
First, the error message itself. The
xbt_assert
in the code result in anabort()
in the application, that is interpreted as an application crash by the model-checker.[0.000000] [mc_ModelChecker/INFO] ************************** [0.000000] [mc_ModelChecker/INFO] ** CRASH IN THE PROGRAM ** [0.000000] [mc_ModelChecker/INFO] ************************** [0.000000] [mc_ModelChecker/INFO] From signal: Aborted [0.000000] [mc_ModelChecker/INFO] A core dump was generated by the system.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 calls we made (put/get) are split in atomic calls (iSend+Wait/iRecv+Wait), and all executions are interleaved. Also, Mc SimGrid reports the first faulty execution it finds: it may not be the shorter possible one.
[0.000000] [mc_ModelChecker/INFO] Counter-example execution trace: [0.000000] [mc_ModelChecker/INFO] 1: iRecv(mbox=0) [0.000000] [mc_ModelChecker/INFO] 2: iSend(mbox=0) [0.000000] [mc_ModelChecker/INFO] 1: WaitComm(from 2 to 1, mbox=0, no timeout) [0.000000] [mc_ModelChecker/INFO] 1: iRecv(mbox=0) [0.000000] [mc_ModelChecker/INFO] 2: WaitComm(from 2 to 1, mbox=0, no timeout) [0.000000] [mc_ModelChecker/INFO] 4: iSend(mbox=0) [0.000000] [mc_ModelChecker/INFO] 1: WaitComm(from 4 to 1, mbox=0, no timeout) [0.000000] [mc_ModelChecker/INFO] 1: iRecv(mbox=0) [0.000000] [mc_ModelChecker/INFO] 3: iSend(mbox=0) [0.000000] [mc_ModelChecker/INFO] 1: WaitComm(from 3 to 1, mbox=0, no timeout)Then, the execution path is given.
[0.000000] [mc_record/INFO] Path = 1;2;1;1;2;4;1;1;3;1
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 withoutsimgrid-mc
. This is becausesimgrid-mc
forbids to use a debugger such as gdb or valgrind on the code during the model-checking. For example, you can trigger the same execution in valgrind as follows:$ valgrind ./ndet-receive-s4u small_platform.xml --cfg=model-check/replay:'1;2;1;1;2;4;1;1;3;1' ==402== Memcheck, a memory error detector ==402== Copyright (C) 2002-2017, and GNU GPL'd, by Julian Seward et al. ==402== Using Valgrind-3.16.1 and LibVEX; rerun with -h for copyright info ==402== Command: ./ndet-receive-s4u small_platform.xml --cfg=model-check/replay:1;2;1;1;2;4;1;1;3;1 ==402== [0.000000] [xbt_cfg/INFO] Configuration change: Set 'model-check/replay' to '1;2;1;1;2;4;1;1;3;1' [0.000000] [mc_record/INFO] path=1;2;1;1;2;4;1;1;3;1 [Jupiter:client:(2) 0.000000] [example/INFO] Sending 1 [Bourassa:client:(3) 0.000000] [example/INFO] Sending 2 [Ginette:client:(4) 0.000000] [example/INFO] Sending 3 [Jupiter:client:(2) 0.000000] [example/INFO] Sent! [Tremblay:server:(1) 0.000000] /source/tutorial/ndet-receive-s4u.cpp:27: [root/CRITICAL] Assertion value_got == 3 failed (some output ignored) ==402== ==402== Process terminating with default action of signal 6 (SIGABRT): dumping core ==402== at 0x550FCE1: raise (raise.c:51) ==402== by 0x54F9536: abort (abort.c:79) ==402== by 0x10C696: server() (ndet-receive-s4u.cpp:27) (more valgrind output ignored)Then, Mc SimGrid displays some statistics about the amount of expanded states (the unique states in which your program was at a given point of the exploration), the visited states (the amount of times we visited another state – the same state may have been visited several times) and the amount of transitions.
[0.000000] [mc_dfs/INFO] DFS exploration ended. 22 unique states visited; 4 backtracks (56 transition replays, 30 states visited overall)
Finally, the application stack trace is displayed as the model-checker sees it. It should be the same as the one displayed from the application side, unless you found a bug our tools.
Using MPI instead of S4U
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
translation of ndet-receive-s4u.cpp
to MPI.
Code of ndet-receive-mpi.c
: click here to open
You can compile and run it on top of SimGrid as follows.
$ smpicc ndet-receive-mpi.c -o ndet-receive-mpi
$ smpirun -np 4 -platform small_platform.xml ndet-receive-mpi
Interestingly enough, the bug is triggered on my machine even without Mc SimGrid, because the simulator happens to use the execution path 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 should use Mc SimGrid to exhaustively explore the state space and trigger the bug in all cases.
$ smpirun -wrapper simgrid-mc -np 4 -platform small_platform.xml ndet-receive-mpi
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 can also trigger a given execution path out of the model-checker, for example to explore it with valgrind.
$ 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
Under the hood
If you want to run such analysis on your own code, out of the provided docker, there is some steps that you should take.
SimGrid should naturally be compiled with model-checking support. This requires a full set of dependencies (documented on the relevant page) and should not be activated by default as there is a small performance penalty for codes using a SimGrid with MC enabled (even if you don’t activate the model-checking at run time).
Also install
libboost-stacktrace-dev
to display nice backtraces from the application side (the one from the model-checking side is available in any case, but it contains less details).Mc SimGrid uses the
ptrace
system call to spy on the verified application. Some versions of Docker forbid the use of this call by default for security reason (it could be used to escape the docker containment with older versions of Linux). If you encounter this issue, you should either update your settings (the security issue was solved in later versions of Linux), or add--cap-add SYS_PTRACE
to the docker parameters, as hinted by the text.
Going further
This tutorial is not complete yet, as there is nothing on reduction techniques. For now, the best source of information on these topics is this old tutorial and that old presentation. But be warned that these source of information are very old: the liveness verification was removed in v3.35, even if these docs still mention it.