SimGrid  3.15
Versatile Simulation of Distributed Systems
Under the Hood


  • Simulation Loop, LMM, sharing -> papers
  • Context Switching, privatization -> papers


S4U classes are designed to be user process interfaces to Maestro resources. We provide an uniform interface to them:

  • automatic reference count with intrusive smart pointers simgrid::s4u::FooPtr (also called simgrid::s4u::Foo::Ptr);
  • manual reference count with intrusive_ptr_add_ref(p), intrusive_ptr_release(p) (which is the interface used by boost::intrusive_ptr);
  • delegation of the operations to a opaque pimpl (which is the Maestro object);
  • the Maestro object and the corresponding S4U object have the same lifetime (and share the same reference count).

The ability to manipulate thge objects thought pointers and have the ability to use explicit reference count management is useful for creating C wrappers to the S4U and should play nicely with other language bindings (such as SWIG-based ones).

Some objects currently live for the whole duration of the simulation and do not have refertence counts. We still provide dummy intrusive_ptr_add_ref(p), intrusive_ptr_release(p) and FooPtr for consistency.

In many cases, we try to have a API which is consistent with the API or corresponding C++ standard classes. For example, the methods of simgrid::s4u::Mutex are based on std::mutex. This has several benefits:

  • we use a proven interface with a well defined and documented semantic;
  • the interface is easy to understand and remember for people used to the C++ standard interface;
  • we can use some standard C++ algorithms and helper classes with our types (simgrid::s4u::Mutex can be used with std::lock, std::unique_lock, etc.).

Example of simgrid::s4u::Actor:

class Actor {
// This is the corresponding maestro object:
friend simgrid::simix::Process;
simgrid::simix::Process* pimpl_ = nullptr;
Actor(simgrid::simix::Process* pimpl) : pimpl_(pimpl) {}
Actor(Actor const&) = delete;
Actor& operator=(Actor const&) = delete;
// Reference count is delegated to the S4u object:
friend void intrusive_ptr_add_ref(Actor* actor)
xbt_assert(actor != nullptr);
friend void intrusive_ptr_release(Actor* actor)
xbt_assert(actor != nullptr);
using Ptr = boost::intrusive_ptr<Actor>;
// Create processes:
static Ptr createActor(const char* name, s4u::Host *host, double killTime, std::function<void()> code);
// [...]
using ActorPtr = Actor::Ptr;

It uses the simgrid::simix::Process as a opaque pimple:

class Process {
std::atomic_int_fast32_t refcount_ { 1 };
// The lifetime of the s4u::Actor is bound to the lifetime of the Process:
Process() : actor_(this) {}
// Reference count:
friend void intrusive_ptr_add_ref(Process* process)
// Atomic operation! Do not split in two instructions!
auto previous = (process->refcount_)++;
xbt_assert(previous != 0);
(void) previous;
friend void intrusive_ptr_release(Process* process)
// Atomic operation! Do not split in two instructions!
auto count = --(process->refcount_);
if (count == 0)
delete process;
// [...]
smx_process_t SIMIX_process_ref(smx_process_t process)
if (process != nullptr)
return process;
void SIMIX_process_unref(smx_process_t process)
if (process != nullptr)

Asynchronous operations


The simgrid::kernel::Future class has been added to SimGrid as an abstraction to represent asynchronous operations in the SimGrid maestro. Its API is based on std::experimental::future from the C++ Extensions for Concurrency Technical Specification:

The expected way to work with simgrid::kernel::Future<T> is to add a completion handler/continuation:

// This code is executed in the maestro context, we cannot block for the result
// to be ready:
simgrid::kernel::Future<std::vector<char>> result = simgrid::kernel::readFile(file);
// Add a completion handler:
result.then([file](simgrid::kernel::Future<std::vector<char>> result) {
// At this point, the operation is complete and we can safely call .get():
try {
std::vector<char> data = result.get();
XBT_DEBUG("Finished reading file %s: length %zu", file.c_str(), data.size());
// If the operation failed, .get() throws an exception:
catch (std::runtime_error& e) {
XBT_ERROR("Could not read file %s", file.c_str());

The SimGrid kernel cannot block so calling .get() or .wait() on a simgrid::kernel::Future<T> which is not ready will deadlock. In practice, the simulator detects this and aborts after reporting an error.

In order to generate your own future, you might need to use a simgrid::kernel::Promise<T>. The promise is a one-way channel which can be used to set the result of an associated simgrid::kernel::Future<T> (with either .set_value() or .set_exception()):

simgrid::kernel::Future<void> kernel_wait_until(double date)
auto promise = std::make_shared<simgrid::kernel::Promise<void>>();
auto future = promise->get_future();
SIMIX_timer_set(date, [promise] {
return future;

Like the experimental futures, we support chaining .then() methods with automatic future unwrapping. You might want to look at some tutorial on C++ futures for more details and examples. Some operations of the proposed experimental futures are currently not implemented in our futures however such as .wait_for(), .wait_until(), shared_future, when_any().


Model Checker

The current implementation of the model-checker uses two distinct processes:

  • the SimGrid model-checker (simgrid-mc) itself lives in the parent process;
  • it spaws a child process for the SimGrid simulator/maestro and the simulated processes.

They communicate using a AF_UNIX SOCK_DGRAM socket and exchange messages defined in mc_protocol.h. The SIMGRID_MC_SOCKET_FD environment variable it set to the file descriptor of this socket in the child process.

The model-checker analyzes, saves and restores the state of the model-checked process using the following techniques:

  • the model-checker reads and writes in the model-checked address space;
  • the model-cheker ptrace()s the model-checked process and is thus able to know the state of the model-checked process if it crashes;
  • DWARF debug informations are used to unwind the stack and identify local variables;
  • a custom heap is enabled in the model-checked process which allows the model checker to know which chunks are allocated and which are freed.

Address space

The AddressSpace is a base class used for both the model-checked process and its snapshots and has methods to read in the corresponding address space:

  • the Process class is a subclass representing the model-checked process;
  • the Snapshot class is a subclass representing a snapshot of the process.

Additional helper class include:

  • Remote<T> is the result of reading a T in a remote AddressSpace. For trivial types (int, etc.), it is convertible t o T;
  • RemotePtr<T> represents the address of an object of type T in some remote AddressSpace (it could be an alias to Remote<T*>).


ELF is a standard executable file and dynamic libraries file format. DWARF is a standard for debug informations. Both are used on GNU/Linux systems and exploited by the model-checker to understand the model-checked process:

  • ObjectInformation represents the informations about a given ELF module (executable or shared-object);
  • Frame represents a subprogram scope (either a subprogram or a scope within the subprogram);
  • Type represents a type (eg. char*, int, std::string) and is referenced by variables (global, variables, parameters), functions (return type), and other types (type of a struct field, etc.);
  • LocationList and DwarfExpression are used to describe the location of variables.