MPI for Distributed Simulation

Parallel and distributed discrete event simulation allows the execution of a single simulation program on multiple processors. By splitting up the simulation into logical processes, LPs, each LP can be executed by a different processor. This simulation methodology enables very large-scale simulations by leveraging increased processing power and memory availability. In order to ensure proper execution of a distributed simulation, message passing between LPs is required. To support distributed simulation in ns-3, the standard Message Passing Interface (MPI) is used, along with a new distributed simulator class. Currently, dividing a simulation for distributed purposes in ns-3 can only occur across point-to-point links.

Current Implementation Details

During the course of a distributed simulation, many packets must cross simulator boundaries. In other words, a packet that originated on one LP is destined for a different LP, and in order to make this transition, a message containing the packet contents must be sent to the remote LP. Upon receiving this message, the remote LP can rebuild the packet and proceed as normal. The process of sending an receiving messages between LPs is handled easily by the new MPI interface in ns-3.

Along with simple message passing between LPs, a distributed simulator is used on each LP to determine which events to process. It is important to process events in time-stamped order to ensure proper simulation execution. If a LP receives a message containing an event from the past, clearly this is an issue, since this event could change other events which have already been executed. To address this problem, two conservative synchronization algorithm with lookahead are used in ns-3. For more information on different synchronization approaches and parallel and distributed simulation in general, please refer to “Parallel and Distributed Simulation Systems” by Richard Fujimoto.

The default parallel synchronization strategy implemented in the DistributedSimulatorImpl class is based on a globally synchronized algorithm using an MPI collective operation to synchronize simulation time across all LPs. A second synchronization strategy based on local communication and null messages is implemented in the NullMessageSimulatorImpl class, For the null message strategy the global all to all gather is not required; LPs only need to communication with LPs that have shared point-to-point links. The algorithm to use is controlled by which the ns-3 global value SimulatorImplementationType.

The best algorithm to use is dependent on the communication and event scheduling pattern for the application. In general, null message synchronization algorithms will scale better due to local communication scaling better than a global all-to-all gather that is required by DistributedSimulatorImpl. There are two known cases where the global synchronization performs better. The first is when most LPs have point-to-point link with most other LPs, in other words the LPs are nearly fully connected. In this case the null message algorithm will generate more message passing traffic than the all-to-all gather. A second case where the global all-to-all gather is more efficient is when there are long periods of simulation time when no events are occurring. The all-to-all gather algorithm is able to quickly determine then next event time globally. The nearest neighbor behavior of the null message algorithm will require more communications to propagate that knowledge; each LP is only aware of neighbor next event times.

Distributing the topology

Currently, the full topology is created on each rank, regardless of the individual node system ids. Only the applications are specific to a rank. For example, consider node 1 on LP 1 and node 2 on LP 2, with a traffic generator on node 1. Both node 1 and node 2 will be created on both LP1 and LP2; however, the traffic generator will only be installed on LP1. While this is not optimal for memory efficiency, it does simplify routing, since all current routing implementations in ns-3 will work with distributed simulation.

Running Distributed Simulations

Prerequisites

Ensure that MPI is installed, as well as mpic++. In Ubuntu repositories, these are openmpi-bin, openmpi-common, openmpi-doc, libopenmpi-dev. In Fedora, these are openmpi and openmpi-devel.

Note:

There is a conflict on some Fedora systems between libotf and openmpi. A possible “quick-fix” is to yum remove libotf before installing openmpi. This will remove conflict, but it will also remove emacs. Alternatively, these steps could be followed to resolve the conflict:

  1. Rename the tiny otfdump which emacs says it needs:

    $ mv /usr/bin/otfdump /usr/bin/otfdump.emacs-version
    
  2. Manually resolve openmpi dependencies:

    $ sudo yum install libgfortran libtorque numactl
    
  3. Download rpm packages:

    openmpi-1.3.1-1.fc11.i586.rpm
    openmpi-devel-1.3.1-1.fc11.i586.rpm
    openmpi-libs-1.3.1-1.fc11.i586.rpm
    openmpi-vt-1.3.1-1.fc11.i586.rpm
    

    from http://mirrors.kernel.org/fedora/releases/11/Everything/i386/os/Packages/

  4. Force the packages in:

    $ sudo rpm -ivh --force \
    openmpi-1.3.1-1.fc11.i586.rpm \
    openmpi-libs-1.3.1-1.fc11.i586.rpm \
    openmpi-devel-1.3.1-1.fc11.i586.rpm \
    openmpi-vt-1.3.1-1.fc11.i586.rpm
    

Also, it may be necessary to add the openmpi bin directory to PATH in order to execute mpic++ and mpirun from the command line. Alternatively, the full path to these executables can be used. Finally, if openmpi complains about the inability to open shared libraries, such as libmpi_cxx.so.0, it may be necessary to add the openmpi lib directory to LD_LIBRARY_PATH.

Here is an example of setting up PATH and LD_LIBRARY_PATH using a bash shell:

  • For a 32-bit Linux distribution:

    $ export PATH=$PATH:/usr/lib/openmpi/bin
    $ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib/openmpi/lib
    

For a 64-bit Linux distribution:

$ export PATH=$PATH:/usr/lib64/openmpi/bin
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib64/openmpi/lib

These lines can be added into ~/.bash_profile or ~/.bashrc to avoid having to retype them when a new shell is opened.

Note 2: There is a separate issue on recent Fedora distributions, which is that the libraries are built with AVX instructions. On older machines or some virtual machines, this results in an illegal instruction being thrown. This is not an ns-3 issue, a simple MPI test case will also fail. The AVX instructions are being called during initialization.

The symptom of this is that attempts to run an ns-3 MPI program will fail with the error: terminated with signal SIGILL. To check if this is the problem, run:

$ grep avx /proc/cpuinfo

and it will not return anything if AVX is not present.

If AVX is not supported, it is recommended to switch to a different MPI implementation such as MPICH:

$ dnf remove openmpi openmpi-devel
$ dnf install mpich mpich-devel environment-modules
$ module load mpi/mpich-x86_64

Building and Running Examples

If you already built ns-3 without MPI enabled, you must re-build:

$ ./ns3 distclean

Configure ns-3 with the –enable-mpi option:

$ ./ns3 configure -d debug --enable-examples --enable-tests --enable-mpi

Ensure that MPI is enabled by checking the optional features shown from the output of configure.

Next, build ns-3:

$ ./ns3

After building ns-3 with mpi enabled, the example programs are now ready to run with mpiexec. It is advised to avoid running ns3 directly with mpiexec; two options that should be more robust are to either use the –command-template way of running the mpiexec program, or to use ./ns3 shell and run the executables directly on the command line. Here are a few examples (from the root ns-3 directory):

$ ./ns3 run simple-distributed --command-template="mpiexec -np 2 %s"
$ ./ns3 run nms-p2p-nix-distributed --command-template="mpiexec -np 2 -machinefile mpihosts %s --nix=0"

An example using the null message synchronization algorithm:

$ ./ns3 run simple-distributed --command-template="mpiexec -np 2 %s --nullmsg"

The np switch is the number of logical processors to use. The machinefile switch is which machines to use. In order to use machinefile, the target file must exist (in this case mpihosts). This can simply contain something like:

localhost
localhost
localhost
...

Or if you have a cluster of machines, you can name them.

The other alternative to command-template is to use ./ns3 shell. Here are the equivalent examples to the above (assuming optimized build profile):

$ ./ns3 shell
$ cd build/src/mpi/examples
$ mpiexec -np 2 ns3-dev-simple-distributed-optimized
$ mpiexec -np 2 -machinefile mpihosts ns3-dev-nms-p2p-nix-distributed-optimized --nix=0
$ mpiexec -np 2 ns3-dev-simple-distributed-optimized --nullmsg

Setting synchronization algorithm to use

The global value SimulatorImplementationType is used to set the synchronization algorithm to use. This value must be set before the MpiInterface::Enable method is invoked if the default DistributedSimulatorImpl is not used. Here is an example code snippet showing how to add a command line argument to control the synchronization algorithm choice::

cmd.AddValue ("nullmsg", "Enable the use of null-message synchronization", nullmsg);
if(nullmsg)
  {
    GlobalValue::Bind ("SimulatorImplementationType",
                       StringValue ("ns3::NullMessageSimulatorImpl"));
  }
else
  {
    GlobalValue::Bind ("SimulatorImplementationType",
                       StringValue ("ns3::DistributedSimulatorImpl"));
  }

// Enable parallel simulator with the command line arguments
MpiInterface::Enable (&argc, &argv);

Creating custom topologies

The example programs in src/mpi/examples give a good idea of how to create different topologies for distributed simulation. The main points are assigning system ids to individual nodes, creating point-to-point links where the simulation should be divided, and installing applications only on the LP associated with the target node.

Assigning system ids to nodes is simple and can be handled two different ways. First, a NodeContainer can be used to create the nodes and assign system ids:

NodeContainer nodes;
nodes.Create (5, 1); // Creates 5 nodes with system id 1.

Alternatively, nodes can be created individually, assigned system ids, and added to a NodeContainer. This is useful if a NodeContainer holds nodes with different system ids:

NodeContainer nodes;
Ptr<Node> node1 = CreateObject<Node> (0); // Create node1 with system id 0
Ptr<Node> node2 = CreateObject<Node> (1); // Create node2 with system id 1
nodes.Add (node1);
nodes.Add (node2);

Next, where the simulation is divided is determined by the placement of point-to-point links. If a point-to-point link is created between two nodes with different system ids, a remote point-to-point link is created, as described in Current Implementation Details.

Finally, installing applications only on the LP associated with the target node is very important. For example, if a traffic generator is to be placed on node 0, which is on LP0, only LP0 should install this application. This is easily accomplished by first checking the simulator system id, and ensuring that it matches the system id of the target node before installing the application.

Tracing During Distributed Simulations

Depending on the system id (rank) of the simulator, the information traced will be different, since traffic originating on one simulator is not seen by another simulator until it reaches nodes specific to that simulator. The easiest way to keep track of different traces is to just name the trace files or pcaps differently, based on the system id of the simulator. For example, something like this should work well, assuming all of these local variables were previously defined:

if (MpiInterface::GetSystemId () == 0)
  {
    pointToPoint.EnablePcapAll ("distributed-rank0");
    phy.EnablePcap ("distributed-rank0", apDevices.Get (0));
    csma.EnablePcap ("distributed-rank0", csmaDevices.Get (0), true);
  }
else if (MpiInterface::GetSystemId () == 1)
  {
    pointToPoint.EnablePcapAll ("distributed-rank1");
    phy.EnablePcap ("distributed-rank1", apDevices.Get (0));
    csma.EnablePcap ("distributed-rank1", csmaDevices.Get (0), true);
  }