granted-time-window-mpi-interface.cc

Go to the documentation of this file.

/*
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 as
 * published by the Free Software Foundation;
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 *
 * Author: George Riley <riley@ece.gatech.edu>
 */
 
/**
 * \file
 * \ingroup mpi
 * Implementation of classes ns3::SentBuffer and ns3::GrantedTimeWindowMpiInterface.
 */
 
// This object contains static methods that provide an easy interface
// to the necessary MPI information.
 
#include "granted-time-window-mpi-interface.h"
 
#include "mpi-interface.h"
#include "mpi-receiver.h"
 
#include "ns3/log.h"
#include "ns3/net-device.h"
#include "ns3/node-list.h"
#include "ns3/node.h"
#include "ns3/nstime.h"
#include "ns3/simulator-impl.h"
#include "ns3/simulator.h"
 
#include <iomanip>
#include <iostream>
#include <list>
#include <mpi.h>
 
namespace ns3
{
 
NS_LOG_COMPONENT_DEFINE("GrantedTimeWindowMpiInterface");
 
NS_OBJECT_ENSURE_REGISTERED(GrantedTimeWindowMpiInterface);
 
SentBuffer::SentBuffer()
{
    m_buffer = nullptr;
    m_request = MPI_REQUEST_NULL;
}
 
SentBuffer::~SentBuffer()
{
    delete[] m_buffer;
}
 
uint8_t*
SentBuffer::GetBuffer()
{
    return m_buffer;
}
 
void
SentBuffer::SetBuffer(uint8_t* buffer)
{
    m_buffer = buffer;
}
 
MPI_Request*
SentBuffer::GetRequest()
{
    return &m_request;
}
 
uint32_t GrantedTimeWindowMpiInterface::g_sid = 0;
uint32_t GrantedTimeWindowMpiInterface::g_size = 1;
bool GrantedTimeWindowMpiInterface::g_enabled = false;
bool GrantedTimeWindowMpiInterface::g_mpiInitCalled = false;
uint32_t GrantedTimeWindowMpiInterface::g_rxCount = 0;
uint32_t GrantedTimeWindowMpiInterface::g_txCount = 0;
std::list<SentBuffer> GrantedTimeWindowMpiInterface::g_pendingTx;
 
MPI_Request* GrantedTimeWindowMpiInterface::g_requests;
char** GrantedTimeWindowMpiInterface::g_pRxBuffers;
MPI_Comm GrantedTimeWindowMpiInterface::g_communicator = MPI_COMM_WORLD;
bool GrantedTimeWindowMpiInterface::g_freeCommunicator = false;
 
TypeId
GrantedTimeWindowMpiInterface::GetTypeId()
{
    static TypeId tid =
        TypeId("ns3::GrantedTimeWindowMpiInterface").SetParent<Object>().SetGroupName("Mpi");
    return tid;
}
 
void
GrantedTimeWindowMpiInterface::Destroy()
{
    NS_LOG_FUNCTION(this);
 
    for (uint32_t i = 0; i < GetSize(); ++i)
    {
        delete[] g_pRxBuffers[i];
    }
    delete[] g_pRxBuffers;
    delete[] g_requests;
 
    g_pendingTx.clear();
}
 
uint32_t
GrantedTimeWindowMpiInterface::GetRxCount()
{
    NS_ASSERT(g_enabled);
    return g_rxCount;
}
 
uint32_t
GrantedTimeWindowMpiInterface::GetTxCount()
{
    NS_ASSERT(g_enabled);
    return g_txCount;
}
 
uint32_t
GrantedTimeWindowMpiInterface::GetSystemId()
{
    NS_ASSERT(g_enabled);
    return g_sid;
}
 
uint32_t
GrantedTimeWindowMpiInterface::GetSize()
{
    NS_ASSERT(g_enabled);
    return g_size;
}
 
bool
GrantedTimeWindowMpiInterface::IsEnabled()
{
    return g_enabled;
}
 
MPI_Comm
GrantedTimeWindowMpiInterface::GetCommunicator()
{
    NS_ASSERT(g_enabled);
    return g_communicator;
}
 
void
GrantedTimeWindowMpiInterface::Enable(int* pargc, char*** pargv)
{
    NS_LOG_FUNCTION(this << pargc << pargv);
 
    NS_ASSERT(g_enabled == false);
 
    // Initialize the MPI interface
    MPI_Init(pargc, pargv);
    Enable(MPI_COMM_WORLD);
    g_mpiInitCalled = true;
    g_enabled = true;
}
 
void
GrantedTimeWindowMpiInterface::Enable(MPI_Comm communicator)
{
    NS_LOG_FUNCTION(this);
 
    NS_ASSERT(g_enabled == false);
 
    // Standard MPI practice is to duplicate the communicator for
    // library to use.  Library communicates in isolated communication
    // context.
    MPI_Comm_dup(communicator, &g_communicator);
    g_freeCommunicator = true;
 
    MPI_Barrier(g_communicator);
 
    int mpiSystemId;
    int mpiSize;
    MPI_Comm_rank(g_communicator, &mpiSystemId);
    MPI_Comm_size(g_communicator, &mpiSize);
    g_sid = mpiSystemId;
    g_size = mpiSize;
 
    g_enabled = true;
    // Post a non-blocking receive for all peers
    g_pRxBuffers = new char*[g_size];
    g_requests = new MPI_Request[g_size];
    for (uint32_t i = 0; i < GetSize(); ++i)
    {
        g_pRxBuffers[i] = new char[MAX_MPI_MSG_SIZE];
        MPI_Irecv(g_pRxBuffers[i],
                  MAX_MPI_MSG_SIZE,
                  MPI_CHAR,
                  MPI_ANY_SOURCE,
                  0,
                  g_communicator,
                  &g_requests[i]);
    }
}
 
void
GrantedTimeWindowMpiInterface::SendPacket(Ptr<Packet> p,
                                          const Time& rxTime,
                                          uint32_t node,
                                          uint32_t dev)
{
    NS_LOG_FUNCTION(this << p << rxTime.GetTimeStep() << node << dev);
 
    SentBuffer sendBuf;
    g_pendingTx.push_back(sendBuf);
    auto i = g_pendingTx.rbegin(); // Points to the last element
 
    uint32_t serializedSize = p->GetSerializedSize();
    auto buffer = new uint8_t[serializedSize + 16];
    i->SetBuffer(buffer);
    // Add the time, dest node and dest device
    uint64_t t = rxTime.GetInteger();
    auto pTime = reinterpret_cast<uint64_t*>(buffer);
    *pTime++ = t;
    auto pData = reinterpret_cast<uint32_t*>(pTime);
    *pData++ = node;
    *pData++ = dev;
    // Serialize the packet
    p->Serialize(reinterpret_cast<uint8_t*>(pData), serializedSize);
 
    // Find the system id for the destination node
    Ptr<Node> destNode = NodeList::GetNode(node);
    uint32_t nodeSysId = destNode->GetSystemId();
 
    MPI_Isend(reinterpret_cast<void*>(i->GetBuffer()),
              serializedSize + 16,
              MPI_CHAR,
              nodeSysId,
              0,
              g_communicator,
              (i->GetRequest()));
    g_txCount++;
}
 
void
GrantedTimeWindowMpiInterface::ReceiveMessages()
{
    NS_LOG_FUNCTION_NOARGS();
 
    // Poll the non-block reads to see if data arrived
    while (true)
    {
        int flag = 0;
        int index = 0;
        MPI_Status status;
 
        MPI_Testany(MpiInterface::GetSize(), g_requests, &index, &flag, &status);
        if (!flag)
        {
            break; // No more messages
        }
        int count;
        MPI_Get_count(&status, MPI_CHAR, &count);
        g_rxCount++; // Count this receive
 
        // Get the meta data first
        auto pTime = reinterpret_cast<uint64_t*>(g_pRxBuffers[index]);
        uint64_t time = *pTime++;
        auto pData = reinterpret_cast<uint32_t*>(pTime);
        uint32_t node = *pData++;
        uint32_t dev = *pData++;
 
        Time rxTime(time);
 
        count -= sizeof(time) + sizeof(node) + sizeof(dev);
 
        Ptr<Packet> p = Create<Packet>(reinterpret_cast<uint8_t*>(pData), count, true);
 
        // Find the correct node/device to schedule receive event
        Ptr<Node> pNode = NodeList::GetNode(node);
        Ptr<MpiReceiver> pMpiRec = nullptr;
        uint32_t nDevices = pNode->GetNDevices();
        for (uint32_t i = 0; i < nDevices; ++i)
        {
            Ptr<NetDevice> pThisDev = pNode->GetDevice(i);
            if (pThisDev->GetIfIndex() == dev)
            {
                pMpiRec = pThisDev->GetObject<MpiReceiver>();
                break;
            }
        }
 
        NS_ASSERT(pNode && pMpiRec);
 
        // Schedule the rx event
        Simulator::ScheduleWithContext(pNode->GetId(),
                                       rxTime - Simulator::Now(),
                                       &MpiReceiver::Receive,
                                       pMpiRec,
                                       p);
 
        // Re-queue the next read
        MPI_Irecv(g_pRxBuffers[index],
                  MAX_MPI_MSG_SIZE,
                  MPI_CHAR,
                  MPI_ANY_SOURCE,
                  0,
                  g_communicator,
                  &g_requests[index]);
    }
}
 
void
GrantedTimeWindowMpiInterface::TestSendComplete()
{
    NS_LOG_FUNCTION_NOARGS();
 
    auto i = g_pendingTx.begin();
    while (i != g_pendingTx.end())
    {
        MPI_Status status;
        int flag = 0;
        MPI_Test(i->GetRequest(), &flag, &status);
        auto current = i; // Save current for erasing
        i++;              // Advance to next
        if (flag)
        { // This message is complete
            g_pendingTx.erase(current);
        }
    }
}
 
void
GrantedTimeWindowMpiInterface::Disable()
{
    NS_LOG_FUNCTION_NOARGS();
 
    if (g_freeCommunicator)
    {
        MPI_Comm_free(&g_communicator);
        g_freeCommunicator = false;
    }
 
    // ns-3 should MPI finalize only if ns-3 was used to initialize
    if (g_mpiInitCalled)
    {
        int flag = 0;
        MPI_Initialized(&flag);
        if (flag)
        {
            MPI_Finalize();
        }
        else
        {
            NS_FATAL_ERROR("Cannot disable MPI environment without Initializing it first");
        }
        g_mpiInitCalled = false;
    }
 
    g_enabled = false;
}
 
} // namespace ns3