Ruby: Improve Change PerfectSwitch's wakeup function

Currently the wakeup function for the PerfectSwitch contains three loops -

loop on number of virtual networks
  loop on number of incoming links
	    loop till all messages for this (link, network) have been routed

With an 8 processor mesh network and Hammer protocol, about 11-12% of the
was observed to have been spent in this function, which is the highest
amongst all the functions. It was found that the innermost loop is executed
about 45 times per invocation of the wakeup function, when each invocation
of the wakeup function processes just about one message.

The patch tries to do away with the redundant executions of the innermost
loop. Counters have been added for each virtual network that record the
number of messages that need to be routed for that virtual network. The
inner loops are only executed when the number of messages for that particular
virtual network > 0. This does away with almost 80% of the executions of the
innermost loop. The function now consumes about 5-6% of the total execution
time.

1 files changed, 149 insertions, 132 deletions

diff --git a/src/mem/ruby/network/simple/PerfectSwitch.cc b/src/mem/ruby/network/simple/PerfectSwitch.cc
index 7229c724f..5c461c63f 100644
--- a/src/mem/ruby/network/simple/PerfectSwitch.cc
+++ b/src/mem/ruby/network/simple/PerfectSwitch.cc
@@ -54,6 +54,11 @@ PerfectSwitch::PerfectSwitch(SwitchID sid, SimpleNetwork* network_ptr)
     m_round_robin_start = 0;
     m_network_ptr = network_ptr;
     m_wakeups_wo_switch = 0;
+
+    for(int i = 0;i < m_virtual_networks;++i)
+    {
+        m_pending_message_count.push_back(0);
+    }
 }
 
 void
@@ -62,12 +67,15 @@ PerfectSwitch::addInPort(const vector<MessageBuffer*>& in)
     assert(in.size() == m_virtual_networks);
     NodeID port = m_in.size();
     m_in.push_back(in);
+
     for (int j = 0; j < m_virtual_networks; j++) {
         m_in[port][j]->setConsumer(this);
         string desc = csprintf("[Queue from port %s %s %s to PerfectSwitch]",
             NodeIDToString(m_switch_id), NodeIDToString(port),
             NodeIDToString(j));
         m_in[port][j]->setDescription(desc);
+        m_in[port][j]->setIncomingLink(port);
+        m_in[port][j]->setVnet(j);
     }
 }
 
@@ -154,161 +162,170 @@ PerfectSwitch::wakeup()
             m_round_robin_start = 0;
         }
 
-        // for all input ports, use round robin scheduling
-        for (int counter = 0; counter < m_in.size(); counter++) {
-            // Round robin scheduling
-            incoming++;
-            if (incoming >= m_in.size()) {
-                incoming = 0;
-            }
+        if(m_pending_message_count[vnet] > 0) {
+            // for all input ports, use round robin scheduling
+            for (int counter = 0; counter < m_in.size(); counter++) {
+                // Round robin scheduling
+                incoming++;
+                if (incoming >= m_in.size()) {
+                    incoming = 0;
+                }
 
-            // temporary vectors to store the routing results
-            vector<LinkID> output_links;
-            vector<NetDest> output_link_destinations;
-
-            // Is there a message waiting?
-            while (m_in[incoming][vnet]->isReady()) {
-                DPRINTF(RubyNetwork, "incoming: %d\n", incoming);
-
-                // Peek at message
-                msg_ptr = m_in[incoming][vnet]->peekMsgPtr();
-                net_msg_ptr = safe_cast<NetworkMessage*>(msg_ptr.get());
-                DPRINTF(RubyNetwork, "Message: %s\n", (*net_msg_ptr));
-
-                output_links.clear();
-                output_link_destinations.clear();
-                NetDest msg_dsts =
-                    net_msg_ptr->getInternalDestination();
-
-                // Unfortunately, the token-protocol sends some
-                // zero-destination messages, so this assert isn't valid
-                // assert(msg_dsts.count() > 0);
-
-                assert(m_link_order.size() == m_routing_table.size());
-                assert(m_link_order.size() == m_out.size());
-
-                if (m_network_ptr->getAdaptiveRouting()) {
-                    if (m_network_ptr->isVNetOrdered(vnet)) {
-                        // Don't adaptively route
-                        for (int out = 0; out < m_out.size(); out++) {
-                            m_link_order[out].m_link = out;
-                            m_link_order[out].m_value = 0;
-                        }
-                    } else {
-                        // Find how clogged each link is
-                        for (int out = 0; out < m_out.size(); out++) {
-                            int out_queue_length = 0;
-                            for (int v = 0; v < m_virtual_networks; v++) {
-                                out_queue_length += m_out[out][v]->getSize();
+                // temporary vectors to store the routing results
+                vector<LinkID> output_links;
+                vector<NetDest> output_link_destinations;
+
+                // Is there a message waiting?
+                while (m_in[incoming][vnet]->isReady()) {
+                    DPRINTF(RubyNetwork, "incoming: %d\n", incoming);
+
+                    // Peek at message
+                    msg_ptr = m_in[incoming][vnet]->peekMsgPtr();
+                    net_msg_ptr = safe_cast<NetworkMessage*>(msg_ptr.get());
+                    DPRINTF(RubyNetwork, "Message: %s\n", (*net_msg_ptr));
+
+                    output_links.clear();
+                    output_link_destinations.clear();
+                    NetDest msg_dsts =
+                        net_msg_ptr->getInternalDestination();
+
+                    // Unfortunately, the token-protocol sends some
+                    // zero-destination messages, so this assert isn't valid
+                    // assert(msg_dsts.count() > 0);
+
+                    assert(m_link_order.size() == m_routing_table.size());
+                    assert(m_link_order.size() == m_out.size());
+
+                    if (m_network_ptr->getAdaptiveRouting()) {
+                        if (m_network_ptr->isVNetOrdered(vnet)) {
+                            // Don't adaptively route
+                            for (int out = 0; out < m_out.size(); out++) {
+                                m_link_order[out].m_link = out;
+                                m_link_order[out].m_value = 0;
+                            }
+                        } else {
+                            // Find how clogged each link is
+                            for (int out = 0; out < m_out.size(); out++) {
+                                int out_queue_length = 0;
+                                for (int v = 0; v < m_virtual_networks; v++) {
+                                    out_queue_length += m_out[out][v]->getSize();
+                                }
+                                int value =
+                                    (out_queue_length << 8) | (random() & 0xff);
+                                m_link_order[out].m_link = out;
+                                m_link_order[out].m_value = value;
                             }
-                            int value =
-                                (out_queue_length << 8) | (random() & 0xff);
-                            m_link_order[out].m_link = out;
-                            m_link_order[out].m_value = value;
+
+                            // Look at the most empty link first
+                            sort(m_link_order.begin(), m_link_order.end());
                         }
+                    }
 
-                        // Look at the most empty link first
-                        sort(m_link_order.begin(), m_link_order.end());
+                    for (int i = 0; i < m_routing_table.size(); i++) {
+                        // pick the next link to look at
+                        int link = m_link_order[i].m_link;
+                        NetDest dst = m_routing_table[link];
+                        DPRINTF(RubyNetwork, "dst: %s\n", dst);
+
+                        if (!msg_dsts.intersectionIsNotEmpty(dst))
+                            continue;
+
+                        // Remember what link we're using
+                        output_links.push_back(link);
+
+                        // Need to remember which destinations need this
+                        // message in another vector.  This Set is the
+                        // intersection of the routing_table entry and the
+                        // current destination set.  The intersection must
+                        // not be empty, since we are inside "if"
+                        output_link_destinations.push_back(msg_dsts.AND(dst));
+
+                        // Next, we update the msg_destination not to
+                        // include those nodes that were already handled
+                        // by this link
+                        msg_dsts.removeNetDest(dst);
                     }
-                }
 
-                for (int i = 0; i < m_routing_table.size(); i++) {
-                    // pick the next link to look at
-                    int link = m_link_order[i].m_link;
-                    NetDest dst = m_routing_table[link];
-                    DPRINTF(RubyNetwork, "dst: %s\n", dst);
-
-                    if (!msg_dsts.intersectionIsNotEmpty(dst))
-                        continue;
-
-                    // Remember what link we're using
-                    output_links.push_back(link);
-
-                    // Need to remember which destinations need this
-                    // message in another vector.  This Set is the
-                    // intersection of the routing_table entry and the
-                    // current destination set.  The intersection must
-                    // not be empty, since we are inside "if"
-                    output_link_destinations.push_back(msg_dsts.AND(dst));
-
-                    // Next, we update the msg_destination not to
-                    // include those nodes that were already handled
-                    // by this link
-                    msg_dsts.removeNetDest(dst);
-                }
+                    assert(msg_dsts.count() == 0);
+                    //assert(output_links.size() > 0);
+
+                    // Check for resources - for all outgoing queues
+                    bool enough = true;
+                    for (int i = 0; i < output_links.size(); i++) {
+                        int outgoing = output_links[i];
+                        if (!m_out[outgoing][vnet]->areNSlotsAvailable(1))
+                            enough = false;
+                        DPRINTF(RubyNetwork, "Checking if node is blocked\n"
+                                "outgoing: %d, vnet: %d, enough: %d\n",
+                                outgoing, vnet, enough);
+                    }
 
-                assert(msg_dsts.count() == 0);
-                //assert(output_links.size() > 0);
-
-                // Check for resources - for all outgoing queues
-                bool enough = true;
-                for (int i = 0; i < output_links.size(); i++) {
-                    int outgoing = output_links[i];
-                    if (!m_out[outgoing][vnet]->areNSlotsAvailable(1))
-                        enough = false;
-                    DPRINTF(RubyNetwork, "Checking if node is blocked\n"
-                            "outgoing: %d, vnet: %d, enough: %d\n",
-                            outgoing, vnet, enough);
-                }
+                    // There were not enough resources
+                    if (!enough) {
+                        g_eventQueue_ptr->scheduleEvent(this, 1);
+                        DPRINTF(RubyNetwork, "Can't deliver message since a node "
+                                "is blocked\n"
+                                "Message: %s\n", (*net_msg_ptr));
+                        break; // go to next incoming port
+                    }
 
-                // There were not enough resources
-                if (!enough) {
-                    g_eventQueue_ptr->scheduleEvent(this, 1);
-                    DPRINTF(RubyNetwork, "Can't deliver message since a node "
-                            "is blocked\n"
-                            "Message: %s\n", (*net_msg_ptr));
-                    break; // go to next incoming port
-                }
+                    MsgPtr unmodified_msg_ptr;
 
-                MsgPtr unmodified_msg_ptr;
+                    if (output_links.size() > 1) {
+                        // If we are sending this message down more than
+                        // one link (size>1), we need to make a copy of
+                        // the message so each branch can have a different
+                        // internal destination we need to create an
+                        // unmodified MsgPtr because the MessageBuffer
+                        // enqueue func will modify the message
 
-                if (output_links.size() > 1) {
-                    // If we are sending this message down more than
-                    // one link (size>1), we need to make a copy of
-                    // the message so each branch can have a different
-                    // internal destination we need to create an
-                    // unmodified MsgPtr because the MessageBuffer
-                    // enqueue func will modify the message
+                        // This magic line creates a private copy of the
+                        // message
+                        unmodified_msg_ptr = msg_ptr->clone();
+                    }
 
-                    // This magic line creates a private copy of the
-                    // message
-                    unmodified_msg_ptr = msg_ptr->clone();
-                }
+                    // Enqueue it - for all outgoing queues
+                    for (int i=0; i<output_links.size(); i++) {
+                        int outgoing = output_links[i];
 
-                // Enqueue it - for all outgoing queues
-                for (int i=0; i<output_links.size(); i++) {
-                    int outgoing = output_links[i];
+                        if (i > 0) {
+                            // create a private copy of the unmodified
+                            // message
+                            msg_ptr = unmodified_msg_ptr->clone();
+                        }
 
-                    if (i > 0) {
-                        // create a private copy of the unmodified
-                        // message
-                        msg_ptr = unmodified_msg_ptr->clone();
-                    }
+                        // Change the internal destination set of the
+                        // message so it knows which destinations this
+                        // link is responsible for.
+                        net_msg_ptr = safe_cast<NetworkMessage*>(msg_ptr.get());
+                        net_msg_ptr->getInternalDestination() =
+                            output_link_destinations[i];
 
-                    // Change the internal destination set of the
-                    // message so it knows which destinations this
-                    // link is responsible for.
-                    net_msg_ptr = safe_cast<NetworkMessage*>(msg_ptr.get());
-                    net_msg_ptr->getInternalDestination() =
-                        output_link_destinations[i];
+                        // Enqeue msg
+                        DPRINTF(RubyNetwork, "Switch: %d enqueuing net msg from "
+                                "inport[%d][%d] to outport [%d][%d] time: %lld.\n",
+                                m_switch_id, incoming, vnet, outgoing, vnet,
+                                g_eventQueue_ptr->getTime());
 
-                    // Enqeue msg
-                    DPRINTF(RubyNetwork, "Switch: %d enqueuing net msg from "
-                            "inport[%d][%d] to outport [%d][%d] time: %lld.\n",
-                            m_switch_id, incoming, vnet, outgoing, vnet,
-                            g_eventQueue_ptr->getTime());
+                        m_out[outgoing][vnet]->enqueue(msg_ptr);
+                    }
 
-                    m_out[outgoing][vnet]->enqueue(msg_ptr);
+                    // Dequeue msg
+                    m_in[incoming][vnet]->pop();
+                    m_pending_message_count[vnet]--;
                 }
-
-                // Dequeue msg
-                m_in[incoming][vnet]->pop();
             }
         }
     }
 }
 
 void
+PerfectSwitch::storeEventInfo(int info)
+{
+    m_pending_message_count[info]++;
+}
+
+void
 PerfectSwitch::printStats(std::ostream& out) const
 {
     out << "PerfectSwitch printStats" << endl;