-
/*
* Copyright (c) 1999-2008 Mark D. Hill and David A. Wood
* All rights reserved.
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
-/*
- * $Id$
- *
- * Description: see Throttle.h
- *
- */
+#include <cassert>
+#include "base/cast.hh"
+#include "base/cprintf.hh"
+#include "debug/RubyNetwork.hh"
#include "mem/ruby/network/simple/Throttle.hh"
-#include "mem/ruby/buffers/MessageBuffer.hh"
+#include "mem/ruby/network/MessageBuffer.hh"
#include "mem/ruby/network/Network.hh"
-#include "mem/ruby/system/System.hh"
#include "mem/ruby/slicc_interface/NetworkMessage.hh"
-#include "mem/protocol/Protocol.hh"
+#include "mem/ruby/system/System.hh"
+
+using namespace std;
-const int HIGH_RANGE = 256;
-const int ADJUST_INTERVAL = 50000;
const int MESSAGE_SIZE_MULTIPLIER = 1000;
//const int BROADCAST_SCALING = 4; // Have a 16p system act like a 64p systems
const int BROADCAST_SCALING = 1;
static int network_message_to_size(NetworkMessage* net_msg_ptr);
-extern std::ostream * debug_cout_ptr;
-
-Throttle::Throttle(int sID, NodeID node, int link_latency, int link_bandwidth_multiplier)
+Throttle::Throttle(int sID, NodeID node, Cycles link_latency,
+ int link_bandwidth_multiplier, int endpoint_bandwidth,
+ ClockedObject *em)
+ : Consumer(em)
{
- init(node, link_latency, link_bandwidth_multiplier);
- m_sID = sID;
+ init(node, link_latency, link_bandwidth_multiplier, endpoint_bandwidth);
+ m_sID = sID;
}
-Throttle::Throttle(NodeID node, int link_latency, int link_bandwidth_multiplier)
+Throttle::Throttle(NodeID node, Cycles link_latency,
+ int link_bandwidth_multiplier, int endpoint_bandwidth,
+ ClockedObject *em)
+ : Consumer(em)
{
- init(node, link_latency, link_bandwidth_multiplier);
- m_sID = 0;
+ init(node, link_latency, link_bandwidth_multiplier, endpoint_bandwidth);
+ m_sID = 0;
}
-void Throttle::init(NodeID node, int link_latency, int link_bandwidth_multiplier)
+void
+Throttle::init(NodeID node, Cycles link_latency,
+ int link_bandwidth_multiplier, int endpoint_bandwidth)
{
- m_node = node;
- m_vnets = 0;
-
- ASSERT(link_bandwidth_multiplier > 0);
- m_link_bandwidth_multiplier = link_bandwidth_multiplier;
- m_link_latency = link_latency;
-
- m_bash_counter = HIGH_RANGE;
- m_bandwidth_since_sample = 0;
- m_last_bandwidth_sample = 0;
- m_wakeups_wo_switch = 0;
- clearStats();
-}
+ m_node = node;
+ assert(link_bandwidth_multiplier > 0);
+ m_link_bandwidth_multiplier = link_bandwidth_multiplier;
-void Throttle::clear()
-{
- for (int counter = 0; counter < m_vnets; counter++) {
- m_in[counter]->clear();
- m_out[counter]->clear();
- }
+ m_link_latency = link_latency;
+ m_endpoint_bandwidth = endpoint_bandwidth;
+
+ m_wakeups_wo_switch = 0;
+ m_link_utilization_proxy = 0;
}
-void Throttle::addLinks(const Vector<MessageBuffer*>& in_vec, const Vector<MessageBuffer*>& out_vec)
+void
+Throttle::addLinks(const map<int, MessageBuffer*>& in_vec,
+ const map<int, MessageBuffer*>& out_vec)
{
- assert(in_vec.size() == out_vec.size());
- for (int i=0; i<in_vec.size(); i++) {
- addVirtualNetwork(in_vec[i], out_vec[i]);
- }
-
- m_message_counters.setSize(MessageSizeType_NUM);
- for (int i=0; i<MessageSizeType_NUM; i++) {
- m_message_counters[i].setSize(in_vec.size());
- for (int j=0; j<m_message_counters[i].size(); j++) {
- m_message_counters[i][j] = 0;
- }
- }
+ assert(in_vec.size() == out_vec.size());
- if (g_PRINT_TOPOLOGY) {
- m_out_link_vec.insertAtBottom(out_vec);
- }
-}
+ for (auto& it : in_vec) {
+ int vnet = it.first;
-void Throttle::addVirtualNetwork(MessageBuffer* in_ptr, MessageBuffer* out_ptr)
-{
- m_units_remaining.insertAtBottom(0);
- m_in.insertAtBottom(in_ptr);
- m_out.insertAtBottom(out_ptr);
-
- // Set consumer and description
- m_in[m_vnets]->setConsumer(this);
- string desc = "[Queue to Throttle " + NodeIDToString(m_sID) + " " + NodeIDToString(m_node) + "]";
- m_in[m_vnets]->setDescription(desc);
- m_vnets++;
-}
+ auto jt = out_vec.find(vnet);
+ assert(jt != out_vec.end());
-void Throttle::wakeup()
-{
- // Limits the number of message sent to a limited number of bytes/cycle.
- assert(getLinkBandwidth() > 0);
- int bw_remaining = getLinkBandwidth();
-
- // Give the highest numbered link priority most of the time
- m_wakeups_wo_switch++;
- int highest_prio_vnet = m_vnets-1;
- int lowest_prio_vnet = 0;
- int counter = 1;
- bool schedule_wakeup = false;
-
- // invert priorities to avoid starvation seen in the component network
- if (m_wakeups_wo_switch > PRIORITY_SWITCH_LIMIT) {
- m_wakeups_wo_switch = 0;
- highest_prio_vnet = 0;
- lowest_prio_vnet = m_vnets-1;
- counter = -1;
- }
+ MessageBuffer *in_ptr = it.second;
+ MessageBuffer *out_ptr = (*jt).second;
- for (int vnet = highest_prio_vnet; (vnet*counter) >= (counter*lowest_prio_vnet); vnet -= counter) {
+ m_in[vnet] = in_ptr;
+ m_out[vnet] = out_ptr;
+ m_units_remaining[vnet] = 0;
+
+ // Set consumer and description
+ in_ptr->setConsumer(this);
+ string desc = "[Queue to Throttle " + to_string(m_sID) + " " +
+ to_string(m_node) + "]";
+ in_ptr->setDescription(desc);
+ }
+}
- assert(m_out[vnet] != NULL);
- assert(m_in[vnet] != NULL);
+void
+Throttle::operateVnet(int vnet, int &bw_remaining, bool &schedule_wakeup,
+ MessageBuffer *in, MessageBuffer *out)
+{
+ assert(out != NULL);
+ assert(in != NULL);
assert(m_units_remaining[vnet] >= 0);
- while ((bw_remaining > 0) && ((m_in[vnet]->isReady()) || (m_units_remaining[vnet] > 0)) && m_out[vnet]->areNSlotsAvailable(1)) {
+ while (bw_remaining > 0 && (in->isReady() || m_units_remaining[vnet] > 0) &&
+ out->areNSlotsAvailable(1)) {
+
+ // See if we are done transferring the previous message on
+ // this virtual network
+ if (m_units_remaining[vnet] == 0 && in->isReady()) {
+ // Find the size of the message we are moving
+ MsgPtr msg_ptr = in->peekMsgPtr();
+ NetworkMessage* net_msg_ptr =
+ safe_cast<NetworkMessage*>(msg_ptr.get());
+ m_units_remaining[vnet] +=
+ network_message_to_size(net_msg_ptr);
+
+ DPRINTF(RubyNetwork, "throttle: %d my bw %d bw spent "
+ "enqueueing net msg %d time: %lld.\n",
+ m_node, getLinkBandwidth(), m_units_remaining[vnet],
+ g_system_ptr->curCycle());
+
+ // Move the message
+ in->dequeue();
+ out->enqueue(msg_ptr, m_link_latency);
+
+ // Count the message
+ m_msg_counts[net_msg_ptr->getMessageSize()][vnet]++;
+ DPRINTF(RubyNetwork, "%s\n", *out);
+ }
+
+ // Calculate the amount of bandwidth we spent on this message
+ int diff = m_units_remaining[vnet] - bw_remaining;
+ m_units_remaining[vnet] = max(0, diff);
+ bw_remaining = max(0, -diff);
+ }
- // See if we are done transferring the previous message on this virtual network
- if (m_units_remaining[vnet] == 0 && m_in[vnet]->isReady()) {
+ if (bw_remaining > 0 && (in->isReady() || m_units_remaining[vnet] > 0) &&
+ !out->areNSlotsAvailable(1)) {
+ DPRINTF(RubyNetwork, "vnet: %d", vnet);
- // Find the size of the message we are moving
- MsgPtr msg_ptr = m_in[vnet]->peekMsgPtr();
- NetworkMessage* net_msg_ptr = dynamic_cast<NetworkMessage*>(msg_ptr.ref());
- m_units_remaining[vnet] += network_message_to_size(net_msg_ptr);
+ // schedule me to wakeup again because I'm waiting for my
+ // output queue to become available
+ schedule_wakeup = true;
+ }
+}
- DEBUG_NEWLINE(NETWORK_COMP,HighPrio);
- DEBUG_MSG(NETWORK_COMP,HighPrio,"throttle: " + int_to_string(m_node)
- + " my bw " + int_to_string(getLinkBandwidth())
- + " bw spent enqueueing net msg " + int_to_string(m_units_remaining[vnet])
- + " time: " + int_to_string(g_eventQueue_ptr->getTime()) + ".");
+void
+Throttle::wakeup()
+{
+ // Limits the number of message sent to a limited number of bytes/cycle.
+ assert(getLinkBandwidth() > 0);
+ int bw_remaining = getLinkBandwidth();
- // Move the message
- m_out[vnet]->enqueue(m_in[vnet]->peekMsgPtr(), m_link_latency);
- m_in[vnet]->pop();
+ m_wakeups_wo_switch++;
+ bool schedule_wakeup = false;
- // Count the message
- m_message_counters[net_msg_ptr->getMessageSize()][vnet]++;
+ // variable for deciding the direction in which to iterate
+ bool iteration_direction = false;
- DEBUG_MSG(NETWORK_COMP,LowPrio,*m_out[vnet]);
- DEBUG_NEWLINE(NETWORK_COMP,HighPrio);
- }
- // Calculate the amount of bandwidth we spent on this message
- int diff = m_units_remaining[vnet] - bw_remaining;
- m_units_remaining[vnet] = max(0, diff);
- bw_remaining = max(0, -diff);
+ // invert priorities to avoid starvation seen in the component network
+ if (m_wakeups_wo_switch > PRIORITY_SWITCH_LIMIT) {
+ m_wakeups_wo_switch = 0;
+ iteration_direction = true;
}
- if ((bw_remaining > 0) && ((m_in[vnet]->isReady()) || (m_units_remaining[vnet] > 0)) && !m_out[vnet]->areNSlotsAvailable(1)) {
- DEBUG_MSG(NETWORK_COMP,LowPrio,vnet);
- schedule_wakeup = true; // schedule me to wakeup again because I'm waiting for my output queue to become available
- }
- }
-
- // We should only wake up when we use the bandwidth
- // assert(bw_remaining != getLinkBandwidth()); // This is only mostly true
-
- // Record that we used some or all of the link bandwidth this cycle
- double ratio = 1.0-(double(bw_remaining)/double(getLinkBandwidth()));
- // If ratio = 0, we used no bandwidth, if ratio = 1, we used all
- linkUtilized(ratio);
-
- // Sample the link bandwidth utilization over a number of cycles
- int bw_used = getLinkBandwidth()-bw_remaining;
- m_bandwidth_since_sample += bw_used;
-
- // FIXME - comment out the bash specific code for faster performance
- // Start Bash code
- // Update the predictor
- Time current_time = g_eventQueue_ptr->getTime();
- while ((current_time - m_last_bandwidth_sample) > ADJUST_INTERVAL) {
- double utilization = m_bandwidth_since_sample/double(ADJUST_INTERVAL * getLinkBandwidth());
-
- if (utilization > g_bash_bandwidth_adaptive_threshold) {
- // Used more bandwidth
- m_bash_counter++;
+ if (iteration_direction) {
+ for (auto& it : m_in) {
+ int vnet = it.first;
+ operateVnet(vnet, bw_remaining, schedule_wakeup,
+ it.second, m_out[vnet]);
+ }
} else {
- // Used less bandwidth
- m_bash_counter--;
+ for (auto it = m_in.rbegin(); it != m_in.rend(); ++it) {
+ int vnet = (*it).first;
+ operateVnet(vnet, bw_remaining, schedule_wakeup,
+ (*it).second, m_out[vnet]);
+ }
}
- // Make sure we don't overflow
- m_bash_counter = min(HIGH_RANGE, m_bash_counter);
- m_bash_counter = max(0, m_bash_counter);
-
- // Reset samples
- m_last_bandwidth_sample += ADJUST_INTERVAL;
- m_bandwidth_since_sample = 0;
- }
- // End Bash code
-
- if ((bw_remaining > 0) && !schedule_wakeup) {
- // We have extra bandwidth and our output buffer was available, so we must not have anything else to do until another message arrives.
- DEBUG_MSG(NETWORK_COMP,LowPrio,*this);
- DEBUG_MSG(NETWORK_COMP,LowPrio,"not scheduled again");
- } else {
- DEBUG_MSG(NETWORK_COMP,LowPrio,*this);
- DEBUG_MSG(NETWORK_COMP,LowPrio,"scheduled again");
- // We are out of bandwidth for this cycle, so wakeup next cycle and continue
- g_eventQueue_ptr->scheduleEvent(this, 1);
- }
-}
+ // We should only wake up when we use the bandwidth
+ // This is only mostly true
+ // assert(bw_remaining != getLinkBandwidth());
-bool Throttle::broadcastBandwidthAvailable(int rand) const
-{
- bool result = !(m_bash_counter > ((HIGH_RANGE/4) + (rand % (HIGH_RANGE/2))));
- return result;
-}
+ // Record that we used some or all of the link bandwidth this cycle
+ double ratio = 1.0 - (double(bw_remaining) / double(getLinkBandwidth()));
-void Throttle::printStats(ostream& out) const
-{
- out << "utilized_percent: " << getUtilization() << endl;
-}
+ // If ratio = 0, we used no bandwidth, if ratio = 1, we used all
+ m_link_utilization_proxy += ratio;
-void Throttle::clearStats()
-{
- m_ruby_start = g_eventQueue_ptr->getTime();
- m_links_utilized = 0.0;
+ if (bw_remaining > 0 && !schedule_wakeup) {
+ // We have extra bandwidth and our output buffer was
+ // available, so we must not have anything else to do until
+ // another message arrives.
+ DPRINTF(RubyNetwork, "%s not scheduled again\n", *this);
+ } else {
+ DPRINTF(RubyNetwork, "%s scheduled again\n", *this);
- for (int i=0; i<m_message_counters.size(); i++) {
- for (int j=0; j<m_message_counters[i].size(); j++) {
- m_message_counters[i][j] = 0;
+ // We are out of bandwidth for this cycle, so wakeup next
+ // cycle and continue
+ scheduleEvent(Cycles(1));
}
- }
}
-void Throttle::printConfig(ostream& out) const
+void
+Throttle::regStats(string parent)
{
-
+ m_link_utilization
+ .name(parent + csprintf(".throttle%i", m_node) + ".link_utilization");
+
+ for (MessageSizeType type = MessageSizeType_FIRST;
+ type < MessageSizeType_NUM; ++type) {
+ m_msg_counts[(unsigned int)type]
+ .init(Network::getNumberOfVirtualNetworks())
+ .name(parent + csprintf(".throttle%i", m_node) + ".msg_count." +
+ MessageSizeType_to_string(type))
+ .flags(Stats::nozero)
+ ;
+ m_msg_bytes[(unsigned int) type]
+ .name(parent + csprintf(".throttle%i", m_node) + ".msg_bytes." +
+ MessageSizeType_to_string(type))
+ .flags(Stats::nozero)
+ ;
+
+ m_msg_bytes[(unsigned int) type] = m_msg_counts[type] * Stats::constant(
+ Network::MessageSizeType_to_int(type));
+ }
}
-double Throttle::getUtilization() const
+void
+Throttle::clearStats()
{
- return (100.0 * double(m_links_utilized)) / (double(g_eventQueue_ptr->getTime()-m_ruby_start));
+ m_link_utilization_proxy = 0;
}
-void Throttle::print(ostream& out) const
+void
+Throttle::collateStats()
{
- out << "[Throttle: " << m_sID << " " << m_node << " bw: " << getLinkBandwidth() << "]";
+ m_link_utilization = 100.0 * m_link_utilization_proxy
+ / (double(g_system_ptr->curCycle() - g_ruby_start));
}
-// Helper function
+void
+Throttle::print(ostream& out) const
+{
+ ccprintf(out, "[%i bw: %i]", m_node, getLinkBandwidth());
+}
-static
-int network_message_to_size(NetworkMessage* net_msg_ptr)
+int
+network_message_to_size(NetworkMessage* net_msg_ptr)
{
- assert(net_msg_ptr != NULL);
+ assert(net_msg_ptr != NULL);
- // Artificially increase the size of broadcast messages
- if (BROADCAST_SCALING > 1) {
- if (net_msg_ptr->getDestination().isBroadcast()) {
- return (MessageSizeType_to_int(net_msg_ptr->getMessageSize()) * MESSAGE_SIZE_MULTIPLIER * BROADCAST_SCALING);
- }
- }
- return (MessageSizeType_to_int(net_msg_ptr->getMessageSize()) * MESSAGE_SIZE_MULTIPLIER);
+ int size = Network::MessageSizeType_to_int(net_msg_ptr->getMessageSize());
+ size *= MESSAGE_SIZE_MULTIPLIER;
+
+ // Artificially increase the size of broadcast messages
+ if (BROADCAST_SCALING > 1 && net_msg_ptr->getDestination().isBroadcast())
+ size *= BROADCAST_SCALING;
+
+ return size;
}