1/*
2 * Copyright (c) 1999-2008 Mark D. Hill and David A. Wood
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions are
7 * met: redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer;
9 * redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution;
12 * neither the name of the copyright holders nor the names of its
13 * contributors may be used to endorse or promote products derived from
14 * this software without specific prior written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
17 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
20 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
26 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 */
28
29#include <cassert>
30
31#include "base/cast.hh"
32#include "base/cprintf.hh"
33#include "debug/RubyNetwork.hh"
34#include "mem/ruby/buffers/MessageBuffer.hh"
35#include "mem/ruby/network/simple/Throttle.hh"
36#include "mem/ruby/network/Network.hh"
37#include "mem/ruby/slicc_interface/NetworkMessage.hh"
38#include "mem/ruby/system/System.hh"
39
40using namespace std;
41
42const int HIGH_RANGE = 256;
43const int ADJUST_INTERVAL = 50000;
44const int MESSAGE_SIZE_MULTIPLIER = 1000;
45//const int BROADCAST_SCALING = 4; // Have a 16p system act like a 64p systems
46const int BROADCAST_SCALING = 1;
47const int PRIORITY_SWITCH_LIMIT = 128;
48
49static int network_message_to_size(NetworkMessage* net_msg_ptr);
50
51Throttle::Throttle(int sID, NodeID node, Cycles link_latency,
52 int link_bandwidth_multiplier, int endpoint_bandwidth,
53 ClockedObject *em)
54 : Consumer(em)
55{
56 init(node, link_latency, link_bandwidth_multiplier, endpoint_bandwidth);
57 m_sID = sID;
58}
59
60Throttle::Throttle(NodeID node, Cycles link_latency,
61 int link_bandwidth_multiplier, int endpoint_bandwidth,
62 ClockedObject *em)
63 : Consumer(em)
64{
65 init(node, link_latency, link_bandwidth_multiplier, endpoint_bandwidth);
66 m_sID = 0;
67}
68
69void
70Throttle::init(NodeID node, Cycles link_latency,
71 int link_bandwidth_multiplier, int endpoint_bandwidth)
72{
73 m_node = node;
74 m_vnets = 0;
75
76 assert(link_bandwidth_multiplier > 0);
77 m_link_bandwidth_multiplier = link_bandwidth_multiplier;
78 m_link_latency = link_latency;
79 m_endpoint_bandwidth = endpoint_bandwidth;
80
81 m_wakeups_wo_switch = 0;
82 clearStats();
83}
84
85void
86Throttle::clear()
87{
88 for (int counter = 0; counter < m_vnets; counter++) {
89 m_in[counter]->clear();
90 m_out[counter]->clear();
91 }
92}
93
94void
95Throttle::addLinks(const std::vector<MessageBuffer*>& in_vec,
96 const std::vector<MessageBuffer*>& out_vec, ClockedObject *em)
96 const std::vector& out_vec)
97{
98 assert(in_vec.size() == out_vec.size());
99 for (int i=0; i<in_vec.size(); i++) {
100 addVirtualNetwork(in_vec[i], out_vec[i], em);
100 addVirtualNetwork(in_vec[i], out_vec[i]);
101 }
102
103 m_message_counters.resize(MessageSizeType_NUM);
104 for (int i = 0; i < MessageSizeType_NUM; i++) {
105 m_message_counters[i].resize(in_vec.size());
106 for (int j = 0; j<m_message_counters[i].size(); j++) {
107 m_message_counters[i][j] = 0;
108 }
109 }
110}
111
112void
113Throttle::addVirtualNetwork(MessageBuffer* in_ptr, MessageBuffer* out_ptr,
114 ClockedObject *em)
113Throttle::addVirtualNetwork(MessageBuffer* in_ptr, MessageBuffer* out_ptr)
114{
115 m_units_remaining.push_back(0);
116 m_in.push_back(in_ptr);
117 m_out.push_back(out_ptr);
118
119 // Set consumer and description
120 m_in[m_vnets]->setConsumer(this);
122 m_in[m_vnets]->setClockObj(em);
121
122 string desc = "[Queue to Throttle " + to_string(m_sID) + " " +
123 to_string(m_node) + "]";
124 m_in[m_vnets]->setDescription(desc);
125 m_vnets++;
126}
127
128void
129Throttle::wakeup()
130{
131 // Limits the number of message sent to a limited number of bytes/cycle.
132 assert(getLinkBandwidth() > 0);
133 int bw_remaining = getLinkBandwidth();
134
135 // Give the highest numbered link priority most of the time
136 m_wakeups_wo_switch++;
137 int highest_prio_vnet = m_vnets-1;
138 int lowest_prio_vnet = 0;
139 int counter = 1;
140 bool schedule_wakeup = false;
141
142 // invert priorities to avoid starvation seen in the component network
143 if (m_wakeups_wo_switch > PRIORITY_SWITCH_LIMIT) {
144 m_wakeups_wo_switch = 0;
145 highest_prio_vnet = 0;
146 lowest_prio_vnet = m_vnets-1;
147 counter = -1;
148 }
149
150 for (int vnet = highest_prio_vnet;
151 (vnet * counter) >= (counter * lowest_prio_vnet);
152 vnet -= counter) {
153
154 assert(m_out[vnet] != NULL);
155 assert(m_in[vnet] != NULL);
156 assert(m_units_remaining[vnet] >= 0);
157
158 while (bw_remaining > 0 &&
159 (m_in[vnet]->isReady() || m_units_remaining[vnet] > 0) &&
160 m_out[vnet]->areNSlotsAvailable(1)) {
161
162 // See if we are done transferring the previous message on
163 // this virtual network
164 if (m_units_remaining[vnet] == 0 && m_in[vnet]->isReady()) {
165 // Find the size of the message we are moving
166 MsgPtr msg_ptr = m_in[vnet]->peekMsgPtr();
167 NetworkMessage* net_msg_ptr =
168 safe_cast<NetworkMessage*>(msg_ptr.get());
169 m_units_remaining[vnet] +=
170 network_message_to_size(net_msg_ptr);
171
172 DPRINTF(RubyNetwork, "throttle: %d my bw %d bw spent "
173 "enqueueing net msg %d time: %lld.\n",
174 m_node, getLinkBandwidth(), m_units_remaining[vnet],
177 g_system_ptr->getTime());
175 g_system_ptr->curCycle());
176
177 // Move the message
178 m_out[vnet]->enqueue(m_in[vnet]->peekMsgPtr(), m_link_latency);
179 m_in[vnet]->pop();
180
181 // Count the message
182 m_message_counters[net_msg_ptr->getMessageSize()][vnet]++;
183
184 DPRINTF(RubyNetwork, "%s\n", *m_out[vnet]);
185 }
186
187 // Calculate the amount of bandwidth we spent on this message
188 int diff = m_units_remaining[vnet] - bw_remaining;
189 m_units_remaining[vnet] = max(0, diff);
190 bw_remaining = max(0, -diff);
191 }
192
193 if (bw_remaining > 0 &&
194 (m_in[vnet]->isReady() || m_units_remaining[vnet] > 0) &&
195 !m_out[vnet]->areNSlotsAvailable(1)) {
196 DPRINTF(RubyNetwork, "vnet: %d", vnet);
197 // schedule me to wakeup again because I'm waiting for my
198 // output queue to become available
199 schedule_wakeup = true;
200 }
201 }
202
203 // We should only wake up when we use the bandwidth
204 // This is only mostly true
205 // assert(bw_remaining != getLinkBandwidth());
206
207 // Record that we used some or all of the link bandwidth this cycle
208 double ratio = 1.0 - (double(bw_remaining) / double(getLinkBandwidth()));
209
210 // If ratio = 0, we used no bandwidth, if ratio = 1, we used all
211 linkUtilized(ratio);
212
213 if (bw_remaining > 0 && !schedule_wakeup) {
214 // We have extra bandwidth and our output buffer was
215 // available, so we must not have anything else to do until
216 // another message arrives.
217 DPRINTF(RubyNetwork, "%s not scheduled again\n", *this);
218 } else {
219 DPRINTF(RubyNetwork, "%s scheduled again\n", *this);
220
221 // We are out of bandwidth for this cycle, so wakeup next
222 // cycle and continue
223 scheduleEvent(Cycles(1));
224 }
225}
226
227void
228Throttle::printStats(ostream& out) const
229{
230 out << "utilized_percent: " << getUtilization() << endl;
231}
232
233void
234Throttle::clearStats()
235{
238 m_ruby_start = g_system_ptr->getTime();
236 m_ruby_start = g_system_ptr->curCycle();
237 m_links_utilized = 0.0;
238
239 for (int i = 0; i < m_message_counters.size(); i++) {
240 for (int j = 0; j < m_message_counters[i].size(); j++) {
241 m_message_counters[i][j] = 0;
242 }
243 }
244}
245
246double
247Throttle::getUtilization() const
248{
249 return 100.0 * double(m_links_utilized) /
252 double(g_system_ptr->getTime()-m_ruby_start);
250 double(g_system_ptr->curCycle()-m_ruby_start);
251}
252
253void
254Throttle::print(ostream& out) const
255{
256 ccprintf(out, "[%i bw: %i]", m_node, getLinkBandwidth());
257}
258
259int
260network_message_to_size(NetworkMessage* net_msg_ptr)
261{
262 assert(net_msg_ptr != NULL);
263
264 int size = Network::MessageSizeType_to_int(net_msg_ptr->getMessageSize());
265 size *= MESSAGE_SIZE_MULTIPLIER;
266
267 // Artificially increase the size of broadcast messages
268 if (BROADCAST_SCALING > 1 && net_msg_ptr->getDestination().isBroadcast())
269 size *= BROADCAST_SCALING;
270
271 return size;
272}