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, int link_latency,
| 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, int link_latency,
|
52 int link_bandwidth_multiplier, int endpoint_bandwidth)
| 52 int link_bandwidth_multiplier, int endpoint_bandwidth,
53 EventManager *em)
54 : Consumer(em)
|
53{
54 init(node, link_latency, link_bandwidth_multiplier, endpoint_bandwidth);
55 m_sID = sID;
56}
57
58Throttle::Throttle(NodeID node, int link_latency,
| 55{
56 init(node, link_latency, link_bandwidth_multiplier, endpoint_bandwidth);
57 m_sID = sID;
58}
59
60Throttle::Throttle(NodeID node, int link_latency,
|
59 int link_bandwidth_multiplier, int endpoint_bandwidth)
| 61 int link_bandwidth_multiplier, int endpoint_bandwidth,
62 EventManager *em)
63 : Consumer(em)
|
60{
61 init(node, link_latency, link_bandwidth_multiplier, endpoint_bandwidth);
62 m_sID = 0;
63}
64
65void
66Throttle::init(NodeID node, int link_latency, int link_bandwidth_multiplier,
67 int endpoint_bandwidth)
68{
69 m_node = node;
70 m_vnets = 0;
71
72 assert(link_bandwidth_multiplier > 0);
73 m_link_bandwidth_multiplier = link_bandwidth_multiplier;
74 m_link_latency = link_latency;
75 m_endpoint_bandwidth = endpoint_bandwidth;
76
77 m_wakeups_wo_switch = 0;
78 clearStats();
79}
80
81void
82Throttle::clear()
83{
84 for (int counter = 0; counter < m_vnets; counter++) {
85 m_in[counter]->clear();
86 m_out[counter]->clear();
87 }
88}
89
90void
91Throttle::addLinks(const std::vector<MessageBuffer*>& in_vec,
92 const std::vector<MessageBuffer*>& out_vec)
93{
94 assert(in_vec.size() == out_vec.size());
95 for (int i=0; i<in_vec.size(); i++) {
96 addVirtualNetwork(in_vec[i], out_vec[i]);
97 }
98
99 m_message_counters.resize(MessageSizeType_NUM);
100 for (int i = 0; i < MessageSizeType_NUM; i++) {
101 m_message_counters[i].resize(in_vec.size());
102 for (int j = 0; j<m_message_counters[i].size(); j++) {
103 m_message_counters[i][j] = 0;
104 }
105 }
106}
107
108void
109Throttle::addVirtualNetwork(MessageBuffer* in_ptr, MessageBuffer* out_ptr)
110{
111 m_units_remaining.push_back(0);
112 m_in.push_back(in_ptr);
113 m_out.push_back(out_ptr);
114
115 // Set consumer and description
116 m_in[m_vnets]->setConsumer(this);
117 string desc = "[Queue to Throttle " + to_string(m_sID) + " " +
118 to_string(m_node) + "]";
119 m_in[m_vnets]->setDescription(desc);
120 m_vnets++;
121}
122
123void
124Throttle::wakeup()
125{
126 // Limits the number of message sent to a limited number of bytes/cycle.
127 assert(getLinkBandwidth() > 0);
128 int bw_remaining = getLinkBandwidth();
129
130 // Give the highest numbered link priority most of the time
131 m_wakeups_wo_switch++;
132 int highest_prio_vnet = m_vnets-1;
133 int lowest_prio_vnet = 0;
134 int counter = 1;
135 bool schedule_wakeup = false;
136
137 // invert priorities to avoid starvation seen in the component network
138 if (m_wakeups_wo_switch > PRIORITY_SWITCH_LIMIT) {
139 m_wakeups_wo_switch = 0;
140 highest_prio_vnet = 0;
141 lowest_prio_vnet = m_vnets-1;
142 counter = -1;
143 }
144
145 for (int vnet = highest_prio_vnet;
146 (vnet * counter) >= (counter * lowest_prio_vnet);
147 vnet -= counter) {
148
149 assert(m_out[vnet] != NULL);
150 assert(m_in[vnet] != NULL);
151 assert(m_units_remaining[vnet] >= 0);
152
153 while (bw_remaining > 0 &&
154 (m_in[vnet]->isReady() || m_units_remaining[vnet] > 0) &&
155 m_out[vnet]->areNSlotsAvailable(1)) {
156
157 // See if we are done transferring the previous message on
158 // this virtual network
159 if (m_units_remaining[vnet] == 0 && m_in[vnet]->isReady()) {
160 // Find the size of the message we are moving
161 MsgPtr msg_ptr = m_in[vnet]->peekMsgPtr();
162 NetworkMessage* net_msg_ptr =
163 safe_cast<NetworkMessage*>(msg_ptr.get());
164 m_units_remaining[vnet] +=
165 network_message_to_size(net_msg_ptr);
166
167 DPRINTF(RubyNetwork, "throttle: %d my bw %d bw spent "
168 "enqueueing net msg %d time: %lld.\n",
169 m_node, getLinkBandwidth(), m_units_remaining[vnet],
170 g_system_ptr->getTime());
171
172 // Move the message
173 m_out[vnet]->enqueue(m_in[vnet]->peekMsgPtr(), m_link_latency);
174 m_in[vnet]->pop();
175
176 // Count the message
177 m_message_counters[net_msg_ptr->getMessageSize()][vnet]++;
178
179 DPRINTF(RubyNetwork, "%s\n", *m_out[vnet]);
180 }
181
182 // Calculate the amount of bandwidth we spent on this message
183 int diff = m_units_remaining[vnet] - bw_remaining;
184 m_units_remaining[vnet] = max(0, diff);
185 bw_remaining = max(0, -diff);
186 }
187
188 if (bw_remaining > 0 &&
189 (m_in[vnet]->isReady() || m_units_remaining[vnet] > 0) &&
190 !m_out[vnet]->areNSlotsAvailable(1)) {
191 DPRINTF(RubyNetwork, "vnet: %d", vnet);
192 // schedule me to wakeup again because I'm waiting for my
193 // output queue to become available
194 schedule_wakeup = true;
195 }
196 }
197
198 // We should only wake up when we use the bandwidth
199 // This is only mostly true
200 // assert(bw_remaining != getLinkBandwidth());
201
202 // Record that we used some or all of the link bandwidth this cycle
203 double ratio = 1.0 - (double(bw_remaining) / double(getLinkBandwidth()));
204
205 // If ratio = 0, we used no bandwidth, if ratio = 1, we used all
206 linkUtilized(ratio);
207
208 if (bw_remaining > 0 && !schedule_wakeup) {
209 // We have extra bandwidth and our output buffer was
210 // available, so we must not have anything else to do until
211 // another message arrives.
212 DPRINTF(RubyNetwork, "%s not scheduled again\n", *this);
213 } else {
214 DPRINTF(RubyNetwork, "%s scheduled again\n", *this);
215
216 // We are out of bandwidth for this cycle, so wakeup next
217 // cycle and continue
218 scheduleEvent(1);
219 }
220}
221
222void
223Throttle::printStats(ostream& out) const
224{
225 out << "utilized_percent: " << getUtilization() << endl;
226}
227
228void
229Throttle::clearStats()
230{
231 m_ruby_start = g_system_ptr->getTime();
232 m_links_utilized = 0.0;
233
234 for (int i = 0; i < m_message_counters.size(); i++) {
235 for (int j = 0; j < m_message_counters[i].size(); j++) {
236 m_message_counters[i][j] = 0;
237 }
238 }
239}
240
241double
242Throttle::getUtilization() const
243{
244 return 100.0 * double(m_links_utilized) /
245 double(g_system_ptr->getTime()-m_ruby_start);
246}
247
248void
249Throttle::print(ostream& out) const
250{
251 ccprintf(out, "[%i bw: %i]", m_node, getLinkBandwidth());
252}
253
254int
255network_message_to_size(NetworkMessage* net_msg_ptr)
256{
257 assert(net_msg_ptr != NULL);
258
259 int size = RubySystem::getNetwork()->
260 MessageSizeType_to_int(net_msg_ptr->getMessageSize());
261 size *= MESSAGE_SIZE_MULTIPLIER;
262
263 // Artificially increase the size of broadcast messages
264 if (BROADCAST_SCALING > 1 && net_msg_ptr->getDestination().isBroadcast())
265 size *= BROADCAST_SCALING;
266
267 return size;
268}
| 64{
65 init(node, link_latency, link_bandwidth_multiplier, endpoint_bandwidth);
66 m_sID = 0;
67}
68
69void
70Throttle::init(NodeID node, int link_latency, int link_bandwidth_multiplier,
71 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)
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]);
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{
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);
121 string desc = "[Queue to Throttle " + to_string(m_sID) + " " +
122 to_string(m_node) + "]";
123 m_in[m_vnets]->setDescription(desc);
124 m_vnets++;
125}
126
127void
128Throttle::wakeup()
129{
130 // Limits the number of message sent to a limited number of bytes/cycle.
131 assert(getLinkBandwidth() > 0);
132 int bw_remaining = getLinkBandwidth();
133
134 // Give the highest numbered link priority most of the time
135 m_wakeups_wo_switch++;
136 int highest_prio_vnet = m_vnets-1;
137 int lowest_prio_vnet = 0;
138 int counter = 1;
139 bool schedule_wakeup = false;
140
141 // invert priorities to avoid starvation seen in the component network
142 if (m_wakeups_wo_switch > PRIORITY_SWITCH_LIMIT) {
143 m_wakeups_wo_switch = 0;
144 highest_prio_vnet = 0;
145 lowest_prio_vnet = m_vnets-1;
146 counter = -1;
147 }
148
149 for (int vnet = highest_prio_vnet;
150 (vnet * counter) >= (counter * lowest_prio_vnet);
151 vnet -= counter) {
152
153 assert(m_out[vnet] != NULL);
154 assert(m_in[vnet] != NULL);
155 assert(m_units_remaining[vnet] >= 0);
156
157 while (bw_remaining > 0 &&
158 (m_in[vnet]->isReady() || m_units_remaining[vnet] > 0) &&
159 m_out[vnet]->areNSlotsAvailable(1)) {
160
161 // See if we are done transferring the previous message on
162 // this virtual network
163 if (m_units_remaining[vnet] == 0 && m_in[vnet]->isReady()) {
164 // Find the size of the message we are moving
165 MsgPtr msg_ptr = m_in[vnet]->peekMsgPtr();
166 NetworkMessage* net_msg_ptr =
167 safe_cast<NetworkMessage*>(msg_ptr.get());
168 m_units_remaining[vnet] +=
169 network_message_to_size(net_msg_ptr);
170
171 DPRINTF(RubyNetwork, "throttle: %d my bw %d bw spent "
172 "enqueueing net msg %d time: %lld.\n",
173 m_node, getLinkBandwidth(), m_units_remaining[vnet],
174 g_system_ptr->getTime());
175
176 // Move the message
177 m_out[vnet]->enqueue(m_in[vnet]->peekMsgPtr(), m_link_latency);
178 m_in[vnet]->pop();
179
180 // Count the message
181 m_message_counters[net_msg_ptr->getMessageSize()][vnet]++;
182
183 DPRINTF(RubyNetwork, "%s\n", *m_out[vnet]);
184 }
185
186 // Calculate the amount of bandwidth we spent on this message
187 int diff = m_units_remaining[vnet] - bw_remaining;
188 m_units_remaining[vnet] = max(0, diff);
189 bw_remaining = max(0, -diff);
190 }
191
192 if (bw_remaining > 0 &&
193 (m_in[vnet]->isReady() || m_units_remaining[vnet] > 0) &&
194 !m_out[vnet]->areNSlotsAvailable(1)) {
195 DPRINTF(RubyNetwork, "vnet: %d", vnet);
196 // schedule me to wakeup again because I'm waiting for my
197 // output queue to become available
198 schedule_wakeup = true;
199 }
200 }
201
202 // We should only wake up when we use the bandwidth
203 // This is only mostly true
204 // assert(bw_remaining != getLinkBandwidth());
205
206 // Record that we used some or all of the link bandwidth this cycle
207 double ratio = 1.0 - (double(bw_remaining) / double(getLinkBandwidth()));
208
209 // If ratio = 0, we used no bandwidth, if ratio = 1, we used all
210 linkUtilized(ratio);
211
212 if (bw_remaining > 0 && !schedule_wakeup) {
213 // We have extra bandwidth and our output buffer was
214 // available, so we must not have anything else to do until
215 // another message arrives.
216 DPRINTF(RubyNetwork, "%s not scheduled again\n", *this);
217 } else {
218 DPRINTF(RubyNetwork, "%s scheduled again\n", *this);
219
220 // We are out of bandwidth for this cycle, so wakeup next
221 // cycle and continue
222 scheduleEvent(1);
223 }
224}
225
226void
227Throttle::printStats(ostream& out) const
228{
229 out << "utilized_percent: " << getUtilization() << endl;
230}
231
232void
233Throttle::clearStats()
234{
235 m_ruby_start = g_system_ptr->getTime();
236 m_links_utilized = 0.0;
237
238 for (int i = 0; i < m_message_counters.size(); i++) {
239 for (int j = 0; j < m_message_counters[i].size(); j++) {
240 m_message_counters[i][j] = 0;
241 }
242 }
243}
244
245double
246Throttle::getUtilization() const
247{
248 return 100.0 * double(m_links_utilized) /
249 double(g_system_ptr->getTime()-m_ruby_start);
250}
251
252void
253Throttle::print(ostream& out) const
254{
255 ccprintf(out, "[%i bw: %i]", m_node, getLinkBandwidth());
256}
257
258int
259network_message_to_size(NetworkMessage* net_msg_ptr)
260{
261 assert(net_msg_ptr != NULL);
262
263 int size = RubySystem::getNetwork()->
264 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}
|