Deleted Added
sdiff udiff text old ( 6433:0f0f0fbef977 ) new ( 6876:a658c315512c )
full compact
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/*
30 This file has been modified by Kevin Moore and Dan Nussbaum of the
31 Scalable Systems Research Group at Sun Microsystems Laboratories
32 (http://research.sun.com/scalable/) to support the Adaptive
33 Transactional Memory Test Platform (ATMTP).
34
35 Please send email to atmtp-interest@sun.com with feedback, questions, or
36 to request future announcements about ATMTP.
37
38 ----------------------------------------------------------------------
39
40 File modification date: 2008-02-23
41
42 ----------------------------------------------------------------------
43*/
44
45/*
46 * Profiler.cc
47 *
48 * Description: See Profiler.hh
49 *
50 * $Id$
51 *
52 */
53
54#include "mem/ruby/profiler/Profiler.hh"
55#include "mem/ruby/profiler/AddressProfiler.hh"
56#include "mem/ruby/system/System.hh"
57#include "mem/ruby/network/Network.hh"
58#include "mem/gems_common/PrioHeap.hh"
59#include "mem/protocol/CacheMsg.hh"
60#include "mem/protocol/Protocol.hh"
61#include "mem/gems_common/util.hh"
62#include "mem/gems_common/Map.hh"
63#include "mem/ruby/common/Debug.hh"
64#include "mem/protocol/MachineType.hh"
65
66// Allows use of times() library call, which determines virtual runtime
67#include <sys/times.h>
68
69extern std::ostream * debug_cout_ptr;
70
71static double process_memory_total();
72static double process_memory_resident();
73
74Profiler::Profiler(const string & name)
75{
76 m_name = name;
77 m_requestProfileMap_ptr = new Map<string, int>;
78
79 m_inst_profiler_ptr = NULL;
80 m_address_profiler_ptr = NULL;
81
82 m_real_time_start_time = time(NULL); // Not reset in clearStats()
83 m_stats_period = 1000000; // Default
84 m_periodic_output_file_ptr = &cerr;
85
86}
87
88Profiler::~Profiler()
89{
90 if (m_periodic_output_file_ptr != &cerr) {
91 delete m_periodic_output_file_ptr;
92 }
93 delete m_requestProfileMap_ptr;
94}
95
96void Profiler::init(const vector<string> & argv, vector<string> memory_control_names)
97{
98 // added by SS
99 vector<string>::iterator it;
100 memory_control_profiler* mcp;
101 m_memory_control_names = memory_control_names;
102// printf ( "Here in Profiler::init \n");
103 for ( it=memory_control_names.begin() ; it < memory_control_names.end(); it++ ){
104// printf ( "Here in Profiler::init memory control name %s \n", (*it).c_str());
105 mcp = new memory_control_profiler;
106 mcp->m_memReq = 0;
107 mcp->m_memBankBusy = 0;
108 mcp->m_memBusBusy = 0;
109 mcp->m_memReadWriteBusy = 0;
110 mcp->m_memDataBusBusy = 0;
111 mcp->m_memTfawBusy = 0;
112 mcp->m_memRefresh = 0;
113 mcp->m_memRead = 0;
114 mcp->m_memWrite = 0;
115 mcp->m_memWaitCycles = 0;
116 mcp->m_memInputQ = 0;
117 mcp->m_memBankQ = 0;
118 mcp->m_memArbWait = 0;
119 mcp->m_memRandBusy = 0;
120 mcp->m_memNotOld = 0;
121
122 mcp->m_banks_per_rank = RubySystem::getMemoryControl((*it).c_str())->getBanksPerRank();
123 mcp->m_ranks_per_dimm = RubySystem::getMemoryControl((*it).c_str())->getRanksPerDimm();
124 mcp->m_dimms_per_channel = RubySystem::getMemoryControl((*it).c_str())->getDimmsPerChannel();
125
126 int totalBanks = mcp->m_banks_per_rank
127 * mcp->m_ranks_per_dimm
128 * mcp->m_dimms_per_channel;
129
130 mcp->m_memBankCount.setSize(totalBanks);
131
132 m_memory_control_profilers [(*it).c_str()] = mcp;
133 }
134
135 clearStats();
136 m_hot_lines = false;
137 m_all_instructions = false;
138
139 for (size_t i=0; i<argv.size(); i+=2) {
140 if ( argv[i] == "hot_lines") {
141 m_hot_lines = (argv[i+1]=="true");
142 } else if ( argv[i] == "all_instructions") {
143 m_all_instructions = (argv[i+1]=="true");
144 }else {
145 cerr << "WARNING: Profiler: Unkown configuration parameter: " << argv[i] << endl;
146 assert(false);
147 }
148 }
149
150 m_address_profiler_ptr = new AddressProfiler;
151 m_address_profiler_ptr -> setHotLines(m_hot_lines);
152 m_address_profiler_ptr -> setAllInstructions(m_all_instructions);
153
154 if (m_all_instructions) {
155 m_inst_profiler_ptr = new AddressProfiler;
156 m_inst_profiler_ptr -> setHotLines(m_hot_lines);
157 m_inst_profiler_ptr -> setAllInstructions(m_all_instructions);
158 }
159}
160
161void Profiler::wakeup()
162{
163 // FIXME - avoid the repeated code
164
165 Vector<integer_t> perProcCycleCount;
166 perProcCycleCount.setSize(RubySystem::getNumberOfSequencers());
167
168 for(int i=0; i < RubySystem::getNumberOfSequencers(); i++) {
169 perProcCycleCount[i] = g_system_ptr->getCycleCount(i) - m_cycles_executed_at_start[i] + 1;
170 // The +1 allows us to avoid division by zero
171 }
172
173 integer_t total_misses = m_perProcTotalMisses.sum();
174 integer_t simics_cycles_executed = perProcCycleCount.sum();
175 integer_t transactions_started = m_perProcStartTransaction.sum();
176 integer_t transactions_ended = m_perProcEndTransaction.sum();
177
178 (*m_periodic_output_file_ptr) << "ruby_cycles: " << g_eventQueue_ptr->getTime()-m_ruby_start << endl;
179 (*m_periodic_output_file_ptr) << "total_misses: " << total_misses << " " << m_perProcTotalMisses << endl;
180 (*m_periodic_output_file_ptr) << "simics_cycles_executed: " << simics_cycles_executed << " " << perProcCycleCount << endl;
181 (*m_periodic_output_file_ptr) << "transactions_started: " << transactions_started << " " << m_perProcStartTransaction << endl;
182 (*m_periodic_output_file_ptr) << "transactions_ended: " << transactions_ended << " " << m_perProcEndTransaction << endl;
183 (*m_periodic_output_file_ptr) << "mbytes_resident: " << process_memory_resident() << endl;
184 (*m_periodic_output_file_ptr) << "mbytes_total: " << process_memory_total() << endl;
185 if (process_memory_total() > 0) {
186 (*m_periodic_output_file_ptr) << "resident_ratio: " << process_memory_resident()/process_memory_total() << endl;
187 }
188 (*m_periodic_output_file_ptr) << "miss_latency: " << m_allMissLatencyHistogram << endl;
189
190 *m_periodic_output_file_ptr << endl;
191
192 if (m_all_instructions) {
193 m_inst_profiler_ptr->printStats(*m_periodic_output_file_ptr);
194 }
195
196 //g_system_ptr->getNetwork()->printStats(*m_periodic_output_file_ptr);
197 g_eventQueue_ptr->scheduleEvent(this, m_stats_period);
198}
199
200void Profiler::setPeriodicStatsFile(const string& filename)
201{
202 cout << "Recording periodic statistics to file '" << filename << "' every "
203 << m_stats_period << " Ruby cycles" << endl;
204
205 if (m_periodic_output_file_ptr != &cerr) {
206 delete m_periodic_output_file_ptr;
207 }
208
209 m_periodic_output_file_ptr = new ofstream(filename.c_str());
210 g_eventQueue_ptr->scheduleEvent(this, 1);
211}
212
213void Profiler::setPeriodicStatsInterval(integer_t period)
214{
215 cout << "Recording periodic statistics every " << m_stats_period << " Ruby cycles" << endl;
216 m_stats_period = period;
217 g_eventQueue_ptr->scheduleEvent(this, 1);
218}
219
220void Profiler::printConfig(ostream& out) const
221{
222 out << endl;
223 out << "Profiler Configuration" << endl;
224 out << "----------------------" << endl;
225 out << "periodic_stats_period: " << m_stats_period << endl;
226}
227
228void Profiler::print(ostream& out) const
229{
230 out << "[Profiler]";
231}
232
233void Profiler::printStats(ostream& out, bool short_stats)
234{
235 out << endl;
236 if (short_stats) {
237 out << "SHORT ";
238 }
239 out << "Profiler Stats" << endl;
240 out << "--------------" << endl;
241
242 time_t real_time_current = time(NULL);
243 double seconds = difftime(real_time_current, m_real_time_start_time);
244 double minutes = seconds/60.0;
245 double hours = minutes/60.0;
246 double days = hours/24.0;
247 Time ruby_cycles = g_eventQueue_ptr->getTime()-m_ruby_start;
248
249 if (!short_stats) {
250 out << "Elapsed_time_in_seconds: " << seconds << endl;
251 out << "Elapsed_time_in_minutes: " << minutes << endl;
252 out << "Elapsed_time_in_hours: " << hours << endl;
253 out << "Elapsed_time_in_days: " << days << endl;
254 out << endl;
255 }
256
257 // print the virtual runtimes as well
258 struct tms vtime;
259 times(&vtime);
260 seconds = (vtime.tms_utime + vtime.tms_stime) / 100.0;
261 minutes = seconds / 60.0;
262 hours = minutes / 60.0;
263 days = hours / 24.0;
264 out << "Virtual_time_in_seconds: " << seconds << endl;
265 out << "Virtual_time_in_minutes: " << minutes << endl;
266 out << "Virtual_time_in_hours: " << hours << endl;
267 out << "Virtual_time_in_days: " << days << endl;
268 out << endl;
269
270 out << "Ruby_current_time: " << g_eventQueue_ptr->getTime() << endl;
271 out << "Ruby_start_time: " << m_ruby_start << endl;
272 out << "Ruby_cycles: " << ruby_cycles << endl;
273 out << endl;
274
275 if (!short_stats) {
276 out << "mbytes_resident: " << process_memory_resident() << endl;
277 out << "mbytes_total: " << process_memory_total() << endl;
278 if (process_memory_total() > 0) {
279 out << "resident_ratio: " << process_memory_resident()/process_memory_total() << endl;
280 }
281 out << endl;
282
283 }
284
285 Vector<integer_t> perProcCycleCount;
286 Vector<double> perProcCyclesPerTrans;
287 Vector<double> perProcMissesPerTrans;
288
289
290 perProcCycleCount.setSize(RubySystem::getNumberOfSequencers());
291 perProcCyclesPerTrans.setSize(RubySystem::getNumberOfSequencers());
292 perProcMissesPerTrans.setSize(RubySystem::getNumberOfSequencers());
293
294 for(int i=0; i < RubySystem::getNumberOfSequencers(); i++) {
295 perProcCycleCount[i] = g_system_ptr->getCycleCount(i) - m_cycles_executed_at_start[i] + 1;
296 // The +1 allows us to avoid division by zero
297
298 int trans = m_perProcEndTransaction[i];
299 if (trans == 0) {
300 perProcCyclesPerTrans[i] = 0;
301 perProcMissesPerTrans[i] = 0;
302 } else {
303 perProcCyclesPerTrans[i] = ruby_cycles / double(trans);
304 perProcMissesPerTrans[i] = m_perProcTotalMisses[i] / double(trans);
305 }
306 }
307
308 integer_t total_misses = m_perProcTotalMisses.sum();
309 integer_t user_misses = m_perProcUserMisses.sum();
310 integer_t supervisor_misses = m_perProcSupervisorMisses.sum();
311 integer_t simics_cycles_executed = perProcCycleCount.sum();
312 integer_t transactions_started = m_perProcStartTransaction.sum();
313 integer_t transactions_ended = m_perProcEndTransaction.sum();
314
315 double cycles_per_transaction = (transactions_ended != 0) ? (RubySystem::getNumberOfSequencers() * double(ruby_cycles)) / double(transactions_ended) : 0;
316 double misses_per_transaction = (transactions_ended != 0) ? double(total_misses) / double(transactions_ended) : 0;
317
318 out << "Total_misses: " << total_misses << endl;
319 out << "total_misses: " << total_misses << " " << m_perProcTotalMisses << endl;
320 out << "user_misses: " << user_misses << " " << m_perProcUserMisses << endl;
321 out << "supervisor_misses: " << supervisor_misses << " " << m_perProcSupervisorMisses << endl;
322 out << endl;
323 out << "ruby_cycles_executed: " << simics_cycles_executed << " " << perProcCycleCount << endl;
324 out << endl;
325 out << "transactions_started: " << transactions_started << " " << m_perProcStartTransaction << endl;
326 out << "transactions_ended: " << transactions_ended << " " << m_perProcEndTransaction << endl;
327 out << "cycles_per_transaction: " << cycles_per_transaction << " " << perProcCyclesPerTrans << endl;
328 out << "misses_per_transaction: " << misses_per_transaction << " " << perProcMissesPerTrans << endl;
329
330 out << endl;
331
332 out << endl;
333
334 vector<string>::iterator it;
335
336 for ( it=m_memory_control_names.begin() ; it < m_memory_control_names.end(); it++ ){
337 long long int m_memReq = m_memory_control_profilers[(*it).c_str()] -> m_memReq;
338 long long int m_memRefresh = m_memory_control_profilers[(*it).c_str()] -> m_memRefresh;
339 long long int m_memInputQ = m_memory_control_profilers[(*it).c_str()] -> m_memInputQ;
340 long long int m_memBankQ = m_memory_control_profilers[(*it).c_str()] -> m_memBankQ;
341 long long int m_memWaitCycles = m_memory_control_profilers[(*it).c_str()] -> m_memWaitCycles;
342 long long int m_memRead = m_memory_control_profilers[(*it).c_str()] -> m_memRead;
343 long long int m_memWrite = m_memory_control_profilers[(*it).c_str()] -> m_memWrite;
344 long long int m_memBankBusy = m_memory_control_profilers[(*it).c_str()] -> m_memBankBusy;
345 long long int m_memRandBusy = m_memory_control_profilers[(*it).c_str()] -> m_memRandBusy;
346 long long int m_memNotOld = m_memory_control_profilers[(*it).c_str()] -> m_memNotOld;
347 long long int m_memArbWait = m_memory_control_profilers[(*it).c_str()] -> m_memArbWait;
348 long long int m_memBusBusy = m_memory_control_profilers[(*it).c_str()] -> m_memBusBusy;
349 long long int m_memTfawBusy = m_memory_control_profilers[(*it).c_str()] -> m_memTfawBusy;
350 long long int m_memReadWriteBusy = m_memory_control_profilers[(*it).c_str()] -> m_memReadWriteBusy;
351 long long int m_memDataBusBusy = m_memory_control_profilers[(*it).c_str()] -> m_memDataBusBusy;
352 Vector<long long int> m_memBankCount = m_memory_control_profilers[(*it).c_str()] -> m_memBankCount;
353
354 if (m_memReq || m_memRefresh) { // if there's a memory controller at all
355 long long int total_stalls = m_memInputQ + m_memBankQ + m_memWaitCycles;
356 double stallsPerReq = total_stalls * 1.0 / m_memReq;
357 out << "Memory control " << (*it) << ":" << endl;
358 out << " memory_total_requests: " << m_memReq << endl; // does not include refreshes
359 out << " memory_reads: " << m_memRead << endl;
360 out << " memory_writes: " << m_memWrite << endl;
361 out << " memory_refreshes: " << m_memRefresh << endl;
362 out << " memory_total_request_delays: " << total_stalls << endl;
363 out << " memory_delays_per_request: " << stallsPerReq << endl;
364 out << " memory_delays_in_input_queue: " << m_memInputQ << endl;
365 out << " memory_delays_behind_head_of_bank_queue: " << m_memBankQ << endl;
366 out << " memory_delays_stalled_at_head_of_bank_queue: " << m_memWaitCycles << endl;
367 // Note: The following "memory stalls" entries are a breakdown of the
368 // cycles which already showed up in m_memWaitCycles. The order is
369 // significant; it is the priority of attributing the cycles.
370 // For example, bank_busy is before arbitration because if the bank was
371 // busy, we didn't even check arbitration.
372 // Note: "not old enough" means that since we grouped waiting heads-of-queues
373 // into batches to avoid starvation, a request in a newer batch
374 // didn't try to arbitrate yet because there are older requests waiting.
375 out << " memory_stalls_for_bank_busy: " << m_memBankBusy << endl;
376 out << " memory_stalls_for_random_busy: " << m_memRandBusy << endl;
377 out << " memory_stalls_for_anti_starvation: " << m_memNotOld << endl;
378 out << " memory_stalls_for_arbitration: " << m_memArbWait << endl;
379 out << " memory_stalls_for_bus: " << m_memBusBusy << endl;
380 out << " memory_stalls_for_tfaw: " << m_memTfawBusy << endl;
381 out << " memory_stalls_for_read_write_turnaround: " << m_memReadWriteBusy << endl;
382 out << " memory_stalls_for_read_read_turnaround: " << m_memDataBusBusy << endl;
383 out << " accesses_per_bank: ";
384 for (int bank=0; bank < m_memBankCount.size(); bank++) {
385 out << m_memBankCount[bank] << " ";
386 //if ((bank % 8) == 7) out << " " << endl;
387 }
388 out << endl;
389 out << endl;
390 }
391 }
392 if (!short_stats) {
393 out << "Busy Controller Counts:" << endl;
394 for(int i=0; i < MachineType_NUM; i++) {
395 for(int j=0; j < MachineType_base_count((MachineType)i); j++) {
396 MachineID machID;
397 machID.type = (MachineType)i;
398 machID.num = j;
399 out << machID << ":" << m_busyControllerCount[i][j] << " ";
400 if ((j+1)%8 == 0) {
401 out << endl;
402 }
403 }
404 out << endl;
405 }
406 out << endl;
407
408 out << "Busy Bank Count:" << m_busyBankCount << endl;
409 out << endl;
410
411 out << "sequencer_requests_outstanding: " << m_sequencer_requests << endl;
412 out << endl;
413 }
414
415 if (!short_stats) {
416 out << "All Non-Zero Cycle Demand Cache Accesses" << endl;
417 out << "----------------------------------------" << endl;
418 out << "miss_latency: " << m_allMissLatencyHistogram << endl;
419 for(int i=0; i<m_missLatencyHistograms.size(); i++) {
420 if (m_missLatencyHistograms[i].size() > 0) {
421 out << "miss_latency_" << RubyRequestType(i) << ": " << m_missLatencyHistograms[i] << endl;
422 }
423 }
424 for(int i=0; i<m_machLatencyHistograms.size(); i++) {
425 if (m_machLatencyHistograms[i].size() > 0) {
426 out << "miss_latency_" << GenericMachineType(i) << ": " << m_machLatencyHistograms[i] << endl;
427 }
428 }
429
430 out << endl;
431
432 out << "All Non-Zero Cycle SW Prefetch Requests" << endl;
433 out << "------------------------------------" << endl;
434 out << "prefetch_latency: " << m_allSWPrefetchLatencyHistogram << endl;
435 for(int i=0; i<m_SWPrefetchLatencyHistograms.size(); i++) {
436 if (m_SWPrefetchLatencyHistograms[i].size() > 0) {
437 out << "prefetch_latency_" << CacheRequestType(i) << ": " << m_SWPrefetchLatencyHistograms[i] << endl;
438 }
439 }
440 for(int i=0; i<m_SWPrefetchMachLatencyHistograms.size(); i++) {
441 if (m_SWPrefetchMachLatencyHistograms[i].size() > 0) {
442 out << "prefetch_latency_" << GenericMachineType(i) << ": " << m_SWPrefetchMachLatencyHistograms[i] << endl;
443 }
444 }
445 out << "prefetch_latency_L2Miss:" << m_SWPrefetchL2MissLatencyHistogram << endl;
446
447 if (m_all_sharing_histogram.size() > 0) {
448 out << "all_sharing: " << m_all_sharing_histogram << endl;
449 out << "read_sharing: " << m_read_sharing_histogram << endl;
450 out << "write_sharing: " << m_write_sharing_histogram << endl;
451
452 out << "all_sharing_percent: "; m_all_sharing_histogram.printPercent(out); out << endl;
453 out << "read_sharing_percent: "; m_read_sharing_histogram.printPercent(out); out << endl;
454 out << "write_sharing_percent: "; m_write_sharing_histogram.printPercent(out); out << endl;
455
456 int64 total_miss = m_cache_to_cache + m_memory_to_cache;
457 out << "all_misses: " << total_miss << endl;
458 out << "cache_to_cache_misses: " << m_cache_to_cache << endl;
459 out << "memory_to_cache_misses: " << m_memory_to_cache << endl;
460 out << "cache_to_cache_percent: " << 100.0 * (double(m_cache_to_cache) / double(total_miss)) << endl;
461 out << "memory_to_cache_percent: " << 100.0 * (double(m_memory_to_cache) / double(total_miss)) << endl;
462 out << endl;
463 }
464
465 if (m_outstanding_requests.size() > 0) {
466 out << "outstanding_requests: "; m_outstanding_requests.printPercent(out); out << endl;
467 out << endl;
468 }
469 }
470
471 if (!short_stats) {
472 out << "Request vs. RubySystem State Profile" << endl;
473 out << "--------------------------------" << endl;
474 out << endl;
475
476 Vector<string> requestProfileKeys = m_requestProfileMap_ptr->keys();
477 requestProfileKeys.sortVector();
478
479 for(int i=0; i<requestProfileKeys.size(); i++) {
480 int temp_int = m_requestProfileMap_ptr->lookup(requestProfileKeys[i]);
481 double percent = (100.0*double(temp_int))/double(m_requests);
482 while (requestProfileKeys[i] != "") {
483 out << setw(10) << string_split(requestProfileKeys[i], ':');
484 }
485 out << setw(11) << temp_int;
486 out << setw(14) << percent << endl;
487 }
488 out << endl;
489
490 out << "filter_action: " << m_filter_action_histogram << endl;
491
492 if (!m_all_instructions) {
493 m_address_profiler_ptr->printStats(out);
494 }
495
496 if (m_all_instructions) {
497 m_inst_profiler_ptr->printStats(out);
498 }
499
500 out << endl;
501 out << "Message Delayed Cycles" << endl;
502 out << "----------------------" << endl;
503 out << "Total_delay_cycles: " << m_delayedCyclesHistogram << endl;
504 out << "Total_nonPF_delay_cycles: " << m_delayedCyclesNonPFHistogram << endl;
505 for (int i = 0; i < m_delayedCyclesVCHistograms.size(); i++) {
506 out << " virtual_network_" << i << "_delay_cycles: " << m_delayedCyclesVCHistograms[i] << endl;
507 }
508
509 printResourceUsage(out);
510 }
511
512}
513
514void Profiler::printResourceUsage(ostream& out) const
515{
516 out << endl;
517 out << "Resource Usage" << endl;
518 out << "--------------" << endl;
519
520 integer_t pagesize = getpagesize(); // page size in bytes
521 out << "page_size: " << pagesize << endl;
522
523 rusage usage;
524 getrusage (RUSAGE_SELF, &usage);
525
526 out << "user_time: " << usage.ru_utime.tv_sec << endl;
527 out << "system_time: " << usage.ru_stime.tv_sec << endl;
528 out << "page_reclaims: " << usage.ru_minflt << endl;
529 out << "page_faults: " << usage.ru_majflt << endl;
530 out << "swaps: " << usage.ru_nswap << endl;
531 out << "block_inputs: " << usage.ru_inblock << endl;
532 out << "block_outputs: " << usage.ru_oublock << endl;
533}
534
535void Profiler::clearStats()
536{
537 m_ruby_start = g_eventQueue_ptr->getTime();
538
539 m_cycles_executed_at_start.setSize(RubySystem::getNumberOfSequencers());
540 for (int i=0; i < RubySystem::getNumberOfSequencers(); i++) {
541 if (g_system_ptr == NULL) {
542 m_cycles_executed_at_start[i] = 0;
543 } else {
544 m_cycles_executed_at_start[i] = g_system_ptr->getCycleCount(i);
545 }
546 }
547
548 m_perProcTotalMisses.setSize(RubySystem::getNumberOfSequencers());
549 m_perProcUserMisses.setSize(RubySystem::getNumberOfSequencers());
550 m_perProcSupervisorMisses.setSize(RubySystem::getNumberOfSequencers());
551 m_perProcStartTransaction.setSize(RubySystem::getNumberOfSequencers());
552 m_perProcEndTransaction.setSize(RubySystem::getNumberOfSequencers());
553
554 for(int i=0; i < RubySystem::getNumberOfSequencers(); i++) {
555 m_perProcTotalMisses[i] = 0;
556 m_perProcUserMisses[i] = 0;
557 m_perProcSupervisorMisses[i] = 0;
558 m_perProcStartTransaction[i] = 0;
559 m_perProcEndTransaction[i] = 0;
560 }
561
562 m_busyControllerCount.setSize(MachineType_NUM); // all machines
563 for(int i=0; i < MachineType_NUM; i++) {
564 m_busyControllerCount[i].setSize(MachineType_base_count((MachineType)i));
565 for(int j=0; j < MachineType_base_count((MachineType)i); j++) {
566 m_busyControllerCount[i][j] = 0;
567 }
568 }
569 m_busyBankCount = 0;
570
571 m_delayedCyclesHistogram.clear();
572 m_delayedCyclesNonPFHistogram.clear();
573 m_delayedCyclesVCHistograms.setSize(RubySystem::getNetwork()->getNumberOfVirtualNetworks());
574 for (int i = 0; i < RubySystem::getNetwork()->getNumberOfVirtualNetworks(); i++) {
575 m_delayedCyclesVCHistograms[i].clear();
576 }
577
578 m_missLatencyHistograms.setSize(RubyRequestType_NUM);
579 for(int i=0; i<m_missLatencyHistograms.size(); i++) {
580 m_missLatencyHistograms[i].clear(200);
581 }
582 m_machLatencyHistograms.setSize(GenericMachineType_NUM+1);
583 for(int i=0; i<m_machLatencyHistograms.size(); i++) {
584 m_machLatencyHistograms[i].clear(200);
585 }
586 m_allMissLatencyHistogram.clear(200);
587
588 m_SWPrefetchLatencyHistograms.setSize(CacheRequestType_NUM);
589 for(int i=0; i<m_SWPrefetchLatencyHistograms.size(); i++) {
590 m_SWPrefetchLatencyHistograms[i].clear(200);
591 }
592 m_SWPrefetchMachLatencyHistograms.setSize(GenericMachineType_NUM+1);
593 for(int i=0; i<m_SWPrefetchMachLatencyHistograms.size(); i++) {
594 m_SWPrefetchMachLatencyHistograms[i].clear(200);
595 }
596 m_allSWPrefetchLatencyHistogram.clear(200);
597
598 m_sequencer_requests.clear();
599 m_read_sharing_histogram.clear();
600 m_write_sharing_histogram.clear();
601 m_all_sharing_histogram.clear();
602 m_cache_to_cache = 0;
603 m_memory_to_cache = 0;
604
605 // clear HashMaps
606 m_requestProfileMap_ptr->clear();
607
608 // count requests profiled
609 m_requests = 0;
610
611 m_outstanding_requests.clear();
612 m_outstanding_persistent_requests.clear();
613
614//added by SS
615 vector<string>::iterator it;
616
617 for ( it=m_memory_control_names.begin() ; it < m_memory_control_names.end(); it++ ){
618 m_memory_control_profilers[(*it).c_str()] -> m_memReq = 0;
619 m_memory_control_profilers[(*it).c_str()] -> m_memBankBusy = 0;
620 m_memory_control_profilers[(*it).c_str()] -> m_memBusBusy = 0;
621 m_memory_control_profilers[(*it).c_str()] -> m_memTfawBusy = 0;
622 m_memory_control_profilers[(*it).c_str()] -> m_memReadWriteBusy = 0;
623 m_memory_control_profilers[(*it).c_str()] -> m_memDataBusBusy = 0;
624 m_memory_control_profilers[(*it).c_str()] -> m_memRefresh = 0;
625 m_memory_control_profilers[(*it).c_str()] -> m_memRead = 0;
626 m_memory_control_profilers[(*it).c_str()] -> m_memWrite = 0;
627 m_memory_control_profilers[(*it).c_str()] -> m_memWaitCycles = 0;
628 m_memory_control_profilers[(*it).c_str()] -> m_memInputQ = 0;
629 m_memory_control_profilers[(*it).c_str()] -> m_memBankQ = 0;
630 m_memory_control_profilers[(*it).c_str()] -> m_memArbWait = 0;
631 m_memory_control_profilers[(*it).c_str()] -> m_memRandBusy = 0;
632 m_memory_control_profilers[(*it).c_str()] -> m_memNotOld = 0;
633
634 for (int bank=0; bank < m_memory_control_profilers[(*it).c_str()] -> m_memBankCount.size(); bank++) {
635 m_memory_control_profilers[(*it).c_str()] -> m_memBankCount[bank] = 0;
636 }
637 }
638 // Flush the prefetches through the system - used so that there are no outstanding requests after stats are cleared
639 //g_eventQueue_ptr->triggerAllEvents();
640
641 // update the start time
642 m_ruby_start = g_eventQueue_ptr->getTime();
643}
644
645void Profiler::addAddressTraceSample(const CacheMsg& msg, NodeID id)
646{
647 if (msg.getType() != CacheRequestType_IFETCH) {
648
649 // Note: The following line should be commented out if you want to
650 // use the special profiling that is part of the GS320 protocol
651
652 // NOTE: Unless PROFILE_HOT_LINES is enabled, nothing will be profiled by the AddressProfiler
653 m_address_profiler_ptr->addTraceSample(msg.getLineAddress(), msg.getProgramCounter(), msg.getType(), msg.getAccessMode(), id, false);
654 }
655}
656
657void Profiler::profileSharing(const Address& addr, AccessType type, NodeID requestor, const Set& sharers, const Set& owner)
658{
659 Set set_contacted(owner);
660 if (type == AccessType_Write) {
661 set_contacted.addSet(sharers);
662 }
663 set_contacted.remove(requestor);
664 int number_contacted = set_contacted.count();
665
666 if (type == AccessType_Write) {
667 m_write_sharing_histogram.add(number_contacted);
668 } else {
669 m_read_sharing_histogram.add(number_contacted);
670 }
671 m_all_sharing_histogram.add(number_contacted);
672
673 if (number_contacted == 0) {
674 m_memory_to_cache++;
675 } else {
676 m_cache_to_cache++;
677 }
678
679}
680
681void Profiler::profileMsgDelay(int virtualNetwork, int delayCycles) {
682 assert(virtualNetwork < m_delayedCyclesVCHistograms.size());
683 m_delayedCyclesHistogram.add(delayCycles);
684 m_delayedCyclesVCHistograms[virtualNetwork].add(delayCycles);
685 if (virtualNetwork != 0) {
686 m_delayedCyclesNonPFHistogram.add(delayCycles);
687 }
688}
689
690// profiles original cache requests including PUTs
691void Profiler::profileRequest(const string& requestStr)
692{
693 m_requests++;
694
695 if (m_requestProfileMap_ptr->exist(requestStr)) {
696 (m_requestProfileMap_ptr->lookup(requestStr))++;
697 } else {
698 m_requestProfileMap_ptr->add(requestStr, 1);
699 }
700}
701
702void Profiler::startTransaction(int cpu)
703{
704 m_perProcStartTransaction[cpu]++;
705}
706
707void Profiler::endTransaction(int cpu)
708{
709 m_perProcEndTransaction[cpu]++;
710}
711
712void Profiler::controllerBusy(MachineID machID)
713{
714 m_busyControllerCount[(int)machID.type][(int)machID.num]++;
715}
716
717void Profiler::profilePFWait(Time waitTime)
718{
719 m_prefetchWaitHistogram.add(waitTime);
720}
721
722void Profiler::bankBusy()
723{
724 m_busyBankCount++;
725}
726
727// non-zero cycle demand request
728void Profiler::missLatency(Time t, RubyRequestType type)
729{
730 m_allMissLatencyHistogram.add(t);
731 m_missLatencyHistograms[type].add(t);
732}
733
734// non-zero cycle prefetch request
735void Profiler::swPrefetchLatency(Time t, CacheRequestType type, GenericMachineType respondingMach)
736{
737 m_allSWPrefetchLatencyHistogram.add(t);
738 m_SWPrefetchLatencyHistograms[type].add(t);
739 m_SWPrefetchMachLatencyHistograms[respondingMach].add(t);
740 if(respondingMach == GenericMachineType_Directory || respondingMach == GenericMachineType_NUM) {
741 m_SWPrefetchL2MissLatencyHistogram.add(t);
742 }
743}
744
745void Profiler::profileTransition(const string& component, NodeID version, Address addr,
746 const string& state, const string& event,
747 const string& next_state, const string& note)
748{
749 const int EVENT_SPACES = 20;
750 const int ID_SPACES = 3;
751 const int TIME_SPACES = 7;
752 const int COMP_SPACES = 10;
753 const int STATE_SPACES = 6;
754
755 if ((g_debug_ptr->getDebugTime() > 0) &&
756 (g_eventQueue_ptr->getTime() >= g_debug_ptr->getDebugTime())) {
757 (* debug_cout_ptr).flags(ios::right);
758 (* debug_cout_ptr) << setw(TIME_SPACES) << g_eventQueue_ptr->getTime() << " ";
759 (* debug_cout_ptr) << setw(ID_SPACES) << version << " ";
760 (* debug_cout_ptr) << setw(COMP_SPACES) << component;
761 (* debug_cout_ptr) << setw(EVENT_SPACES) << event << " ";
762
763 (* debug_cout_ptr).flags(ios::right);
764 (* debug_cout_ptr) << setw(STATE_SPACES) << state;
765 (* debug_cout_ptr) << ">";
766 (* debug_cout_ptr).flags(ios::left);
767 (* debug_cout_ptr) << setw(STATE_SPACES) << next_state;
768
769 (* debug_cout_ptr) << " " << addr << " " << note;
770
771 (* debug_cout_ptr) << endl;
772 }
773}
774
775// Helper function
776static double process_memory_total()
777{
778 const double MULTIPLIER = 4096.0/(1024.0*1024.0); // 4kB page size, 1024*1024 bytes per MB,
779 ifstream proc_file;
780 proc_file.open("/proc/self/statm");
781 int total_size_in_pages = 0;
782 int res_size_in_pages = 0;
783 proc_file >> total_size_in_pages;
784 proc_file >> res_size_in_pages;
785 return double(total_size_in_pages)*MULTIPLIER; // size in megabytes
786}
787
788static double process_memory_resident()
789{
790 const double MULTIPLIER = 4096.0/(1024.0*1024.0); // 4kB page size, 1024*1024 bytes per MB,
791 ifstream proc_file;
792 proc_file.open("/proc/self/statm");
793 int total_size_in_pages = 0;
794 int res_size_in_pages = 0;
795 proc_file >> total_size_in_pages;
796 proc_file >> res_size_in_pages;
797 return double(res_size_in_pages)*MULTIPLIER; // size in megabytes
798}
799
800void Profiler::rubyWatch(int id){
801 //int rn_g1 = 0;//SIMICS_get_register_number(id, "g1");
802 uint64 tr = 0;//SIMICS_read_register(id, rn_g1);
803 Address watch_address = Address(tr);
804 const int ID_SPACES = 3;
805 const int TIME_SPACES = 7;
806
807 (* debug_cout_ptr).flags(ios::right);
808 (* debug_cout_ptr) << setw(TIME_SPACES) << g_eventQueue_ptr->getTime() << " ";
809 (* debug_cout_ptr) << setw(ID_SPACES) << id << " "
810 << "RUBY WATCH "
811 << watch_address
812 << endl;
813
814 if(!m_watch_address_list_ptr->exist(watch_address)){
815 m_watch_address_list_ptr->add(watch_address, 1);
816 }
817}
818
819bool Profiler::watchAddress(Address addr){
820 if (m_watch_address_list_ptr->exist(addr))
821 return true;
822 else
823 return false;
824}
825
826int64 Profiler::getTotalTransactionsExecuted() const {
827 return m_perProcEndTransaction.sum();
828}
829
830// For MemoryControl:
831void Profiler::profileMemReq(string name, int bank) {
832// printf("name is %s", name.c_str());
833 assert(m_memory_control_profilers.count(name) == 1);
834 m_memory_control_profilers[name] -> m_memReq++;
835 m_memory_control_profilers[name] -> m_memBankCount[bank]++;
836}
837void Profiler::profileMemBankBusy(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memBankBusy++; }
838void Profiler::profileMemBusBusy(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memBusBusy++; }
839void Profiler::profileMemReadWriteBusy(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memReadWriteBusy++; }
840void Profiler::profileMemDataBusBusy(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memDataBusBusy++; }
841void Profiler::profileMemTfawBusy(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memTfawBusy++; }
842void Profiler::profileMemRefresh(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memRefresh++; }
843void Profiler::profileMemRead(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memRead++; }
844void Profiler::profileMemWrite(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memWrite++; }
845void Profiler::profileMemWaitCycles(string name, int cycles) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memWaitCycles += cycles; }
846void Profiler::profileMemInputQ(string name, int cycles) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memInputQ += cycles; }
847void Profiler::profileMemBankQ(string name, int cycles) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memBankQ += cycles; }
848void Profiler::profileMemArbWait(string name, int cycles) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memArbWait += cycles; }
849void Profiler::profileMemRandBusy(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memRandBusy++; }
850void Profiler::profileMemNotOld(string name) { assert(m_memory_control_profilers.count(name) == 1); m_memory_control_profilers[name] -> m_memNotOld++; }
851