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; --- 28 unchanged lines hidden (view full) --- 37 38 ---------------------------------------------------------------------- 39 40 File modification date: 2008-02-23 41 42 ---------------------------------------------------------------------- 43*/ 44 |
45// Allows use of times() library call, which determines virtual runtime 46#include <sys/resource.h> 47#include <sys/times.h> 48 |
49#include "mem/gems_common/Map.hh" |
50#include "mem/gems_common/PrioHeap.hh" |
51#include "mem/gems_common/util.hh" |
52#include "mem/protocol/CacheMsg.hh" |
53#include "mem/protocol/MachineType.hh" |
54#include "mem/protocol/Protocol.hh" |
55#include "mem/ruby/common/Debug.hh" |
56#include "mem/ruby/network/Network.hh" 57#include "mem/ruby/profiler/AddressProfiler.hh" 58#include "mem/ruby/profiler/Profiler.hh" |
59#include "mem/ruby/system/System.hh" |
60#include "mem/ruby/system/System.hh" |
61 |
62extern std::ostream* debug_cout_ptr; |
63 64static double process_memory_total(); 65static double process_memory_resident(); 66 67Profiler::Profiler(const Params *p) 68 : SimObject(p) 69{ |
70 m_requestProfileMap_ptr = new Map<string, int>; |
71 |
72 m_inst_profiler_ptr = NULL; 73 m_address_profiler_ptr = NULL; |
74 |
75 m_real_time_start_time = time(NULL); // Not reset in clearStats() 76 m_stats_period = 1000000; // Default 77 m_periodic_output_file_ptr = &cerr; |
78 |
79 m_hot_lines = p->hot_lines; 80 m_all_instructions = p->all_instructions; |
81 |
82 m_num_of_sequencers = p->num_of_sequencers; |
83 |
84 m_hot_lines = false; 85 m_all_instructions = false; |
86 |
87 m_address_profiler_ptr = new AddressProfiler(m_num_of_sequencers); 88 m_address_profiler_ptr->setHotLines(m_hot_lines); 89 m_address_profiler_ptr->setAllInstructions(m_all_instructions); |
90 |
91 if (m_all_instructions) { 92 m_inst_profiler_ptr = new AddressProfiler(m_num_of_sequencers); 93 m_inst_profiler_ptr->setHotLines(m_hot_lines); 94 m_inst_profiler_ptr->setAllInstructions(m_all_instructions); 95 } |
96} 97 98Profiler::~Profiler() 99{ |
100 if (m_periodic_output_file_ptr != &cerr) { 101 delete m_periodic_output_file_ptr; 102 } |
103 |
104 delete m_requestProfileMap_ptr; |
105} 106 |
107void 108Profiler::wakeup() |
109{ |
110 // FIXME - avoid the repeated code |
111 |
112 Vector<integer_t> perProcCycleCount; 113 perProcCycleCount.setSize(m_num_of_sequencers); |
114 |
115 for (int i = 0; i < m_num_of_sequencers; i++) { 116 perProcCycleCount[i] = 117 g_system_ptr->getCycleCount(i) - m_cycles_executed_at_start[i] + 1; 118 // The +1 allows us to avoid division by zero 119 } |
120 |
121 ostream &out = *m_periodic_output_file_ptr; |
122 |
123 out << "ruby_cycles: " << g_eventQueue_ptr->getTime()-m_ruby_start << endl 124 << "mbytes_resident: " << process_memory_resident() << endl 125 << "mbytes_total: " << process_memory_total() << endl; |
126 |
127 if (process_memory_total() > 0) { 128 out << "resident_ratio: " 129 << process_memory_resident() / process_memory_total() << endl; 130 } |
131 |
132 out << "miss_latency: " << m_allMissLatencyHistogram << endl; |
133 |
134 out << endl; |
135 |
136 if (m_all_instructions) { 137 m_inst_profiler_ptr->printStats(out); 138 } |
139 |
140 //g_system_ptr->getNetwork()->printStats(out); 141 g_eventQueue_ptr->scheduleEvent(this, m_stats_period); |
142} 143 |
144void 145Profiler::setPeriodicStatsFile(const string& filename) |
146{ |
147 cout << "Recording periodic statistics to file '" << filename << "' every " 148 << m_stats_period << " Ruby cycles" << endl; |
149 |
150 if (m_periodic_output_file_ptr != &cerr) { 151 delete m_periodic_output_file_ptr; 152 } |
153 |
154 m_periodic_output_file_ptr = new ofstream(filename.c_str()); 155 g_eventQueue_ptr->scheduleEvent(this, 1); |
156} 157 |
158void 159Profiler::setPeriodicStatsInterval(integer_t period) |
160{ |
161 cout << "Recording periodic statistics every " << m_stats_period 162 << " Ruby cycles" << endl; |
163 |
164 m_stats_period = period; 165 g_eventQueue_ptr->scheduleEvent(this, 1); |
166} 167 |
168void 169Profiler::printConfig(ostream& out) const |
170{ |
171 out << endl; 172 out << "Profiler Configuration" << endl; 173 out << "----------------------" << endl; 174 out << "periodic_stats_period: " << m_stats_period << endl; |
175} 176 |
177void 178Profiler::print(ostream& out) const |
179{ |
180 out << "[Profiler]"; |
181} 182 |
183void 184Profiler::printStats(ostream& out, bool short_stats) |
185{ |
186 out << endl; |
187 if (short_stats) { 188 out << "SHORT "; 189 } 190 out << "Profiler Stats" << endl; 191 out << "--------------" << endl; |
192 |
193 time_t real_time_current = time(NULL); 194 double seconds = difftime(real_time_current, m_real_time_start_time); 195 double minutes = seconds / 60.0; 196 double hours = minutes / 60.0; 197 double days = hours / 24.0; 198 Time ruby_cycles = g_eventQueue_ptr->getTime()-m_ruby_start; |
199 |
200 if (!short_stats) { 201 out << "Elapsed_time_in_seconds: " << seconds << endl; 202 out << "Elapsed_time_in_minutes: " << minutes << endl; 203 out << "Elapsed_time_in_hours: " << hours << endl; 204 out << "Elapsed_time_in_days: " << days << endl; 205 out << endl; |
206 } |
207 208 // print the virtual runtimes as well 209 struct tms vtime; 210 times(&vtime); 211 seconds = (vtime.tms_utime + vtime.tms_stime) / 100.0; 212 minutes = seconds / 60.0; 213 hours = minutes / 60.0; 214 days = hours / 24.0; 215 out << "Virtual_time_in_seconds: " << seconds << endl; 216 out << "Virtual_time_in_minutes: " << minutes << endl; 217 out << "Virtual_time_in_hours: " << hours << endl; 218 out << "Virtual_time_in_days: " << days << endl; |
219 out << endl; 220 |
221 out << "Ruby_current_time: " << g_eventQueue_ptr->getTime() << endl; 222 out << "Ruby_start_time: " << m_ruby_start << endl; 223 out << "Ruby_cycles: " << ruby_cycles << endl; 224 out << endl; |
225 |
226 if (!short_stats) { 227 out << "mbytes_resident: " << process_memory_resident() << endl; 228 out << "mbytes_total: " << process_memory_total() << endl; 229 if (process_memory_total() > 0) { 230 out << "resident_ratio: " 231 << process_memory_resident()/process_memory_total() << endl; 232 } 233 out << endl; 234 } |
235 |
236 Vector<integer_t> perProcCycleCount; 237 perProcCycleCount.setSize(m_num_of_sequencers); |
238 |
239 for (int i = 0; i < m_num_of_sequencers; i++) { 240 perProcCycleCount[i] = 241 g_system_ptr->getCycleCount(i) - m_cycles_executed_at_start[i] + 1; 242 // The +1 allows us to avoid division by zero 243 } |
244 |
245 out << "ruby_cycles_executed: " << perProcCycleCount << endl; |
246 |
247 out << endl; 248 |
249 if (!short_stats) { 250 out << "Busy Controller Counts:" << endl; 251 for (int i = 0; i < MachineType_NUM; i++) { 252 int size = MachineType_base_count((MachineType)i); 253 for (int j = 0; j < size; j++) { 254 MachineID machID; 255 machID.type = (MachineType)i; 256 machID.num = j; 257 out << machID << ":" << m_busyControllerCount[i][j] << " "; 258 if ((j + 1) % 8 == 0) { 259 out << endl; 260 } 261 } 262 out << endl; 263 } 264 out << endl; |
265 |
266 out << "Busy Bank Count:" << m_busyBankCount << endl; 267 out << endl; |
268 |
269 out << "sequencer_requests_outstanding: " 270 << m_sequencer_requests << endl; 271 out << endl; |
272 } |
273 |
274 if (!short_stats) { 275 out << "All Non-Zero Cycle Demand Cache Accesses" << endl; 276 out << "----------------------------------------" << endl; 277 out << "miss_latency: " << m_allMissLatencyHistogram << endl; 278 for (int i = 0; i < m_missLatencyHistograms.size(); i++) { 279 if (m_missLatencyHistograms[i].size() > 0) { 280 out << "miss_latency_" << RubyRequestType(i) << ": " 281 << m_missLatencyHistograms[i] << endl; 282 } 283 } 284 for (int i = 0; i < m_machLatencyHistograms.size(); i++) { 285 if (m_machLatencyHistograms[i].size() > 0) { 286 out << "miss_latency_" << GenericMachineType(i) << ": " 287 << m_machLatencyHistograms[i] << endl; 288 } 289 } |
290 |
291 out << endl; |
292 |
293 out << "All Non-Zero Cycle SW Prefetch Requests" << endl; 294 out << "------------------------------------" << endl; 295 out << "prefetch_latency: " << m_allSWPrefetchLatencyHistogram << endl; 296 for (int i = 0; i < m_SWPrefetchLatencyHistograms.size(); i++) { 297 if (m_SWPrefetchLatencyHistograms[i].size() > 0) { 298 out << "prefetch_latency_" << CacheRequestType(i) << ": " 299 << m_SWPrefetchLatencyHistograms[i] << endl; 300 } 301 } 302 for (int i = 0; i < m_SWPrefetchMachLatencyHistograms.size(); i++) { 303 if (m_SWPrefetchMachLatencyHistograms[i].size() > 0) { 304 out << "prefetch_latency_" << GenericMachineType(i) << ": " 305 << m_SWPrefetchMachLatencyHistograms[i] << endl; 306 } 307 } 308 out << "prefetch_latency_L2Miss:" 309 << m_SWPrefetchL2MissLatencyHistogram << endl; |
310 |
311 if (m_all_sharing_histogram.size() > 0) { 312 out << "all_sharing: " << m_all_sharing_histogram << endl; 313 out << "read_sharing: " << m_read_sharing_histogram << endl; 314 out << "write_sharing: " << m_write_sharing_histogram << endl; |
315 |
316 out << "all_sharing_percent: "; 317 m_all_sharing_histogram.printPercent(out); 318 out << endl; |
319 |
320 out << "read_sharing_percent: "; 321 m_read_sharing_histogram.printPercent(out); 322 out << endl; |
323 |
324 out << "write_sharing_percent: "; 325 m_write_sharing_histogram.printPercent(out); 326 out << endl; |
327 |
328 int64 total_miss = m_cache_to_cache + m_memory_to_cache; 329 out << "all_misses: " << total_miss << endl; 330 out << "cache_to_cache_misses: " << m_cache_to_cache << endl; 331 out << "memory_to_cache_misses: " << m_memory_to_cache << endl; 332 out << "cache_to_cache_percent: " 333 << 100.0 * (double(m_cache_to_cache) / double(total_miss)) 334 << endl; 335 out << "memory_to_cache_percent: " 336 << 100.0 * (double(m_memory_to_cache) / double(total_miss)) 337 << endl; 338 out << endl; 339 } |
340 |
341 if (m_outstanding_requests.size() > 0) { 342 out << "outstanding_requests: "; 343 m_outstanding_requests.printPercent(out); 344 out << endl; 345 out << endl; 346 } |
347 } |
348 |
349 if (!short_stats) { 350 out << "Request vs. RubySystem State Profile" << endl; 351 out << "--------------------------------" << endl; 352 out << endl; |
353 |
354 Vector<string> requestProfileKeys = m_requestProfileMap_ptr->keys(); 355 requestProfileKeys.sortVector(); |
356 |
357 for (int i = 0; i < requestProfileKeys.size(); i++) { 358 int temp_int = 359 m_requestProfileMap_ptr->lookup(requestProfileKeys[i]); 360 double percent = (100.0 * double(temp_int)) / double(m_requests); 361 while (requestProfileKeys[i] != "") { 362 out << setw(10) << string_split(requestProfileKeys[i], ':'); 363 } 364 out << setw(11) << temp_int; 365 out << setw(14) << percent << endl; 366 } 367 out << endl; |
368 |
369 out << "filter_action: " << m_filter_action_histogram << endl; |
370 |
371 if (!m_all_instructions) { 372 m_address_profiler_ptr->printStats(out); 373 } |
374 |
375 if (m_all_instructions) { 376 m_inst_profiler_ptr->printStats(out); 377 } 378 379 out << endl; 380 out << "Message Delayed Cycles" << endl; 381 out << "----------------------" << endl; 382 out << "Total_delay_cycles: " << m_delayedCyclesHistogram << endl; 383 out << "Total_nonPF_delay_cycles: " 384 << m_delayedCyclesNonPFHistogram << endl; 385 for (int i = 0; i < m_delayedCyclesVCHistograms.size(); i++) { 386 out << " virtual_network_" << i << "_delay_cycles: " 387 << m_delayedCyclesVCHistograms[i] << endl; 388 } 389 390 printResourceUsage(out); 391 } |
392} 393 |
394void 395Profiler::printResourceUsage(ostream& out) const |
396{ |
397 out << endl; 398 out << "Resource Usage" << endl; 399 out << "--------------" << endl; |
400 |
401 integer_t pagesize = getpagesize(); // page size in bytes 402 out << "page_size: " << pagesize << endl; |
403 |
404 rusage usage; 405 getrusage (RUSAGE_SELF, &usage); |
406 |
407 out << "user_time: " << usage.ru_utime.tv_sec << endl; 408 out << "system_time: " << usage.ru_stime.tv_sec << endl; 409 out << "page_reclaims: " << usage.ru_minflt << endl; 410 out << "page_faults: " << usage.ru_majflt << endl; 411 out << "swaps: " << usage.ru_nswap << endl; 412 out << "block_inputs: " << usage.ru_inblock << endl; 413 out << "block_outputs: " << usage.ru_oublock << endl; |
414} 415 |
416void 417Profiler::clearStats() |
418{ |
419 m_ruby_start = g_eventQueue_ptr->getTime(); |
420 |
421 m_cycles_executed_at_start.setSize(m_num_of_sequencers); 422 for (int i = 0; i < m_num_of_sequencers; i++) { 423 if (g_system_ptr == NULL) { 424 m_cycles_executed_at_start[i] = 0; 425 } else { 426 m_cycles_executed_at_start[i] = g_system_ptr->getCycleCount(i); 427 } |
428 } |
429 |
430 m_busyControllerCount.setSize(MachineType_NUM); // all machines 431 for (int i = 0; i < MachineType_NUM; i++) { 432 int size = MachineType_base_count((MachineType)i); 433 m_busyControllerCount[i].setSize(size); 434 for (int j = 0; j < size; j++) { 435 m_busyControllerCount[i][j] = 0; 436 } |
437 } |
438 m_busyBankCount = 0; |
439 |
440 m_delayedCyclesHistogram.clear(); 441 m_delayedCyclesNonPFHistogram.clear(); 442 int size = RubySystem::getNetwork()->getNumberOfVirtualNetworks(); 443 m_delayedCyclesVCHistograms.setSize(size); 444 for (int i = 0; i < size; i++) { 445 m_delayedCyclesVCHistograms[i].clear(); 446 } |
447 |
448 m_missLatencyHistograms.setSize(RubyRequestType_NUM); 449 for (int i = 0; i < m_missLatencyHistograms.size(); i++) { 450 m_missLatencyHistograms[i].clear(200); 451 } 452 m_machLatencyHistograms.setSize(GenericMachineType_NUM+1); 453 for (int i = 0; i < m_machLatencyHistograms.size(); i++) { 454 m_machLatencyHistograms[i].clear(200); 455 } 456 m_allMissLatencyHistogram.clear(200); |
457 |
458 m_SWPrefetchLatencyHistograms.setSize(CacheRequestType_NUM); 459 for (int i = 0; i < m_SWPrefetchLatencyHistograms.size(); i++) { 460 m_SWPrefetchLatencyHistograms[i].clear(200); 461 } 462 m_SWPrefetchMachLatencyHistograms.setSize(GenericMachineType_NUM+1); 463 for (int i = 0; i < m_SWPrefetchMachLatencyHistograms.size(); i++) { 464 m_SWPrefetchMachLatencyHistograms[i].clear(200); 465 } 466 m_allSWPrefetchLatencyHistogram.clear(200); |
467 |
468 m_sequencer_requests.clear(); 469 m_read_sharing_histogram.clear(); 470 m_write_sharing_histogram.clear(); 471 m_all_sharing_histogram.clear(); 472 m_cache_to_cache = 0; 473 m_memory_to_cache = 0; |
474 |
475 // clear HashMaps 476 m_requestProfileMap_ptr->clear(); |
477 |
478 // count requests profiled 479 m_requests = 0; |
480 |
481 m_outstanding_requests.clear(); 482 m_outstanding_persistent_requests.clear(); |
483 |
484 // Flush the prefetches through the system - used so that there 485 // are no outstanding requests after stats are cleared 486 //g_eventQueue_ptr->triggerAllEvents(); |
487 |
488 // update the start time 489 m_ruby_start = g_eventQueue_ptr->getTime(); |
490} 491 |
492void 493Profiler::addAddressTraceSample(const CacheMsg& msg, NodeID id) |
494{ |
495 if (msg.getType() != CacheRequestType_IFETCH) { 496 // Note: The following line should be commented out if you 497 // want to use the special profiling that is part of the GS320 498 // protocol |
499 |
500 // NOTE: Unless PROFILE_HOT_LINES is enabled, nothing will be 501 // profiled by the AddressProfiler 502 m_address_profiler_ptr-> 503 addTraceSample(msg.getLineAddress(), msg.getProgramCounter(), 504 msg.getType(), msg.getAccessMode(), id, false); 505 } |
506} 507 |
508void 509Profiler::profileSharing(const Address& addr, AccessType type, 510 NodeID requestor, const Set& sharers, 511 const Set& owner) |
512{ |
513 Set set_contacted(owner); 514 if (type == AccessType_Write) { 515 set_contacted.addSet(sharers); 516 } 517 set_contacted.remove(requestor); 518 int number_contacted = set_contacted.count(); |
519 |
520 if (type == AccessType_Write) { 521 m_write_sharing_histogram.add(number_contacted); 522 } else { 523 m_read_sharing_histogram.add(number_contacted); 524 } 525 m_all_sharing_histogram.add(number_contacted); |
526 |
527 if (number_contacted == 0) { 528 m_memory_to_cache++; 529 } else { 530 m_cache_to_cache++; 531 } |
532} 533 |
534void 535Profiler::profileMsgDelay(int virtualNetwork, int delayCycles) 536{ 537 assert(virtualNetwork < m_delayedCyclesVCHistograms.size()); 538 m_delayedCyclesHistogram.add(delayCycles); 539 m_delayedCyclesVCHistograms[virtualNetwork].add(delayCycles); 540 if (virtualNetwork != 0) { 541 m_delayedCyclesNonPFHistogram.add(delayCycles); 542 } |
543} 544 545// profiles original cache requests including PUTs |
546void 547Profiler::profileRequest(const string& requestStr) |
548{ |
549 m_requests++; |
550 |
551 if (m_requestProfileMap_ptr->exist(requestStr)) { 552 (m_requestProfileMap_ptr->lookup(requestStr))++; 553 } else { 554 m_requestProfileMap_ptr->add(requestStr, 1); 555 } |
556} 557 |
558void 559Profiler::controllerBusy(MachineID machID) |
560{ |
561 m_busyControllerCount[(int)machID.type][(int)machID.num]++; |
562} 563 |
564void 565Profiler::profilePFWait(Time waitTime) |
566{ |
567 m_prefetchWaitHistogram.add(waitTime); |
568} 569 |
570void 571Profiler::bankBusy() |
572{ |
573 m_busyBankCount++; |
574} 575 576// non-zero cycle demand request |
577void 578Profiler::missLatency(Time t, RubyRequestType type) |
579{ |
580 m_allMissLatencyHistogram.add(t); 581 m_missLatencyHistograms[type].add(t); |
582} 583 584// non-zero cycle prefetch request |
585void 586Profiler::swPrefetchLatency(Time t, CacheRequestType type, 587 GenericMachineType respondingMach) |
588{ |
589 m_allSWPrefetchLatencyHistogram.add(t); 590 m_SWPrefetchLatencyHistograms[type].add(t); 591 m_SWPrefetchMachLatencyHistograms[respondingMach].add(t); 592 if (respondingMach == GenericMachineType_Directory || 593 respondingMach == GenericMachineType_NUM) { 594 m_SWPrefetchL2MissLatencyHistogram.add(t); 595 } |
596} 597 |
598void 599Profiler::profileTransition(const string& component, NodeID version, 600 Address addr, const string& state, const string& event, 601 const string& next_state, const string& note) |
602{ |
603 const int EVENT_SPACES = 20; 604 const int ID_SPACES = 3; 605 const int TIME_SPACES = 7; 606 const int COMP_SPACES = 10; 607 const int STATE_SPACES = 6; |
608 |
609 if (g_debug_ptr->getDebugTime() <= 0 || 610 g_eventQueue_ptr->getTime() < g_debug_ptr->getDebugTime()) 611 return; |
612 |
613 ostream &out = *debug_cout_ptr; 614 out.flags(ios::right); 615 out << setw(TIME_SPACES) << g_eventQueue_ptr->getTime() << " "; 616 out << setw(ID_SPACES) << version << " "; 617 out << setw(COMP_SPACES) << component; 618 out << setw(EVENT_SPACES) << event << " "; |
619 |
620 out.flags(ios::right); 621 out << setw(STATE_SPACES) << state; 622 out << ">"; 623 out.flags(ios::left); 624 out << setw(STATE_SPACES) << next_state; |
625 |
626 out << " " << addr << " " << note; 627 628 out << endl; |
629} 630 631// Helper function |
632static double 633process_memory_total() |
634{ |
635 // 4kB page size, 1024*1024 bytes per MB, 636 const double MULTIPLIER = 4096.0 / (1024.0 * 1024.0); 637 ifstream proc_file; 638 proc_file.open("/proc/self/statm"); 639 int total_size_in_pages = 0; 640 int res_size_in_pages = 0; 641 proc_file >> total_size_in_pages; 642 proc_file >> res_size_in_pages; 643 return double(total_size_in_pages) * MULTIPLIER; // size in megabytes |
644} 645 |
646static double 647process_memory_resident() |
648{ |
649 // 4kB page size, 1024*1024 bytes per MB, 650 const double MULTIPLIER = 4096.0 / (1024.0 * 1024.0); 651 ifstream proc_file; 652 proc_file.open("/proc/self/statm"); 653 int total_size_in_pages = 0; 654 int res_size_in_pages = 0; 655 proc_file >> total_size_in_pages; 656 proc_file >> res_size_in_pages; 657 return double(res_size_in_pages) * MULTIPLIER; // size in megabytes |
658} 659 |
660void 661Profiler::rubyWatch(int id) 662{ |
663 uint64 tr = 0; 664 Address watch_address = Address(tr); 665 const int ID_SPACES = 3; 666 const int TIME_SPACES = 7; 667 |
668 ostream &out = *debug_cout_ptr; |
669 |
670 out.flags(ios::right); 671 out << setw(TIME_SPACES) << g_eventQueue_ptr->getTime() << " "; 672 out << setw(ID_SPACES) << id << " " 673 << "RUBY WATCH " << watch_address << endl; 674 675 if (!m_watch_address_list_ptr->exist(watch_address)) { 676 m_watch_address_list_ptr->add(watch_address, 1); |
677 } 678} 679 |
680bool 681Profiler::watchAddress(Address addr) 682{ |
683 if (m_watch_address_list_ptr->exist(addr)) |
684 return true; |
685 else |
686 return false; |
687} 688 689Profiler * 690RubyProfilerParams::create() 691{ 692 return new Profiler(this); 693} |