/*
- * Copyright (c) 1999-2008 Mark D. Hill and David A. Wood
+ * Copyright (c) 1999-2013 Mark D. Hill and David A. Wood
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
----------------------------------------------------------------------
*/
-/*
- * Profiler.cc
- *
- * Description: See Profiler.hh
- *
- * $Id$
- *
- */
-
#include "mem/ruby/profiler/Profiler.hh"
-#include "mem/ruby/profiler/AddressProfiler.hh"
-#include "mem/ruby/system/System.hh"
-#include "mem/ruby/network/Network.hh"
-#include "mem/gems_common/PrioHeap.hh"
-#include "mem/protocol/CacheMsg.hh"
-#include "mem/protocol/Protocol.hh"
-#include "mem/gems_common/util.hh"
-#include "mem/gems_common/Map.hh"
-#include "mem/ruby/common/Debug.hh"
-#include "mem/protocol/MachineType.hh"
-#include "mem/ruby/system/System.hh"
+#include <sys/types.h>
+#include <unistd.h>
-// Allows use of times() library call, which determines virtual runtime
-#include <sys/times.h>
+#include <algorithm>
+#include <fstream>
-extern std::ostream * debug_cout_ptr;
-
-static double process_memory_total();
-static double process_memory_resident();
-
-Profiler::Profiler(const Params *p)
- : SimObject(p)
-{
- m_requestProfileMap_ptr = new Map<string, int>;
-
- m_inst_profiler_ptr = NULL;
- m_address_profiler_ptr = NULL;
-
- m_real_time_start_time = time(NULL); // Not reset in clearStats()
- m_stats_period = 1000000; // Default
- m_periodic_output_file_ptr = &cerr;
+#include "base/stl_helpers.hh"
+#include "base/str.hh"
+#include "mem/ruby/network/Network.hh"
+#include "mem/ruby/profiler/AddressProfiler.hh"
+#include "mem/ruby/protocol/MachineType.hh"
+#include "mem/ruby/protocol/RubyRequest.hh"
+
+/**
+ * the profiler uses GPUCoalescer code even
+ * though the GPUCoalescer is not built for
+ * all ISAs, which can lead to run/link time
+ * errors. here we guard the coalescer code
+ * with ifdefs as there is no easy way to
+ * refactor this code without removing
+ * GPUCoalescer stats from the profiler.
+ *
+ * eventually we should use probe points
+ * here, but until then these ifdefs will
+ * serve.
+ */
+#ifdef BUILD_GPU
+#include "mem/ruby/system/GPUCoalescer.hh"
- m_hot_lines = p->hot_lines;
- m_all_instructions = p->all_instructions;
+#endif
- m_num_of_sequencers = p->num_of_sequencers;
+#include "mem/ruby/system/Sequencer.hh"
- m_hot_lines = false;
- m_all_instructions = false;
+using namespace std;
+using m5::stl_helpers::operator<<;
- m_address_profiler_ptr = new AddressProfiler(m_num_of_sequencers);
- m_address_profiler_ptr -> setHotLines(m_hot_lines);
- m_address_profiler_ptr -> setAllInstructions(m_all_instructions);
+Profiler::Profiler(const RubySystemParams &p, RubySystem *rs)
+ : m_ruby_system(rs), m_hot_lines(p.hot_lines),
+ m_all_instructions(p.all_instructions),
+ m_num_vnets(p.number_of_virtual_networks),
+ rubyProfilerStats(rs, this)
+{
+ m_address_profiler_ptr = new AddressProfiler(p.num_of_sequencers, this);
+ m_address_profiler_ptr->setHotLines(m_hot_lines);
+ m_address_profiler_ptr->setAllInstructions(m_all_instructions);
- if (m_all_instructions) {
- m_inst_profiler_ptr = new AddressProfiler(m_num_of_sequencers);
- m_inst_profiler_ptr -> setHotLines(m_hot_lines);
- m_inst_profiler_ptr -> setAllInstructions(m_all_instructions);
- }
+ if (m_all_instructions) {
+ m_inst_profiler_ptr = new AddressProfiler(p.num_of_sequencers, this);
+ m_inst_profiler_ptr->setHotLines(m_hot_lines);
+ m_inst_profiler_ptr->setAllInstructions(m_all_instructions);
+ }
}
Profiler::~Profiler()
{
- if (m_periodic_output_file_ptr != &cerr) {
- delete m_periodic_output_file_ptr;
- }
-
- delete m_requestProfileMap_ptr;
}
-void Profiler::wakeup()
+Profiler::
+ProfilerStats::ProfilerStats(Stats::Group *parent, Profiler *profiler)
+ : Stats::Group(parent),
+ perRequestTypeStats(parent),
+ perMachineTypeStats(parent),
+ perRequestTypeMachineTypeStats(parent),
+ ADD_STAT(delayHistogram, "delay histogram for all message"),
+ ADD_STAT(m_outstandReqHistSeqr, ""),
+ ADD_STAT(m_outstandReqHistCoalsr, ""),
+ ADD_STAT(m_latencyHistSeqr, ""),
+ ADD_STAT(m_latencyHistCoalsr, ""),
+ ADD_STAT(m_hitLatencyHistSeqr, ""),
+ ADD_STAT(m_missLatencyHistSeqr, ""),
+ ADD_STAT(m_missLatencyHistCoalsr, "")
{
- // FIXME - avoid the repeated code
-
- Vector<integer_t> perProcCycleCount;
- perProcCycleCount.setSize(m_num_of_sequencers);
-
- for(int i=0; i < m_num_of_sequencers; i++) {
- perProcCycleCount[i] = g_system_ptr->getCycleCount(i) - m_cycles_executed_at_start[i] + 1;
- // The +1 allows us to avoid division by zero
- }
-
- integer_t total_misses = m_perProcTotalMisses.sum();
- integer_t simics_cycles_executed = perProcCycleCount.sum();
- integer_t transactions_started = m_perProcStartTransaction.sum();
- integer_t transactions_ended = m_perProcEndTransaction.sum();
-
- (*m_periodic_output_file_ptr) << "ruby_cycles: "
- << g_eventQueue_ptr->getTime()-m_ruby_start
- << endl;
-
- (*m_periodic_output_file_ptr) << "total_misses: "
- << total_misses
- << " "
- << m_perProcTotalMisses
- << endl;
-
- (*m_periodic_output_file_ptr) << "simics_cycles_executed: "
- << simics_cycles_executed
- << " "
- << perProcCycleCount
- << endl;
-
- (*m_periodic_output_file_ptr) << "transactions_started: "
- << transactions_started
- << " "
- << m_perProcStartTransaction
- << endl;
-
- (*m_periodic_output_file_ptr) << "transactions_ended: "
- << transactions_ended
- << " "
- << m_perProcEndTransaction
- << endl;
-
- (*m_periodic_output_file_ptr) << "mbytes_resident: "
- << process_memory_resident()
- << endl;
-
- (*m_periodic_output_file_ptr) << "mbytes_total: "
- << process_memory_total()
- << endl;
-
- if (process_memory_total() > 0) {
- (*m_periodic_output_file_ptr) << "resident_ratio: "
- << process_memory_resident()/process_memory_total()
- << endl;
- }
-
- (*m_periodic_output_file_ptr) << "miss_latency: "
- << m_allMissLatencyHistogram
- << endl;
-
- *m_periodic_output_file_ptr << endl;
-
- if (m_all_instructions) {
- m_inst_profiler_ptr->printStats(*m_periodic_output_file_ptr);
- }
-
- //g_system_ptr->getNetwork()->printStats(*m_periodic_output_file_ptr);
- g_eventQueue_ptr->scheduleEvent(this, m_stats_period);
-}
+ delayHistogram
+ .init(10)
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ for (int i = 0; i < profiler->m_num_vnets; i++) {
+ delayVCHistogram.push_back(new Stats::Histogram(this));
+ delayVCHistogram[i]
+ ->init(10)
+ .name(csprintf("delayVCHist.vnet_%i", i))
+ .desc(csprintf("delay histogram for vnet_%i", i))
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+ }
-void Profiler::setPeriodicStatsFile(const string& filename)
-{
- cout << "Recording periodic statistics to file '" << filename << "' every "
- << m_stats_period << " Ruby cycles" << endl;
+ m_outstandReqHistSeqr
+ .init(10)
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
- if (m_periodic_output_file_ptr != &cerr) {
- delete m_periodic_output_file_ptr;
- }
+ m_outstandReqHistCoalsr
+ .init(10)
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
- m_periodic_output_file_ptr = new ofstream(filename.c_str());
- g_eventQueue_ptr->scheduleEvent(this, 1);
-}
+ m_latencyHistSeqr
+ .init(10)
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
-void Profiler::setPeriodicStatsInterval(integer_t period)
-{
- cout << "Recording periodic statistics every " << m_stats_period
- << " Ruby cycles" << endl;
+ m_latencyHistCoalsr
+ .init(10)
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
- m_stats_period = period;
- g_eventQueue_ptr->scheduleEvent(this, 1);
-}
+ m_hitLatencyHistSeqr
+ .init(10)
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
-void Profiler::printConfig(ostream& out) const
-{
- out << endl;
- out << "Profiler Configuration" << endl;
- out << "----------------------" << endl;
- out << "periodic_stats_period: " << m_stats_period << endl;
+ m_missLatencyHistSeqr
+ .init(10)
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_missLatencyHistCoalsr
+ .init(10)
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
}
-void Profiler::print(ostream& out) const
+Profiler::ProfilerStats::
+PerRequestTypeStats::PerRequestTypeStats(Stats::Group *parent)
+ : Stats::Group(parent, "RequestType")
{
- out << "[Profiler]";
+ for (int i = 0; i < RubyRequestType_NUM; i++) {
+ m_typeLatencyHistSeqr.push_back(new Stats::Histogram(this));
+ m_typeLatencyHistSeqr[i]
+ ->init(10)
+ .name(csprintf("%s.latency_hist_seqr", RubyRequestType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_typeLatencyHistCoalsr.push_back(new Stats::Histogram(this));
+ m_typeLatencyHistCoalsr[i]
+ ->init(10)
+ .name(csprintf("%s.latency_hist_coalsr", RubyRequestType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_hitTypeLatencyHistSeqr.push_back(new Stats::Histogram(this));
+ m_hitTypeLatencyHistSeqr[i]
+ ->init(10)
+ .name(csprintf("%s.hit_latency_hist_seqr", RubyRequestType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_missTypeLatencyHistSeqr.push_back(new Stats::Histogram(this));
+ m_missTypeLatencyHistSeqr[i]
+ ->init(10)
+ .name(csprintf("%s.miss_latency_hist_seqr", RubyRequestType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_missTypeLatencyHistCoalsr.push_back(new Stats::Histogram(this));
+ m_missTypeLatencyHistCoalsr[i]
+ ->init(10)
+ .name(csprintf("%s.miss_latency_hist_coalsr", RubyRequestType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+ }
}
-void Profiler::printStats(ostream& out, bool short_stats)
+Profiler::ProfilerStats::
+PerMachineTypeStats::PerMachineTypeStats(Stats::Group *parent)
+ : Stats::Group(parent, "MachineType")
{
- out << endl;
- if (short_stats) {
- out << "SHORT ";
- }
- out << "Profiler Stats" << endl;
- out << "--------------" << endl;
-
- time_t real_time_current = time(NULL);
- double seconds = difftime(real_time_current, m_real_time_start_time);
- double minutes = seconds/60.0;
- double hours = minutes/60.0;
- double days = hours/24.0;
- Time ruby_cycles = g_eventQueue_ptr->getTime()-m_ruby_start;
-
- if (!short_stats) {
- out << "Elapsed_time_in_seconds: " << seconds << endl;
- out << "Elapsed_time_in_minutes: " << minutes << endl;
- out << "Elapsed_time_in_hours: " << hours << endl;
- out << "Elapsed_time_in_days: " << days << endl;
- out << endl;
- }
-
- // print the virtual runtimes as well
- struct tms vtime;
- times(&vtime);
- seconds = (vtime.tms_utime + vtime.tms_stime) / 100.0;
- minutes = seconds / 60.0;
- hours = minutes / 60.0;
- days = hours / 24.0;
- out << "Virtual_time_in_seconds: " << seconds << endl;
- out << "Virtual_time_in_minutes: " << minutes << endl;
- out << "Virtual_time_in_hours: " << hours << endl;
- out << "Virtual_time_in_days: " << days << endl;
- out << endl;
-
- out << "Ruby_current_time: " << g_eventQueue_ptr->getTime() << endl;
- out << "Ruby_start_time: " << m_ruby_start << endl;
- out << "Ruby_cycles: " << ruby_cycles << endl;
- out << endl;
-
- if (!short_stats) {
- out << "mbytes_resident: " << process_memory_resident() << endl;
- out << "mbytes_total: " << process_memory_total() << endl;
- if (process_memory_total() > 0) {
- out << "resident_ratio: "
- << process_memory_resident()/process_memory_total() << endl;
+ for (int i = 0; i < MachineType_NUM; i++) {
+ m_hitMachLatencyHistSeqr.push_back(new Stats::Histogram(this));
+ m_hitMachLatencyHistSeqr[i]
+ ->init(10)
+ .name(csprintf("%s.hit_mach_latency_hist_seqr", MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_missMachLatencyHistSeqr.push_back(new Stats::Histogram(this));
+ m_missMachLatencyHistSeqr[i]
+ ->init(10)
+ .name(csprintf("%s.miss_mach_latency_hist_seqr", MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_missMachLatencyHistCoalsr.push_back(new Stats::Histogram(this));
+ m_missMachLatencyHistCoalsr[i]
+ ->init(10)
+ .name(csprintf("%s.miss_mach_latency_hist_coalsr",
+ MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_IssueToInitialDelayHistSeqr.push_back(new Stats::Histogram(this));
+ m_IssueToInitialDelayHistSeqr[i]
+ ->init(10)
+ .name(csprintf(
+ "%s.miss_latency_hist_seqr.issue_to_initial_request",
+ MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_IssueToInitialDelayHistCoalsr.push_back(new Stats::Histogram(this));
+ m_IssueToInitialDelayHistCoalsr[i]
+ ->init(10)
+ .name(csprintf(
+ "%s.miss_latency_hist_coalsr.issue_to_initial_request",
+ MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_InitialToForwardDelayHistSeqr.push_back(new Stats::Histogram(this));
+ m_InitialToForwardDelayHistSeqr[i]
+ ->init(10)
+ .name(csprintf("%s.miss_latency_hist_seqr.initial_to_forward",
+ MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_InitialToForwardDelayHistCoalsr
+ .push_back(new Stats::Histogram(this));
+ m_InitialToForwardDelayHistCoalsr[i]
+ ->init(10)
+ .name(csprintf("%s.miss_latency_hist_coalsr.initial_to_forward",
+ MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_ForwardToFirstResponseDelayHistSeqr
+ .push_back(new Stats::Histogram(this));
+
+ m_ForwardToFirstResponseDelayHistSeqr[i]
+ ->init(10)
+ .name(csprintf(
+ "%s.miss_latency_hist_seqr.forward_to_first_response",
+ MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_ForwardToFirstResponseDelayHistCoalsr
+ .push_back(new Stats::Histogram(this));
+ m_ForwardToFirstResponseDelayHistCoalsr[i]
+ ->init(10)
+ .name(csprintf(
+ "%s.miss_latency_hist_coalsr.forward_to_first_response",
+ MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_FirstResponseToCompletionDelayHistSeqr
+ .push_back(new Stats::Histogram(this));
+ m_FirstResponseToCompletionDelayHistSeqr[i]
+ ->init(10)
+ .name(csprintf(
+ "%s.miss_latency_hist_seqr.first_response_to_completion",
+ MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_FirstResponseToCompletionDelayHistCoalsr
+ .push_back(new Stats::Histogram(this));
+ m_FirstResponseToCompletionDelayHistCoalsr[i]
+ ->init(10)
+ .name(csprintf(
+ "%s.miss_latency_hist_coalsr.first_response_to_completion",
+ MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_IncompleteTimesSeqr.push_back(new Stats::Scalar(this));
+ m_IncompleteTimesSeqr[i]
+ ->name(csprintf("%s.incomplete_times_seqr", MachineType(i)))
+ .desc("")
+ .flags(Stats::nozero);
}
- out << endl;
-
- }
-
- Vector<integer_t> perProcCycleCount;
- Vector<double> perProcCyclesPerTrans;
- Vector<double> perProcMissesPerTrans;
-
-
- perProcCycleCount.setSize(m_num_of_sequencers);
- perProcCyclesPerTrans.setSize(m_num_of_sequencers);
- perProcMissesPerTrans.setSize(m_num_of_sequencers);
-
- for(int i=0; i < m_num_of_sequencers; i++) {
- perProcCycleCount[i] = g_system_ptr->getCycleCount(i) - m_cycles_executed_at_start[i] + 1;
- // The +1 allows us to avoid division by zero
+}
- int trans = m_perProcEndTransaction[i];
- if (trans == 0) {
- perProcCyclesPerTrans[i] = 0;
- perProcMissesPerTrans[i] = 0;
- } else {
- perProcCyclesPerTrans[i] = ruby_cycles / double(trans);
- perProcMissesPerTrans[i] = m_perProcTotalMisses[i] / double(trans);
- }
- }
-
- integer_t total_misses = m_perProcTotalMisses.sum();
- integer_t user_misses = m_perProcUserMisses.sum();
- integer_t supervisor_misses = m_perProcSupervisorMisses.sum();
- integer_t simics_cycles_executed = perProcCycleCount.sum();
- integer_t transactions_started = m_perProcStartTransaction.sum();
- integer_t transactions_ended = m_perProcEndTransaction.sum();
-
- double cycles_per_transaction = (transactions_ended != 0) ? (m_num_of_sequencers * double(ruby_cycles)) / double(transactions_ended) : 0;
- double misses_per_transaction = (transactions_ended != 0) ? double(total_misses) / double(transactions_ended) : 0;
-
- out << "Total_misses: " << total_misses << endl;
- out << "total_misses: " << total_misses << " " << m_perProcTotalMisses << endl;
- out << "user_misses: " << user_misses << " " << m_perProcUserMisses << endl;
- out << "supervisor_misses: " << supervisor_misses << " " << m_perProcSupervisorMisses << endl;
- out << endl;
- out << "ruby_cycles_executed: " << simics_cycles_executed << " " << perProcCycleCount << endl;
- out << endl;
- out << "transactions_started: " << transactions_started << " " << m_perProcStartTransaction << endl;
- out << "transactions_ended: " << transactions_ended << " " << m_perProcEndTransaction << endl;
- out << "cycles_per_transaction: " << cycles_per_transaction << " " << perProcCyclesPerTrans << endl;
- out << "misses_per_transaction: " << misses_per_transaction << " " << perProcMissesPerTrans << endl;
-
- out << endl;
-
- out << endl;
-
- if (!short_stats) {
- out << "Busy Controller Counts:" << endl;
- for(int i=0; i < MachineType_NUM; i++) {
- for(int j=0; j < MachineType_base_count((MachineType)i); j++) {
- MachineID machID;
- machID.type = (MachineType)i;
- machID.num = j;
- out << machID << ":" << m_busyControllerCount[i][j] << " ";
- if ((j+1)%8 == 0) {
- out << endl;
+Profiler::ProfilerStats::
+PerRequestTypeMachineTypeStats::
+PerRequestTypeMachineTypeStats(Stats::Group *parent)
+ : Stats::Group(parent, "RequestTypeMachineType")
+{
+ for (int i = 0; i < RubyRequestType_NUM; i++) {
+ m_hitTypeMachLatencyHistSeqr
+ .push_back(std::vector<Stats::Histogram *>());
+ m_missTypeMachLatencyHistSeqr
+ .push_back(std::vector<Stats::Histogram *>());
+ m_missTypeMachLatencyHistCoalsr
+ .push_back(std::vector<Stats::Histogram *>());
+
+ for (int j = 0; j < MachineType_NUM; j++) {
+ m_hitTypeMachLatencyHistSeqr[i]
+ .push_back(new Stats::Histogram(this));
+ m_hitTypeMachLatencyHistSeqr[i][j]
+ ->init(10)
+ .name(csprintf("%s.%s.hit_type_mach_latency_hist_seqr",
+ RubyRequestType(i), MachineType(j)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_missTypeMachLatencyHistSeqr[i]
+ .push_back(new Stats::Histogram(this));
+ m_missTypeMachLatencyHistSeqr[i][j]
+ ->init(10)
+ .name(csprintf("%s.%s.miss_type_mach_latency_hist_seqr",
+ RubyRequestType(i), MachineType(j)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
+
+ m_missTypeMachLatencyHistCoalsr[i]
+ .push_back(new Stats::Histogram(this));
+ m_missTypeMachLatencyHistCoalsr[i][j]
+ ->init(10)
+ .name(csprintf("%s.%s.miss_type_mach_latency_hist_coalsr",
+ RubyRequestType(i), MachineType(j)))
+ .desc("")
+ .flags(Stats::nozero | Stats::pdf | Stats::oneline);
}
- }
- out << endl;
- }
- out << endl;
-
- out << "Busy Bank Count:" << m_busyBankCount << endl;
- out << endl;
-
- out << "sequencer_requests_outstanding: " << m_sequencer_requests << endl;
- out << endl;
- }
-
- if (!short_stats) {
- out << "All Non-Zero Cycle Demand Cache Accesses" << endl;
- out << "----------------------------------------" << endl;
- out << "miss_latency: " << m_allMissLatencyHistogram << endl;
- for(int i=0; i<m_missLatencyHistograms.size(); i++) {
- if (m_missLatencyHistograms[i].size() > 0) {
- out << "miss_latency_" << RubyRequestType(i) << ": " << m_missLatencyHistograms[i] << endl;
- }
- }
- for(int i=0; i<m_machLatencyHistograms.size(); i++) {
- if (m_machLatencyHistograms[i].size() > 0) {
- out << "miss_latency_" << GenericMachineType(i) << ": " << m_machLatencyHistograms[i] << endl;
- }
- }
-
- out << endl;
-
- out << "All Non-Zero Cycle SW Prefetch Requests" << endl;
- out << "------------------------------------" << endl;
- out << "prefetch_latency: " << m_allSWPrefetchLatencyHistogram << endl;
- for(int i=0; i<m_SWPrefetchLatencyHistograms.size(); i++) {
- if (m_SWPrefetchLatencyHistograms[i].size() > 0) {
- out << "prefetch_latency_" << CacheRequestType(i) << ": " << m_SWPrefetchLatencyHistograms[i] << endl;
- }
}
- for(int i=0; i<m_SWPrefetchMachLatencyHistograms.size(); i++) {
- if (m_SWPrefetchMachLatencyHistograms[i].size() > 0) {
- out << "prefetch_latency_" << GenericMachineType(i) << ": " << m_SWPrefetchMachLatencyHistograms[i] << endl;
- }
- }
- out << "prefetch_latency_L2Miss:" << m_SWPrefetchL2MissLatencyHistogram << endl;
-
- if (m_all_sharing_histogram.size() > 0) {
- out << "all_sharing: " << m_all_sharing_histogram << endl;
- out << "read_sharing: " << m_read_sharing_histogram << endl;
- out << "write_sharing: " << m_write_sharing_histogram << endl;
-
- out << "all_sharing_percent: "; m_all_sharing_histogram.printPercent(out); out << endl;
- out << "read_sharing_percent: "; m_read_sharing_histogram.printPercent(out); out << endl;
- out << "write_sharing_percent: "; m_write_sharing_histogram.printPercent(out); out << endl;
-
- int64 total_miss = m_cache_to_cache + m_memory_to_cache;
- out << "all_misses: " << total_miss << endl;
- out << "cache_to_cache_misses: " << m_cache_to_cache << endl;
- out << "memory_to_cache_misses: " << m_memory_to_cache << endl;
- out << "cache_to_cache_percent: " << 100.0 * (double(m_cache_to_cache) / double(total_miss)) << endl;
- out << "memory_to_cache_percent: " << 100.0 * (double(m_memory_to_cache) / double(total_miss)) << endl;
- out << endl;
- }
-
- if (m_outstanding_requests.size() > 0) {
- out << "outstanding_requests: "; m_outstanding_requests.printPercent(out); out << endl;
- out << endl;
- }
- }
-
- if (!short_stats) {
- out << "Request vs. RubySystem State Profile" << endl;
- out << "--------------------------------" << endl;
- out << endl;
-
- Vector<string> requestProfileKeys = m_requestProfileMap_ptr->keys();
- requestProfileKeys.sortVector();
-
- for(int i=0; i<requestProfileKeys.size(); i++) {
- int temp_int = m_requestProfileMap_ptr->lookup(requestProfileKeys[i]);
- double percent = (100.0*double(temp_int))/double(m_requests);
- while (requestProfileKeys[i] != "") {
- out << setw(10) << string_split(requestProfileKeys[i], ':');
- }
- out << setw(11) << temp_int;
- out << setw(14) << percent << endl;
- }
- out << endl;
-
- out << "filter_action: " << m_filter_action_histogram << endl;
+}
+void
+Profiler::collateStats()
+{
if (!m_all_instructions) {
- m_address_profiler_ptr->printStats(out);
+ m_address_profiler_ptr->collateStats();
}
if (m_all_instructions) {
- m_inst_profiler_ptr->printStats(out);
- }
-
- out << endl;
- out << "Message Delayed Cycles" << endl;
- out << "----------------------" << endl;
- out << "Total_delay_cycles: " << m_delayedCyclesHistogram << endl;
- out << "Total_nonPF_delay_cycles: " << m_delayedCyclesNonPFHistogram << endl;
- for (int i = 0; i < m_delayedCyclesVCHistograms.size(); i++) {
- out << " virtual_network_" << i << "_delay_cycles: " << m_delayedCyclesVCHistograms[i] << endl;
+ m_inst_profiler_ptr->collateStats();
}
- printResourceUsage(out);
- }
+ for (uint32_t i = 0; i < MachineType_NUM; i++) {
+ for (map<uint32_t, AbstractController*>::iterator it =
+ m_ruby_system->m_abstract_controls[i].begin();
+ it != m_ruby_system->m_abstract_controls[i].end(); ++it) {
-}
-
-void Profiler::printResourceUsage(ostream& out) const
-{
- out << endl;
- out << "Resource Usage" << endl;
- out << "--------------" << endl;
-
- integer_t pagesize = getpagesize(); // page size in bytes
- out << "page_size: " << pagesize << endl;
-
- rusage usage;
- getrusage (RUSAGE_SELF, &usage);
-
- out << "user_time: " << usage.ru_utime.tv_sec << endl;
- out << "system_time: " << usage.ru_stime.tv_sec << endl;
- out << "page_reclaims: " << usage.ru_minflt << endl;
- out << "page_faults: " << usage.ru_majflt << endl;
- out << "swaps: " << usage.ru_nswap << endl;
- out << "block_inputs: " << usage.ru_inblock << endl;
- out << "block_outputs: " << usage.ru_oublock << endl;
-}
+ AbstractController *ctr = (*it).second;
+ rubyProfilerStats.delayHistogram.add(ctr->getDelayHist());
-void Profiler::clearStats()
-{
- m_ruby_start = g_eventQueue_ptr->getTime();
-
- m_cycles_executed_at_start.setSize(m_num_of_sequencers);
- for (int i=0; i < m_num_of_sequencers; i++) {
- if (g_system_ptr == NULL) {
- m_cycles_executed_at_start[i] = 0;
- } else {
- m_cycles_executed_at_start[i] = g_system_ptr->getCycleCount(i);
- }
- }
-
- m_perProcTotalMisses.setSize(m_num_of_sequencers);
- m_perProcUserMisses.setSize(m_num_of_sequencers);
- m_perProcSupervisorMisses.setSize(m_num_of_sequencers);
- m_perProcStartTransaction.setSize(m_num_of_sequencers);
- m_perProcEndTransaction.setSize(m_num_of_sequencers);
-
- for(int i=0; i < m_num_of_sequencers; i++) {
- m_perProcTotalMisses[i] = 0;
- m_perProcUserMisses[i] = 0;
- m_perProcSupervisorMisses[i] = 0;
- m_perProcStartTransaction[i] = 0;
- m_perProcEndTransaction[i] = 0;
- }
-
- m_busyControllerCount.setSize(MachineType_NUM); // all machines
- for(int i=0; i < MachineType_NUM; i++) {
- m_busyControllerCount[i].setSize(MachineType_base_count((MachineType)i));
- for(int j=0; j < MachineType_base_count((MachineType)i); j++) {
- m_busyControllerCount[i][j] = 0;
+ for (uint32_t i = 0; i < m_num_vnets; i++) {
+ rubyProfilerStats.
+ delayVCHistogram[i]->add(ctr->getDelayVCHist(i));
+ }
+ }
}
- }
- m_busyBankCount = 0;
-
- m_delayedCyclesHistogram.clear();
- m_delayedCyclesNonPFHistogram.clear();
- m_delayedCyclesVCHistograms.setSize(RubySystem::getNetwork()->getNumberOfVirtualNetworks());
- for (int i = 0; i < RubySystem::getNetwork()->getNumberOfVirtualNetworks(); i++) {
- m_delayedCyclesVCHistograms[i].clear();
- }
-
- m_missLatencyHistograms.setSize(RubyRequestType_NUM);
- for(int i=0; i<m_missLatencyHistograms.size(); i++) {
- m_missLatencyHistograms[i].clear(200);
- }
- m_machLatencyHistograms.setSize(GenericMachineType_NUM+1);
- for(int i=0; i<m_machLatencyHistograms.size(); i++) {
- m_machLatencyHistograms[i].clear(200);
- }
- m_allMissLatencyHistogram.clear(200);
-
- m_SWPrefetchLatencyHistograms.setSize(CacheRequestType_NUM);
- for(int i=0; i<m_SWPrefetchLatencyHistograms.size(); i++) {
- m_SWPrefetchLatencyHistograms[i].clear(200);
- }
- m_SWPrefetchMachLatencyHistograms.setSize(GenericMachineType_NUM+1);
- for(int i=0; i<m_SWPrefetchMachLatencyHistograms.size(); i++) {
- m_SWPrefetchMachLatencyHistograms[i].clear(200);
- }
- m_allSWPrefetchLatencyHistogram.clear(200);
-
- m_sequencer_requests.clear();
- m_read_sharing_histogram.clear();
- m_write_sharing_histogram.clear();
- m_all_sharing_histogram.clear();
- m_cache_to_cache = 0;
- m_memory_to_cache = 0;
-
- // clear HashMaps
- m_requestProfileMap_ptr->clear();
-
- // count requests profiled
- m_requests = 0;
-
- m_outstanding_requests.clear();
- m_outstanding_persistent_requests.clear();
-
- // Flush the prefetches through the system - used so that there are no outstanding requests after stats are cleared
- //g_eventQueue_ptr->triggerAllEvents();
-
- // update the start time
- m_ruby_start = g_eventQueue_ptr->getTime();
-}
-
-void Profiler::addAddressTraceSample(const CacheMsg& msg, NodeID id)
-{
- if (msg.getType() != CacheRequestType_IFETCH) {
-
- // Note: The following line should be commented out if you want to
- // use the special profiling that is part of the GS320 protocol
-
- // NOTE: Unless PROFILE_HOT_LINES is enabled, nothing will be profiled by the AddressProfiler
- m_address_profiler_ptr->addTraceSample(msg.getLineAddress(), msg.getProgramCounter(), msg.getType(), msg.getAccessMode(), id, false);
- }
-}
-
-void Profiler::profileSharing(const Address& addr, AccessType type, NodeID requestor, const Set& sharers, const Set& owner)
-{
- Set set_contacted(owner);
- if (type == AccessType_Write) {
- set_contacted.addSet(sharers);
- }
- set_contacted.remove(requestor);
- int number_contacted = set_contacted.count();
-
- if (type == AccessType_Write) {
- m_write_sharing_histogram.add(number_contacted);
- } else {
- m_read_sharing_histogram.add(number_contacted);
- }
- m_all_sharing_histogram.add(number_contacted);
-
- if (number_contacted == 0) {
- m_memory_to_cache++;
- } else {
- m_cache_to_cache++;
- }
-
-}
-
-void Profiler::profileMsgDelay(int virtualNetwork, int delayCycles) {
- assert(virtualNetwork < m_delayedCyclesVCHistograms.size());
- m_delayedCyclesHistogram.add(delayCycles);
- m_delayedCyclesVCHistograms[virtualNetwork].add(delayCycles);
- if (virtualNetwork != 0) {
- m_delayedCyclesNonPFHistogram.add(delayCycles);
- }
-}
-
-// profiles original cache requests including PUTs
-void Profiler::profileRequest(const string& requestStr)
-{
- m_requests++;
-
- if (m_requestProfileMap_ptr->exist(requestStr)) {
- (m_requestProfileMap_ptr->lookup(requestStr))++;
- } else {
- m_requestProfileMap_ptr->add(requestStr, 1);
- }
-}
-
-void Profiler::startTransaction(int cpu)
-{
- m_perProcStartTransaction[cpu]++;
-}
-
-void Profiler::endTransaction(int cpu)
-{
- m_perProcEndTransaction[cpu]++;
-}
-
-void Profiler::controllerBusy(MachineID machID)
-{
- m_busyControllerCount[(int)machID.type][(int)machID.num]++;
-}
-void Profiler::profilePFWait(Time waitTime)
-{
- m_prefetchWaitHistogram.add(waitTime);
-}
-
-void Profiler::bankBusy()
-{
- m_busyBankCount++;
-}
-
-// non-zero cycle demand request
-void Profiler::missLatency(Time t, RubyRequestType type)
-{
- m_allMissLatencyHistogram.add(t);
- m_missLatencyHistograms[type].add(t);
-}
-
-// non-zero cycle prefetch request
-void Profiler::swPrefetchLatency(Time t, CacheRequestType type, GenericMachineType respondingMach)
-{
- m_allSWPrefetchLatencyHistogram.add(t);
- m_SWPrefetchLatencyHistograms[type].add(t);
- m_SWPrefetchMachLatencyHistograms[respondingMach].add(t);
- if(respondingMach == GenericMachineType_Directory || respondingMach == GenericMachineType_NUM) {
- m_SWPrefetchL2MissLatencyHistogram.add(t);
- }
-}
-
-void Profiler::profileTransition(const string& component, NodeID version, Address addr,
- const string& state, const string& event,
- const string& next_state, const string& note)
-{
- const int EVENT_SPACES = 20;
- const int ID_SPACES = 3;
- const int TIME_SPACES = 7;
- const int COMP_SPACES = 10;
- const int STATE_SPACES = 6;
-
- if ((g_debug_ptr->getDebugTime() > 0) &&
- (g_eventQueue_ptr->getTime() >= g_debug_ptr->getDebugTime())) {
- (* debug_cout_ptr).flags(ios::right);
- (* debug_cout_ptr) << setw(TIME_SPACES) << g_eventQueue_ptr->getTime() << " ";
- (* debug_cout_ptr) << setw(ID_SPACES) << version << " ";
- (* debug_cout_ptr) << setw(COMP_SPACES) << component;
- (* debug_cout_ptr) << setw(EVENT_SPACES) << event << " ";
-
- (* debug_cout_ptr).flags(ios::right);
- (* debug_cout_ptr) << setw(STATE_SPACES) << state;
- (* debug_cout_ptr) << ">";
- (* debug_cout_ptr).flags(ios::left);
- (* debug_cout_ptr) << setw(STATE_SPACES) << next_state;
-
- (* debug_cout_ptr) << " " << addr << " " << note;
-
- (* debug_cout_ptr) << endl;
- }
-}
+ for (uint32_t i = 0; i < MachineType_NUM; i++) {
+ for (map<uint32_t, AbstractController*>::iterator it =
+ m_ruby_system->m_abstract_controls[i].begin();
+ it != m_ruby_system->m_abstract_controls[i].end(); ++it) {
+
+ AbstractController *ctr = (*it).second;
+ Sequencer *seq = ctr->getCPUSequencer();
+ if (seq != NULL) {
+ rubyProfilerStats.
+ m_outstandReqHistSeqr.add(seq->getOutstandReqHist());
+ }
+#ifdef BUILD_GPU
+ GPUCoalescer *coal = ctr->getGPUCoalescer();
+ if (coal != NULL) {
+ rubyProfilerStats.
+ m_outstandReqHistCoalsr.add(coal->getOutstandReqHist());
+ }
+#endif
+ }
+ }
-// Helper function
-static double process_memory_total()
-{
- const double MULTIPLIER = 4096.0/(1024.0*1024.0); // 4kB page size, 1024*1024 bytes per MB,
- ifstream proc_file;
- proc_file.open("/proc/self/statm");
- int total_size_in_pages = 0;
- int res_size_in_pages = 0;
- proc_file >> total_size_in_pages;
- proc_file >> res_size_in_pages;
- return double(total_size_in_pages)*MULTIPLIER; // size in megabytes
+ for (uint32_t i = 0; i < MachineType_NUM; i++) {
+ for (map<uint32_t, AbstractController*>::iterator it =
+ m_ruby_system->m_abstract_controls[i].begin();
+ it != m_ruby_system->m_abstract_controls[i].end(); ++it) {
+
+ AbstractController *ctr = (*it).second;
+ Sequencer *seq = ctr->getCPUSequencer();
+ if (seq != NULL) {
+ // add all the latencies
+ rubyProfilerStats.
+ m_latencyHistSeqr.add(seq->getLatencyHist());
+ rubyProfilerStats.
+ m_hitLatencyHistSeqr.add(seq->getHitLatencyHist());
+ rubyProfilerStats.
+ m_missLatencyHistSeqr.add(seq->getMissLatencyHist());
+
+ // add the per request type latencies
+ for (uint32_t j = 0; j < RubyRequestType_NUM; ++j) {
+ rubyProfilerStats
+ .perRequestTypeStats
+ .m_typeLatencyHistSeqr[j]
+ ->add(seq->getTypeLatencyHist(j));
+ rubyProfilerStats
+ .perRequestTypeStats
+ .m_hitTypeLatencyHistSeqr[j]
+ ->add(seq->getHitTypeLatencyHist(j));
+ rubyProfilerStats
+ .perRequestTypeStats
+ .m_missTypeLatencyHistSeqr[j]
+ ->add(seq->getMissTypeLatencyHist(j));
+ }
+
+ // add the per machine type miss latencies
+ for (uint32_t j = 0; j < MachineType_NUM; ++j) {
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_hitMachLatencyHistSeqr[j]
+ ->add(seq->getHitMachLatencyHist(j));
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_missMachLatencyHistSeqr[j]
+ ->add(seq->getMissMachLatencyHist(j));
+
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_IssueToInitialDelayHistSeqr[j]
+ ->add(seq->getIssueToInitialDelayHist(MachineType(j)));
+
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_InitialToForwardDelayHistSeqr[j]
+ ->add(seq
+ ->getInitialToForwardDelayHist(MachineType(j)));
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_ForwardToFirstResponseDelayHistSeqr[j]
+ ->add(seq
+ ->getForwardRequestToFirstResponseHist(
+ MachineType(j)));
+
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_FirstResponseToCompletionDelayHistSeqr[j]
+ ->add(seq
+ ->getFirstResponseToCompletionDelayHist(
+ MachineType(j)));
+
+ *(rubyProfilerStats
+ .perMachineTypeStats
+ .m_IncompleteTimesSeqr[j]) +=
+ seq->getIncompleteTimes(MachineType(j));
+ }
+
+ // add the per (request, machine) type miss latencies
+ for (uint32_t j = 0; j < RubyRequestType_NUM; j++) {
+ for (uint32_t k = 0; k < MachineType_NUM; k++) {
+ rubyProfilerStats
+ .perRequestTypeMachineTypeStats
+ .m_hitTypeMachLatencyHistSeqr[j][k]->add(
+ seq->getHitTypeMachLatencyHist(j,k));
+ rubyProfilerStats
+ .perRequestTypeMachineTypeStats
+ .m_missTypeMachLatencyHistSeqr[j][k]->add(
+ seq->getMissTypeMachLatencyHist(j,k));
+ }
+ }
+ }
+#ifdef BUILD_GPU
+ GPUCoalescer *coal = ctr->getGPUCoalescer();
+ if (coal != NULL) {
+ // add all the latencies
+ rubyProfilerStats.
+ m_latencyHistCoalsr.add(coal->getLatencyHist());
+ rubyProfilerStats.
+ m_missLatencyHistCoalsr.add(coal->getMissLatencyHist());
+
+ // add the per request type latencies
+ for (uint32_t j = 0; j < RubyRequestType_NUM; ++j) {
+ rubyProfilerStats
+ .perRequestTypeStats
+ .m_typeLatencyHistCoalsr[j]
+ ->add(coal->getTypeLatencyHist(j));
+ rubyProfilerStats
+ .perRequestTypeStats
+ .m_missTypeLatencyHistCoalsr[j]
+ ->add(coal->getMissTypeLatencyHist(j));
+ }
+
+ // add the per machine type miss latencies
+ for (uint32_t j = 0; j < MachineType_NUM; ++j) {
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_missMachLatencyHistCoalsr[j]
+ ->add(coal->getMissMachLatencyHist(j));
+
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_IssueToInitialDelayHistCoalsr[j]
+ ->add(coal->getIssueToInitialDelayHist(
+ MachineType(j)));
+
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_InitialToForwardDelayHistCoalsr[j]
+ ->add(coal->getInitialToForwardDelayHist(
+ MachineType(j)));
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_ForwardToFirstResponseDelayHistCoalsr[j]
+ ->add(coal->getForwardRequestToFirstResponseHist(
+ MachineType(j)));
+
+ rubyProfilerStats
+ .perMachineTypeStats
+ .m_FirstResponseToCompletionDelayHistCoalsr[j]
+ ->add(coal->getFirstResponseToCompletionDelayHist(
+ MachineType(j)));
+ }
+
+ // add the per (request, machine) type miss latencies
+ for (uint32_t j = 0; j < RubyRequestType_NUM; j++) {
+ for (uint32_t k = 0; k < MachineType_NUM; k++) {
+ rubyProfilerStats
+ .perRequestTypeMachineTypeStats
+ .m_missTypeMachLatencyHistCoalsr[j][k]
+ ->add(coal->getMissTypeMachLatencyHist(j,k));
+ }
+ }
+ }
+#endif
+ }
+ }
}
-static double process_memory_resident()
+void
+Profiler::addAddressTraceSample(const RubyRequest& msg, NodeID id)
{
- const double MULTIPLIER = 4096.0/(1024.0*1024.0); // 4kB page size, 1024*1024 bytes per MB,
- ifstream proc_file;
- proc_file.open("/proc/self/statm");
- int total_size_in_pages = 0;
- int res_size_in_pages = 0;
- proc_file >> total_size_in_pages;
- proc_file >> res_size_in_pages;
- return double(res_size_in_pages)*MULTIPLIER; // size in megabytes
-}
-
-void Profiler::rubyWatch(int id){
- //int rn_g1 = 0;//SIMICS_get_register_number(id, "g1");
- uint64 tr = 0;//SIMICS_read_register(id, rn_g1);
- Address watch_address = Address(tr);
- const int ID_SPACES = 3;
- const int TIME_SPACES = 7;
-
- (* debug_cout_ptr).flags(ios::right);
- (* debug_cout_ptr) << setw(TIME_SPACES) << g_eventQueue_ptr->getTime() << " ";
- (* debug_cout_ptr) << setw(ID_SPACES) << id << " "
- << "RUBY WATCH "
- << watch_address
- << endl;
-
- if(!m_watch_address_list_ptr->exist(watch_address)){
- m_watch_address_list_ptr->add(watch_address, 1);
+ if (msg.getType() != RubyRequestType_IFETCH) {
+ // Note: The following line should be commented out if you
+ // want to use the special profiling that is part of the GS320
+ // protocol
+
+ // NOTE: Unless PROFILE_HOT_LINES is enabled, nothing will be
+ // profiled by the AddressProfiler
+ m_address_profiler_ptr->
+ addTraceSample(msg.getLineAddress(), msg.getProgramCounter(),
+ msg.getType(), msg.getAccessMode(), id, false);
}
}
-bool Profiler::watchAddress(Address addr){
- if (m_watch_address_list_ptr->exist(addr))
- return true;
- else
- return false;
-}
-
-int64 Profiler::getTotalTransactionsExecuted() const {
- return m_perProcEndTransaction.sum();
-}
-
-
-Profiler *
-RubyProfilerParams::create()
-{
- return new Profiler(this);
-}
projects / gem5.git / blobdiff