Ruby: Resurrect Cache Warmup Capability

This patch resurrects ruby's cache warmup capability. It essentially
makes use of all the infrastructure that was added to the controllers,
memories and the cache recorder.

src/mem/ruby/buffers/MessageBuffer.cc

@@ -198,7 +198,11 @@ MessageBuffer::enqueue(MsgPtr message, Time delta)
                   m_last_arrival_time * g_eventQueue_ptr->getClock());
         }
     }
-    m_last_arrival_time = arrival_time;
+
+    // If running a cache trace, don't worry about the last arrival checks
+    if (!g_system_ptr->m_warmup_enabled) {
+        m_last_arrival_time = arrival_time;
+    }
 
     // compute the delay cycles and set enqueue time
     Message* msg_ptr = message.get();

src/mem/ruby/system/DMASequencer.hh

@@ -55,6 +55,9 @@ class DMASequencer : public RubyPort
     /* external interface */
     RequestStatus makeRequest(PacketPtr pkt);
     bool busy() { return m_is_busy;}
+    int outstandingCount() const { return (m_is_busy ? 1 : 0); }
+    bool isDeadlockEventScheduled() const { return false; }
+    void descheduleDeadlockEvent() {}
 
     /* SLICC callback */
     void dataCallback(const DataBlock & dblk);

src/mem/ruby/system/DirectoryMemory.cc

@@ -58,6 +58,7 @@ DirectoryMemory::init()
 
     if (m_use_map) {
         m_sparseMemory = new SparseMemory(m_map_levels);
+        g_system_ptr->registerSparseMemory(m_sparseMemory);
     } else {
         m_entries = new AbstractEntry*[m_num_entries];
         for (int i = 0; i < m_num_entries; i++)

src/mem/ruby/system/RubyPort.cc

@@ -27,11 +27,11 @@
  */
 
 #include "cpu/testers/rubytest/RubyTester.hh"
+#include "debug/Config.hh"
 #include "debug/Ruby.hh"
 #include "mem/protocol/AccessPermission.hh"
 #include "mem/ruby/slicc_interface/AbstractController.hh"
 #include "mem/ruby/system/RubyPort.hh"
-#include "mem/physical.hh"
 
 RubyPort::RubyPort(const Params *p)
     : MemObject(p)
@@ -51,6 +51,8 @@ RubyPort::RubyPort(const Params *p)
     m_usingRubyTester = p->using_ruby_tester;
     access_phys_mem = p->access_phys_mem;
 
+    drainEvent = NULL;
+
     ruby_system = p->ruby_system;
     waitingOnSequencer = false;
 }
@@ -510,6 +512,82 @@ RubyPort::ruby_hit_callback(PacketPtr pkt)
             (*i)->sendRetry();
         }
     }
+
+    testDrainComplete();
+}
+
+void
+RubyPort::testDrainComplete()
+{
+    //If we weren't able to drain before, we might be able to now.
+    if (drainEvent != NULL) {
+        unsigned int drainCount = getDrainCount(drainEvent);
+        DPRINTF(Config, "Drain count: %u\n", drainCount);
+        if (drainCount == 0) {
+            drainEvent->process();
+            // Clear the drain event once we're done with it.
+            drainEvent = NULL;
+        }
+    }
+}
+
+unsigned int
+RubyPort::getDrainCount(Event *de)
+{
+    int count = 0;
+    //
+    // If the sequencer is not empty, then requests need to drain.
+    // The outstandingCount is the number of requests outstanding and thus the
+    // number of times M5's timing port will process the drain event.
+    //
+    count += outstandingCount();
+
+    DPRINTF(Config, "outstanding count %d\n", outstandingCount());
+
+    // To simplify the draining process, the sequencer's deadlock detection
+    // event should have been descheduled.
+    assert(isDeadlockEventScheduled() == false);
+
+    if (pio_port != NULL) {
+        count += pio_port->drain(de);
+        DPRINTF(Config, "count after pio check %d\n", count);
+    }
+    if (physMemPort != NULL) {
+        count += physMemPort->drain(de);
+        DPRINTF(Config, "count after physmem check %d\n", count);
+    }
+
+    for (CpuPortIter p_iter = cpu_ports.begin(); p_iter != cpu_ports.end();
+         p_iter++) {
+        M5Port* cpu_port = *p_iter;
+        count += cpu_port->drain(de);
+        DPRINTF(Config, "count after cpu port check %d\n", count);
+    }
+
+    DPRINTF(Config, "final count %d\n", count);
+
+    return count;
+}
+
+unsigned int
+RubyPort::drain(Event *de)
+{
+    if (isDeadlockEventScheduled()) {
+        descheduleDeadlockEvent();
+    }
+
+    int count = getDrainCount(de);
+
+    // Set status
+    if (count != 0) {
+        drainEvent = de;
+
+        changeState(SimObject::Draining);
+        return count;
+    }
+
+    changeState(SimObject::Drained);
+    return 0;
 }
 
 void

src/mem/ruby/system/RubyPort.hh

@@ -33,7 +33,6 @@
 #include <string>
 
 #include "mem/protocol/RequestStatus.hh"
-#include "mem/ruby/slicc_interface/RubyRequest.hh"
 #include "mem/ruby/system/System.hh"
 #include "mem/mem_object.hh"
 #include "mem/physical.hh"
@@ -115,17 +114,23 @@ class RubyPort : public MemObject
     Port *getPort(const std::string &if_name, int idx);
 
     virtual RequestStatus makeRequest(PacketPtr pkt) = 0;
+    virtual int outstandingCount() const = 0;
+    virtual bool isDeadlockEventScheduled() const = 0;
+    virtual void descheduleDeadlockEvent() = 0;
 
     //
     // Called by the controller to give the sequencer a pointer.
     // A pointer to the controller is needed for atomic support.
     //
     void setController(AbstractController* _cntrl) { m_controller = _cntrl; }
+    int getId() { return m_version; }
+    unsigned int drain(Event *de);
 
   protected:
     const std::string m_name;
     void ruby_hit_callback(PacketPtr pkt);
     void hit(PacketPtr pkt);
+    void testDrainComplete();
 
     int m_version;
     AbstractController* m_controller;
@@ -143,6 +148,8 @@ class RubyPort : public MemObject
         }
     }
 
+    unsigned int getDrainCount(Event *de);
+
     uint16_t m_port_id;
     uint64_t m_request_cnt;
 
@@ -152,6 +159,8 @@ class RubyPort : public MemObject
     typedef std::vector<M5Port*>::iterator CpuPortIter;
     std::vector<M5Port*> cpu_ports;
 
+    Event *drainEvent;
+
     PhysicalMemory* physmem;
     RubySystem* ruby_system;
 

src/mem/ruby/system/Sequencer.cc

@@ -519,7 +519,11 @@ Sequencer::hitCallback(SequencerRequest* srequest,
     }
 
     // update the data
-    if (pkt->getPtr<uint8_t>(true) != NULL) {
+    if (g_system_ptr->m_warmup_enabled) {
+        assert(pkt->getPtr<uint8_t>(false) != NULL);
+        data.setData(pkt->getPtr<uint8_t>(false),
+                     request_address.getOffset(), pkt->getSize());
+    } else if (pkt->getPtr<uint8_t>(true) != NULL) {
         if ((type == RubyRequestType_LD) ||
             (type == RubyRequestType_IFETCH) ||
             (type == RubyRequestType_RMW_Read) ||
@@ -551,8 +555,17 @@ Sequencer::hitCallback(SequencerRequest* srequest,
         testerSenderState->subBlock->mergeFrom(data);
     }
 
-    ruby_hit_callback(pkt);
     delete srequest;
+
+    if (g_system_ptr->m_warmup_enabled) {
+        delete pkt;
+        g_system_ptr->m_cache_recorder->enqueueNextFetchRequest();
+    } else if (g_system_ptr->m_cooldown_enabled) {
+        delete pkt;
+        g_system_ptr->m_cache_recorder->enqueueNextFlushRequest();
+    } else {
+        ruby_hit_callback(pkt);
+    }
 }
 
 bool

src/mem/ruby/system/Sequencer.hh

@@ -39,8 +39,6 @@
 #include "mem/ruby/system/RubyPort.hh"
 
 class DataBlock;
-class CacheMsg;
-class MachineID;
 class CacheMemory;
 
 class RubySequencerParams;
@@ -100,6 +98,18 @@ class Sequencer : public RubyPort, public Consumer
 
     RequestStatus makeRequest(PacketPtr pkt);
     bool empty() const;
+    int outstandingCount() const { return m_outstanding_count; }
+    bool
+    isDeadlockEventScheduled() const
+    {
+        return deadlockCheckEvent.scheduled();
+    }
+
+    void
+    descheduleDeadlockEvent()
+    {
+        deschedule(deadlockCheckEvent);
+    }
 
     void print(std::ostream& out) const;
     void printStats(std::ostream& out) const;

src/mem/ruby/system/System.cc

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1999-2008 Mark D. Hill and David A. Wood
+ * Copyright (c) 1999-2011 Mark D. Hill and David A. Wood
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -26,15 +26,19 @@
  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
+#include <fcntl.h>
+#include <zlib.h>
+
+#include <cstdio>
+
 #include "base/intmath.hh"
 #include "base/output.hh"
-#include "mem/ruby/buffers/MessageBuffer.hh"
+#include "debug/RubySystem.hh"
 #include "mem/ruby/common/Address.hh"
 #include "mem/ruby/network/Network.hh"
 #include "mem/ruby/profiler/Profiler.hh"
-#include "mem/ruby/slicc_interface/AbstractController.hh"
-#include "mem/ruby/system/MemoryVector.hh"
 #include "mem/ruby/system/System.hh"
+#include "sim/simulate.hh"
 
 using namespace std;
 
@@ -86,6 +90,8 @@ RubySystem::RubySystem(const Params *p)
     //
     RubyExitCallback* rubyExitCB = new RubyExitCallback(p->stats_filename);
     registerExitCallback(rubyExitCB);
+    m_warmup_enabled = false;
+    m_cooldown_enabled = false;
 }
 
 void
@@ -112,6 +118,12 @@ RubySystem::registerAbstractController(AbstractController* cntrl)
   m_abs_cntrl_vec.push_back(cntrl);
 }
 
+void
+RubySystem::registerSparseMemory(SparseMemory* s)
+{
+    m_sparse_memory_vector.push_back(s);
+}
+
 RubySystem::~RubySystem()
 {
     delete m_network_ptr;
@@ -157,10 +169,144 @@ RubySystem::printStats(ostream& out)
     m_network_ptr->printStats(out);
 }
 
+void
+RubySystem::writeCompressedTrace(uint8* raw_data, string filename,
+                                 uint64 uncompressed_trace_size)
+{
+    // Create the checkpoint file for the memory
+    string thefile = Checkpoint::dir() + "/" + filename.c_str();
+
+    int fd = creat(thefile.c_str(), 0664);
+    if (fd < 0) {
+        perror("creat");
+        fatal("Can't open memory trace file '%s'\n", filename);
+    }
+
+    gzFile compressedMemory = gzdopen(fd, "wb");
+    if (compressedMemory == NULL)
+        fatal("Insufficient memory to allocate compression state for %s\n",
+              filename);
+
+    if (gzwrite(compressedMemory, raw_data, uncompressed_trace_size) !=
+        uncompressed_trace_size) {
+        fatal("Write failed on memory trace file '%s'\n", filename);
+    }
+
+    if (gzclose(compressedMemory)) {
+        fatal("Close failed on memory trace file '%s'\n", filename);
+    }
+    delete raw_data;
+}
+
 void
 RubySystem::serialize(std::ostream &os)
 {
+    m_cooldown_enabled = true;
 
+    vector<Sequencer*> sequencer_map;
+    Sequencer* sequencer_ptr = NULL;
+    int cntrl_id = -1;
+
+
+    for (int cntrl = 0; cntrl < m_abs_cntrl_vec.size(); cntrl++) {
+        sequencer_map.push_back(m_abs_cntrl_vec[cntrl]->getSequencer());
+        if (sequencer_ptr == NULL) {
+            sequencer_ptr = sequencer_map[cntrl];
+            cntrl_id = cntrl;
+        }
+    }
+
+    assert(sequencer_ptr != NULL);
+
+    for (int cntrl = 0; cntrl < m_abs_cntrl_vec.size(); cntrl++) {
+        if (sequencer_map[cntrl] == NULL) {
+            sequencer_map[cntrl] = sequencer_ptr;
+        }
+    }
+
+    // Create the CacheRecorder and record the cache trace
+    m_cache_recorder = new CacheRecorder(NULL, 0, sequencer_map);
+
+    for (int cntrl = 0; cntrl < m_abs_cntrl_vec.size(); cntrl++) {
+        m_abs_cntrl_vec[cntrl]->recordCacheTrace(cntrl, m_cache_recorder);
+    }
+
+    // save the current tick value
+    Tick curtick_original = curTick();
+    // save the event queue head
+    Event* eventq_head = eventq->replaceHead(NULL);
+
+    // Schedule an event to start cache cooldown
+    RubyEvent* e = new RubyEvent(this);
+    schedule(e,curTick());
+    simulate();
+
+    // Restore eventq head
+    eventq_head = eventq->replaceHead(eventq_head);
+    // Restore curTick
+    curTick(curtick_original);
+
+    uint8* raw_data = NULL;
+
+    if (m_mem_vec_ptr != NULL) {
+        uint64 memory_trace_size = m_mem_vec_ptr->collatePages(raw_data);
+
+        string memory_trace_file = name() + ".memory.gz";
+        writeCompressedTrace(raw_data, memory_trace_file,
+                             memory_trace_size);
+
+        SERIALIZE_SCALAR(memory_trace_file);
+        SERIALIZE_SCALAR(memory_trace_size);
+
+    } else {
+        for (int i = 0; i < m_sparse_memory_vector.size(); ++i) {
+            m_sparse_memory_vector[i]->recordBlocks(cntrl_id,
+                                                    m_cache_recorder);
+        }
+    }
+
+    // Aggergate the trace entries together into a single array
+    raw_data = new uint8_t[4096];
+    uint64 cache_trace_size = m_cache_recorder->aggregateRecords(&raw_data,
+                                                                 4096);
+    string cache_trace_file = name() + ".cache.gz";
+    writeCompressedTrace(raw_data, cache_trace_file, cache_trace_size);
+
+    SERIALIZE_SCALAR(cache_trace_file);
+    SERIALIZE_SCALAR(cache_trace_size);
+
+    m_cooldown_enabled = false;
+}
+
+void
+RubySystem::readCompressedTrace(string filename, uint8*& raw_data,
+                                uint64& uncompressed_trace_size)
+{
+    // Read the trace file
+    gzFile compressedTrace;
+
+    // trace file
+    int fd = open(filename.c_str(), O_RDONLY);
+    if (fd < 0) {
+        perror("open");
+        fatal("Unable to open trace file %s", filename);
+    }
+
+    compressedTrace = gzdopen(fd, "rb");
+    if (compressedTrace == NULL) {
+        fatal("Insufficient memory to allocate compression state for %s\n",
+              filename);
+    }
+
+    raw_data = new uint8_t[uncompressed_trace_size];
+    if (gzread(compressedTrace, raw_data, uncompressed_trace_size) <
+            uncompressed_trace_size) {
+        fatal("Unable to read complete trace from file %s\n", filename);
+    }
+
+    if (gzclose(compressedTrace)) {
+        fatal("Failed to close cache trace file '%s'\n", filename);
+    }
 }
 
 void
@@ -172,6 +318,88 @@ RubySystem::unserialize(Checkpoint *cp, const string &section)
     // value of curTick()
     //
     clearStats();
+    uint8* uncompressed_trace = NULL;
+
+    if (m_mem_vec_ptr != NULL) {
+        string memory_trace_file;
+        uint64 memory_trace_size = 0;
+
+        UNSERIALIZE_SCALAR(memory_trace_file);
+        UNSERIALIZE_SCALAR(memory_trace_size);
+        memory_trace_file = cp->cptDir + "/" + memory_trace_file;
+
+        readCompressedTrace(memory_trace_file, uncompressed_trace,
+                            memory_trace_size);
+        m_mem_vec_ptr->populatePages(uncompressed_trace);
+
+        delete uncompressed_trace;
+        uncompressed_trace = NULL;
+    }
+
+    string cache_trace_file;
+    uint64 cache_trace_size = 0;
+
+    UNSERIALIZE_SCALAR(cache_trace_file);
+    UNSERIALIZE_SCALAR(cache_trace_size);
+    cache_trace_file = cp->cptDir + "/" + cache_trace_file;
+
+    readCompressedTrace(cache_trace_file, uncompressed_trace,
+                        cache_trace_size);
+    m_warmup_enabled = true;
+
+    vector<Sequencer*> sequencer_map;
+    Sequencer* t = NULL;
+    for (int cntrl = 0; cntrl < m_abs_cntrl_vec.size(); cntrl++) {
+        sequencer_map.push_back(m_abs_cntrl_vec[cntrl]->getSequencer());
+        if(t == NULL) t = sequencer_map[cntrl];
+    }
+
+    assert(t != NULL);
+
+    for (int cntrl = 0; cntrl < m_abs_cntrl_vec.size(); cntrl++) {
+        if (sequencer_map[cntrl] == NULL) {
+            sequencer_map[cntrl] = t;
+        }
+    }
+
+    m_cache_recorder = new CacheRecorder(uncompressed_trace, cache_trace_size,
+                                         sequencer_map);
+}
+
+void
+RubySystem::startup()
+{
+    if (m_warmup_enabled) {
+        // save the current tick value
+        Tick curtick_original = curTick();
+        // save the event queue head
+        Event* eventq_head = eventq->replaceHead(NULL);
+        // set curTick to 0
+        curTick(0);
+
+        // Schedule an event to start cache warmup
+        RubyEvent* e = new RubyEvent(this);
+        schedule(e,curTick());
+        simulate();
+
+        delete m_cache_recorder;
+        m_cache_recorder = NULL;
+        m_warmup_enabled = false;
+        // Restore eventq head
+        eventq_head = eventq->replaceHead(eventq_head);
+        // Restore curTick
+        curTick(curtick_original);
+    }
+}
+
+void
+RubySystem::RubyEvent::process()
+{
+    if (ruby_system->m_warmup_enabled) {
+        ruby_system->m_cache_recorder->enqueueNextFetchRequest();
+    }  else if (ruby_system->m_cooldown_enabled) {
+        ruby_system->m_cache_recorder->enqueueNextFlushRequest();
+    }
 }
 
 void
@@ -181,11 +409,6 @@ RubySystem::clearStats() const
     m_network_ptr->clearStats();
 }
 
-void
-RubySystem::recordCacheContents(CacheRecorder& tr) const
-{
-}
-
 #ifdef CHECK_COHERENCE
 // This code will check for cases if the given cache block is exclusive in
 // one node and shared in another-- a coherence violation

src/mem/ruby/system/System.hh

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1999-2008 Mark D. Hill and David A. Wood
+ * Copyright (c) 1999-2012 Mark D. Hill and David A. Wood
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -38,20 +38,34 @@
 #include "base/callback.hh"
 #include "mem/ruby/common/Global.hh"
 #include "mem/ruby/eventqueue/RubyEventQueue.hh"
-#include "mem/ruby/system/RubyPort.hh"
+#include "mem/ruby/recorder/CacheRecorder.hh"
 #include "mem/ruby/slicc_interface/AbstractController.hh"
+#include "mem/ruby/system/MemoryVector.hh"
+#include "mem/ruby/system/SparseMemory.hh"
 #include "params/RubySystem.hh"
 #include "sim/sim_object.hh"
 
-class AbstractController;
-class CacheRecorder;
-class MemoryVector;
 class Network;
 class Profiler;
 
 class RubySystem : public SimObject
 {
   public:
+    class RubyEvent : public Event
+    {
+      public:
+        RubyEvent(RubySystem* _ruby_system)
+        {
+            ruby_system = _ruby_system;
+        }
+      private:
+        void process();
+
+        RubySystem* ruby_system;
+    };
+
+    friend class RubyEvent;
+
     typedef RubySystemParams Params;
     RubySystem(const Params *p);
     ~RubySystem();
@@ -92,7 +106,6 @@ class RubySystem : public SimObject
         return m_mem_vec_ptr;
     }
 
-    void recordCacheContents(CacheRecorder& tr) const;
     static void printConfig(std::ostream& out);
     static void printStats(std::ostream& out);
     void clearStats() const;
@@ -106,12 +119,15 @@ class RubySystem : public SimObject
 
     void print(std::ostream& out) const;
 
-    virtual void serialize(std::ostream &os);
-    virtual void unserialize(Checkpoint *cp, const std::string &section);
+    void serialize(std::ostream &os);
+    void unserialize(Checkpoint *cp, const std::string &section);
+    void process();
+    void startup();
 
     void registerNetwork(Network*);
     void registerProfiler(Profiler*);
     void registerAbstractController(AbstractController*);
+    void registerSparseMemory(SparseMemory*);
 
   private:
     // Private copy constructor and assignment operator
@@ -121,6 +137,11 @@ class RubySystem : public SimObject
     void init();
 
     static void printSystemConfig(std::ostream& out);
+    void readCompressedTrace(std::string filename,
+                             uint8*& raw_data,
+                             uint64& uncompressed_trace_size);
+    void writeCompressedTrace(uint8* raw_data, std::string file,
+                              uint64 uncompressed_trace_size);
 
   private:
     // configuration parameters
@@ -131,13 +152,16 @@ class RubySystem : public SimObject
     static int m_block_size_bits;
     static uint64 m_memory_size_bytes;
     static int m_memory_size_bits;
-
     static Network* m_network_ptr;
 
   public:
     static Profiler* m_profiler_ptr;
     static MemoryVector* m_mem_vec_ptr;
     std::vector<AbstractController*> m_abs_cntrl_vec;
+    bool m_warmup_enabled;
+    bool m_cooldown_enabled;
+    CacheRecorder* m_cache_recorder;
+    std::vector<SparseMemory*> m_sparse_memory_vector;
 };
 
 inline std::ostream&