edits
authorPeizhao Ou <peizhaoo@uci.edu>
Wed, 11 Feb 2015 02:11:00 +0000 (18:11 -0800)
committerPeizhao Ou <peizhaoo@uci.edu>
Wed, 11 Feb 2015 02:11:00 +0000 (18:11 -0800)
benchmarks.mk
concurrent-hashmap/Makefile
concurrent-hashmap/hashmap.h
concurrent-hashmap/hashmap_wildcard.h [new file with mode: 0644]
concurrent-hashmap/main_wildcard.cc [new file with mode: 0644]
mpmc-queue/note.txt [new file with mode: 0644]

index 7f82f920f3c4df24d5ee651a5b7daccc628c55c9..819248ee5624aa78ecf3a4709364181338317994 100644 (file)
@@ -9,7 +9,7 @@ LIB_NAME = model
 LIB_SO = lib$(LIB_NAME).so
 
 BASE = ../..
-INCLUDE = -I$(BASE)/include -I../include
+INCLUDE = -I$(BASE)/include -I../include -I$(BASE)/scfence -I$(BASE)
 
 # C preprocessor flags
 CPPFLAGS += $(INCLUDE) -g
index 4db5bca0e536d44f4bb95aeaef3017adbd7060b4..d39e8315c683ebb8041c9b2353a40e0737be69b2 100644 (file)
@@ -1,11 +1,17 @@
 include ../benchmarks.mk
 
-TESTS := table 
+TESTS := table table_wildcard1
 
 all: $(TESTS)
 
 table: main.cc hashmap.h
        $(CXX) -o $@ $^ $(SPEC_OBJ) $(CXXFLAGS) -std=c++0x $(LDFLAGS)
 
+table_wildcard1: main_wildcard1.cc hashmap_wildcard.h
+       $(CXX) -o $@ $^ $(SPEC_OBJ) $(CXXFLAGS) -std=c++0x $(LDFLAGS)
+
+table_wildcard2: main_wildcard1.cc hashmap_wildcard.h
+       $(CXX) -o $@ $^ $(SPEC_OBJ) $(CXXFLAGS) -std=c++0x $(LDFLAGS)
+
 clean:
        rm -f *.o *.d $(TESTS)
index a66005ccbe9d66039029e3574795fc4d4aff06a0..936f284635a1b71946b1fe867c964852aace5e04 100644 (file)
@@ -14,7 +14,7 @@
 #include <stdlib.h>
 #include <mutex>
 
-#include "cdsannotate.h"
+#include "sc_annotation.h"
 
 #define relaxed memory_order_relaxed
 #define release memory_order_release
@@ -164,7 +164,11 @@ class HashMap {
                // lock, we ignore this operation for the SC analysis, and otherwise we
                // take it into consideration
                
+               SC_BEGIN();
                Entry *firstPtr = first->load(acquire);
+               SC_KEEP();
+               SC_END();
+
                e = firstPtr;
                while (e != NULL) {
                        if (e->hash == hash && eq(key, e->key)) {
diff --git a/concurrent-hashmap/hashmap_wildcard.h b/concurrent-hashmap/hashmap_wildcard.h
new file mode 100644 (file)
index 0000000..557a918
--- /dev/null
@@ -0,0 +1,307 @@
+#ifndef _HASHMAP_H
+#define _HASHMAP_H
+
+#include <iostream>
+#include <atomic>
+#include "stdio.h" 
+//#include <common.h>
+#ifdef STANDALONE
+#include <assert.h>
+#define MODEL_ASSERT assert 
+#else
+#include <model-assert.h>
+#endif
+#include <stdlib.h>
+#include <mutex>
+
+#include "sc_annotation.h"
+#include "wildcard.h"
+
+#define relaxed memory_order_relaxed
+#define release memory_order_release
+#define acquire memory_order_acquire
+#define acq_rel memory_order_acq_rel
+#define seq_cst memory_order_seq_cst
+
+using namespace std;
+
+/**
+       For the sake of simplicity, we do not use template but some toy structs to
+       represent the Key and Value.
+*/
+struct Key {
+       // Probably represent the coordinate (x, y, z)
+       int x;
+       int y;
+       int z;
+
+       int hashCode() {
+               return x + 31 * y + 31 * 31 * z;
+       }
+
+       bool equals(Key *other) {
+               if (!other)
+                       return false;
+               return x == other->x && y == other->y && z == other->z;
+       }
+
+       Key(int x, int y, int z) :
+               x(x),
+               y(y),
+               z(z)
+       {
+
+       }
+};
+
+struct Value {
+       // Probably represent the speed vector (vX, vY, vZ)
+       int vX;
+       int vY;
+       int vZ;
+
+       Value(int vX, int vY, int vZ) :
+               vX(vX),
+               vY(vY),
+               vZ(vZ)
+       {
+
+       }
+
+       bool equals(Value *other) {
+               if (!other)
+                       return false;
+               return vX == other->vX && vY == other->vY && vZ == other->vZ;
+       }
+};
+
+class Entry {
+       public:
+       Key *key;
+       atomic<Value*> value;
+       int hash;
+       atomic<Entry*> next;
+
+       Entry(int h, Key *k, Value *v, Entry *n) {
+               this->hash = h;
+               this->key = k;
+               this->value.store(v, wildcard(1)); // relaxed
+               this->next.store(n, wildcard(2)); // relaxed
+       }
+};
+
+class Segment {
+       public:
+       int count;
+       mutex segMutex;
+
+       void lock() {
+               segMutex.lock();
+       }
+
+       void unlock() {
+               segMutex.unlock();
+       }
+
+       Segment() {
+               this->count = 0;
+       }
+};
+
+class HashMap {
+       public:
+       atomic<Entry*> *table;
+
+       int capacity;
+       int size;
+
+       static const int CONCURRENCY_LEVEL = 16;
+
+       static const int SEGMENT_MASK = CONCURRENCY_LEVEL - 1;
+
+       Segment *segments[CONCURRENCY_LEVEL];
+
+       static const int DEFAULT_INITIAL_CAPACITY = 16;
+
+       // Not gonna consider resize now...
+       
+       HashMap() {
+               this->size = 0;
+               this->capacity = DEFAULT_INITIAL_CAPACITY;
+               this->table = new atomic<Entry*>[capacity];
+               for (int i = 0; i < capacity; i++) {
+                       atomic_init(&table[i], NULL);
+               }
+               for (int i = 0; i < CONCURRENCY_LEVEL; i++) {
+                       segments[i] = new Segment;
+               }
+       }
+
+       int hashKey(Key *x) {
+               int h = x->hashCode();
+               // Use logical right shift
+               unsigned int tmp = (unsigned int) h;
+               return ((h << 7) - h + (tmp >> 9) + (tmp >> 17));
+       }
+
+       bool eq(Key *x, Key *y) {
+               return x == y || x->equals(y);
+       }
+
+       Value* get(Key *key) {
+               MODEL_ASSERT (key);
+               int hash = hashKey(key);
+
+               // Try first without locking...
+               atomic<Entry*> *tab = table;
+               int index = hash & (capacity - 1);
+               atomic<Entry*> *first = &tab[index];
+               Entry *e;
+               Value *res = NULL;
+
+               // Should be a load acquire
+               // This load action here makes it problematic for the SC analysis, what
+               // we need to do is as follows: if the get() method ever acquires the
+               // lock, we ignore this operation for the SC analysis, and otherwise we
+               // take it into consideration
+               
+               SC_BEGIN();
+               Entry *firstPtr = first->load(wildcard(3)); // acquire
+               SC_KEEP();
+               SC_END();
+
+               e = firstPtr;
+               while (e != NULL) {
+                       if (e->hash == hash && eq(key, e->key)) {
+                               res = e->value.load(wildcard(4)); // seq_cst
+                               if (res != NULL)
+                                       return res;
+                               else
+                                       break;
+                               // Loading the next entry
+                               e = e->next.load(wildcard(5)); // acquire
+                       }
+               }
+       
+               // Recheck under synch if key apparently not there or interference
+               Segment *seg = segments[hash & SEGMENT_MASK];
+               seg->lock(); // Critical region begins
+               // Not considering resize now, so ignore the reload of table...
+
+               // Synchronized by locking, no need to be load acquire
+               Entry *newFirstPtr = first->load(wildcard(6)); // relaxed
+               if (e != NULL || firstPtr != newFirstPtr) {
+                       e = newFirstPtr;
+                       while (e != NULL) {
+                               if (e->hash == hash && eq(key, e->key)) {
+                                       res = e->value.load(wildcard(7)); // seq_cst
+                                       seg->unlock(); // Critical region ends
+                                       return res;
+                               }
+                               // Synchronized by locking
+                               e = e->next.load(wildcard(8)); // relaxed
+                       }
+               }
+               seg->unlock(); // Critical region ends
+               return NULL;
+       }
+
+       Value* put(Key *key, Value *value) {
+               // Don't allow NULL key or value
+               MODEL_ASSERT (key && value);
+
+               int hash = hashKey(key);
+               Segment *seg = segments[hash & SEGMENT_MASK];
+               atomic<Entry*> *tab;
+
+               seg->lock(); // Critical region begins
+               tab = table;
+               int index = hash & (capacity - 1);
+
+               atomic<Entry*> *first = &tab[index];
+               Entry *e;
+               Value *oldValue = NULL;
+       
+               // The written of the entry is synchronized by locking
+               Entry *firstPtr = first->load(wildcard(9)); // relaxed
+               e = firstPtr;
+               while (e != NULL) {
+                       if (e->hash == hash && eq(key, e->key)) {
+                               // FIXME: This could be a relaxed (because locking synchronize
+                               // with the previous put())?? 
+                               oldValue = e->value.load(wildcard(10)); // acquire
+                               e->value.store(value, wildcard(11)); // seq_cst
+                               seg->unlock(); // Don't forget to unlock before return
+                               return oldValue;
+                       }
+                       // Synchronized by locking
+                       e = e->next.load(wildcard(12)); // relaxed
+               }
+
+               // Add to front of list
+               Entry *newEntry = new Entry(hash, key, value, firstPtr);
+               // Publish the newEntry to others
+               first->store(newEntry, wildcard(13)); // release
+               seg->unlock(); // Critical region ends
+               return NULL;
+       }
+
+       Value* remove(Key *key, Value *value) {
+               MODEL_ASSERT (key);
+               int hash = hashKey(key);
+               Segment *seg = segments[hash & SEGMENT_MASK];
+               atomic<Entry*> *tab;
+
+               seg->lock(); // Critical region begins
+               tab = table;
+               int index = hash & (capacity - 1);
+
+               atomic<Entry*> *first = &tab[index];
+               Entry *e;
+               Value *oldValue = NULL;
+       
+               // The written of the entry is synchronized by locking
+               Entry *firstPtr = first->load(relaxed);
+               e = firstPtr;
+
+               while (true) {
+                       if (e != NULL) {
+                               seg->unlock(); // Don't forget to unlock
+                               return NULL;
+                       }
+                       if (e->hash == hash && eq(key, e->key))
+                               break;
+                       // Synchronized by locking
+                       e = e->next.load(relaxed);
+               }
+
+               // FIXME: This could be a relaxed (because locking synchronize
+               // with the previous put())?? 
+               oldValue = e->value.load(acquire);
+               // If the value parameter is NULL, we will remove the entry anyway
+               if (value != NULL && value->equals(oldValue)) {
+                       seg->unlock();
+                       return NULL;
+               }
+
+               // Force the get() to grab the lock and retry
+               e->value.store(NULL, relaxed);
+
+               // The strategy to remove the entry is to keep the entries after the
+               // removed one and copy the ones before it
+               Entry *head = e->next.load(relaxed);
+               Entry *p;
+               p = first->load(relaxed);
+               while (p != e) {
+                       head = new Entry(p->hash, p->key, p->value.load(relaxed), head);
+                       p = p->next.load(relaxed);
+               }
+
+               // Publish the new head to readers 
+               first->store(head, release);
+               seg->unlock(); // Critical region ends
+               return oldValue;
+       }
+};
+
+#endif
diff --git a/concurrent-hashmap/main_wildcard.cc b/concurrent-hashmap/main_wildcard.cc
new file mode 100644 (file)
index 0000000..4a6f904
--- /dev/null
@@ -0,0 +1,58 @@
+#include <iostream>
+#include <threads.h>
+#include "hashmap_wildcard.h"
+
+HashMap *table;
+
+Key *k1, *k2;
+Value *r1, *r2, *r3, *r4, *v1, *v2;
+
+void printKey(Key *key) {
+       if (key)
+               printf("pos = (%d, %d, %d)\n", key->x, key->y, key->z);
+       else
+               printf("pos = NULL\n");
+}
+
+void printValue(Value *value) {
+       if (value)
+               printf("velocity = (%d, %d, %d)\n", value->vX, value->vY, value->vZ);
+       else
+               printf("velocity = NULL\n");
+}
+
+void threadA(void *arg) {
+       k1 = new Key(1, 1, 1);
+       k2 = new Key(3, 4, 5);
+       v1 = new Value(10, 10, 10);
+       r1 = table->put(k1, v1);
+       //printValue(r1);
+       r2 = table->get(k2);
+       printf("Thrd A:\n");
+       printValue(r2);
+}
+
+void threadB(void *arg) {
+       k1 = new Key(1, 1, 1);
+       k2 = new Key(3, 4, 5);
+       v2 = new Value(30, 40, 50);
+       r3 = table->put(k2, v2);
+       //printValue(r3);
+       r4 = table->get(k1);
+       printf("Thrd B:\n");
+       printValue(r4);
+}
+
+int user_main(int argc, char *argv[]) {
+       thrd_t t1, t2;
+       table = new HashMap;
+
+       thrd_create(&t1, threadA, NULL);
+       thrd_create(&t2, threadB, NULL);
+       thrd_join(t1);
+       thrd_join(t2);
+       
+       return 0;
+}
+
+
diff --git a/mpmc-queue/note.txt b/mpmc-queue/note.txt
new file mode 100644 (file)
index 0000000..e8597d3
--- /dev/null
@@ -0,0 +1,20 @@
+#1 Why this data structure is non-SC (but it's SC at abstraction).
+
+The underlying pattern of the non-SCness are as the following:
+
+T1                             T2
+x = 1;                 y = 1;
+r1 = y;                        r2 = x;
+
+Suppose there are two identical threads, both having a writer followed by a
+reader, as follows:
+
+T1                                                     T2
+m_written.CAS;                         m_rdwr.CAS;
+
+m_rdwr.CAS;                                    m_written.CAS; // In write_publish()
+
+m_written.load; // a           m_rdwr.load; // The first in the read_fetch() // b
+
+a & b can both read old value and it's non-SC. However, the later loads and CAS
+in the loops will finally fix this problem.