adapt threadcached and ycsb to work with multi-instanced Montage and …

Makefile

@@ -75,7 +75,7 @@ LDIRS+=$(NVM_MALLOC_DIR)
 # libraries to link against (are expected to be in the above directories,
 # or they are system default)
 # LIBS :=-l:libjemalloc.so.2 -lstm -lparharness -lpthread -lhwloc -lm -lrt
-LIBS :=-l:libjemalloc.so.2 -lpthread -lhwloc -lralloc -lgomp -latomic
+LIBS :=-ljemalloc -lpthread -lhwloc -lralloc -lgomp -latomic
 LIBS+=-lnvmmalloc
 # directories that should be built first using recursive make.
 # You should avoid this in general, but it's useful for building

ext/ralloc/src/ralloc.cpp

@@ -116,18 +116,13 @@ void* Ralloc::reallocate(void* ptr, size_t new_size){
     return new_ptr;
 }
 
-struct RallocHolder{
-    Ralloc* ralloc_instance;
-    inline int init(int thd_num, const char* _id, uint64_t size) {
-        ralloc_instance = new Ralloc(thd_num, _id,size);
-        return (int)ralloc_instance->is_restart();
-    }
-    ~RallocHolder(){ 
-        delete ralloc_instance;
-    }
-};
+int RallocHolder::init(int thd_num, const char* _id, uint64_t size){
+    ralloc_instance = new Ralloc(thd_num, _id,size);
+    RP_set_tid(0);// set tid for main thread
+    return (int)ralloc_instance->is_restart();
+}
 
-static RallocHolder _holder;
+RallocHolder _holder;
 /* 
  * mmap the existing heap file corresponding to id. aka restart,
  * 		and if multiple heaps exist, print out and let user select;

ext/ralloc/src/ralloc.hpp

@@ -111,6 +111,10 @@ class Ralloc{
         return 0;
     }
 
+    inline bool is_initialized(){
+        return initialized;
+    }
+
     static void set_tid(int tid_){
         // Wentao: we deliberately allow tid to be set more than once
         // assert((tid==-1 || tid==0) && "tid set more than once!");
@@ -119,18 +123,25 @@ class Ralloc{
     }
 };
 
+// RAII ralloc global instance holder for global API
+struct RallocHolder{
+    Ralloc* ralloc_instance;
+    int init(int thd_num, const char* _id, uint64_t size);
+    ~RallocHolder(){ 
+        delete ralloc_instance;
+    }
+};
+
+extern RallocHolder _holder;
+
 /* return 1 if it's a restart, otherwise 0. */
 extern "C" int RP_init(const char* _id, uint64_t size = 5*1024*1024*1024ULL, int thd_num = 100);
 
 template<class T>
 T* RP_get_root(uint64_t i){
-    #if 0
-    assert(ralloc::initialized);
-    return ralloc::base_md->get_root<T>(i);
-    #endif
+    return _holder.ralloc_instance->get_root<T>(i);
 }
 
-
 std::vector<InuseRecovery::iterator> RP_recover(int n = 1);
 extern "C"{
 #else /* __cplusplus ends */

ext/threadcached/Makefile

@@ -73,7 +73,7 @@ obj/libthreadcached.a: $(PROT_OBJ)
 obj/libthreadcached_nvm.a: $(PROT_OBJ_NVM) $(MONTAGE_PATH)/ext/ralloc/libralloc.a
 	ar -rcs $@ $^
 
-obj/libthreadcached_montage.a: $(PROT_OBJ_MONTAGE) $(MONTAGE_PATH)/ext/ralloc/libralloc.a $(MONTAGE_PATH)/lib/libPDSHarness.a
+obj/libthreadcached_montage.a: $(PROT_OBJ_MONTAGE) $(MONTAGE_PATH)/ext/ralloc/libralloc.a $(MONTAGE_PATH)/lib/libMontage.a
 	ar -rcs $@ $^
 
 obj/%.o: src/%.cc

ext/threadcached/include/memcached.h

@@ -402,6 +402,8 @@ extern unsigned stats_id;
 
 using namespace pds;
 struct item : public PBlk{
+  item(){};
+  item(const item& oth): PBlk(oth){};
 #else
 struct item{
 #endif
@@ -415,7 +417,7 @@ struct item{
   int             nbytes;     /* size of data */
   unsigned short  refcount;
   uint8_t         nsuffix;    /* length of flags-and-length string */
-  uint8_t         it_flags;   /* ITEM_* above */
+  uint8_t         it_flags=0;   /* ITEM_* above */
   uint8_t         slabs_clsid;/* which slab class we're in */
   uint8_t         nkey;       /* key length, w/terminating null and padding */
   /* this odd type prevents type-punning issues when we do

ext/threadcached/include/pku_memcached.h

@@ -106,7 +106,7 @@ void memcached_start_server();
 memcached_return_t
 memcached_end ();
 
-void memcached_init();
+void memcached_init(int thd_num);
 
 void memcached_close();
 

ext/threadcached/src/items.cc

@@ -188,8 +188,11 @@ item *do_item_alloc_pull(const size_t ntotal, const unsigned int id) {
     uint64_t total_bytes = 0;
     /* Try to reclaim memory first */
     lru_pull_tail(id, COLD_LRU, 0, 0, 0, NULL);
+#ifdef MONTAGE
+    it = (item*)PMALLOC(ntotal);
+#else
     it = (item*)pm_malloc(ntotal);
-
+#endif
     if (it == NULL) {
       // We send '0' in for "total_bytes" as this routine is always
       // pulling to evict, or forcing HOT -> COLD migration.
@@ -666,7 +669,7 @@ int lru_pull_tail(const int orig_id, const int cur_lru,
      */
     switch (cur_lru) {
       case HOT_LRU:
-        limit = total_bytes * 20 / 100;
+        limit = total_bytes * 20 / 100; __attribute__ ((fallthrough));
       case WARM_LRU:
         if (limit == 0)
           limit = total_bytes * 40 / 100;

ext/threadcached/src/itoa_ljust.cc

@@ -91,15 +91,15 @@ static inline int digits( uint32_t u, unsigned k, int* d, char** p, int n ) {
 
 static inline char* itoa(uint32_t u, char* p, int d, int n) {
     switch(n) {
-    case 10: d  = u / 100000000; p = out2( d, p );
-    case  9: u -= d * 100000000;
-    case  8: d  = u /   1000000; p = out2( d, p );
-    case  7: u -= d *   1000000;
-    case  6: d  = u /     10000; p = out2( d, p );
-    case  5: u -= d *     10000;
-    case  4: d  = u /       100; p = out2( d, p );
-    case  3: u -= d *       100;
-    case  2: d  = u /         1; p = out2( d, p );
+    case 10: d  = u / 100000000; p = out2( d, p ); __attribute__ ((fallthrough));
+    case  9: u -= d * 100000000; __attribute__ ((fallthrough));
+    case  8: d  = u /   1000000; p = out2( d, p ); __attribute__ ((fallthrough));
+    case  7: u -= d *   1000000; __attribute__ ((fallthrough));
+    case  6: d  = u /     10000; p = out2( d, p ); __attribute__ ((fallthrough));
+    case  5: u -= d *     10000; __attribute__ ((fallthrough));
+    case  4: d  = u /       100; p = out2( d, p ); __attribute__ ((fallthrough));
+    case  3: u -= d *       100; __attribute__ ((fallthrough));
+    case  2: d  = u /         1; p = out2( d, p ); __attribute__ ((fallthrough));
     case  1: ;
     }
     *p = '\0';

ext/threadcached/src/murmur3_hash.cc

@@ -105,8 +105,8 @@ uint32_t MurmurHash3_x86_32 ( const void * key, size_t length)
 
   switch(length & 3)
   {
-  case 3: k1 ^= tail[2] << 16;
-  case 2: k1 ^= tail[1] << 8;
+  case 3: k1 ^= tail[2] << 16; __attribute__ ((fallthrough));
+  case 2: k1 ^= tail[1] << 8; __attribute__ ((fallthrough));
   case 1: k1 ^= tail[0];
           k1 *= c1; k1 = ROTL32(k1,15); k1 *= c2; h1 ^= k1;
   };

ext/threadcached/src/pku_memcached.cc

@@ -403,15 +403,18 @@ HODOR_FUNC_EXPORT(memcached_start_server, 0);
 #include <errno.h>
 #include <unistd.h>
 bool server_flag = false;
-void memcached_init(){
+void memcached_init(int thd_num){
   if (!run_once){
     run_once = true;
   } else return;
+// Wt: montage init is done before memcached_init
+// TODO: move montage inti into memcached_init
+// this part is initializing the global Ralloc heap for transient allocation
 #ifdef RALLOC
   char* heap_prefix = (char*) malloc(L_cuserid+6);
   cuserid(heap_prefix);
   strcat(heap_prefix, "_memcached");
-  is_restart = RP_init(heap_prefix, MEMORY_MAX);
+  is_restart = RP_init(heap_prefix, MEMORY_MAX, thd_num);
   free(heap_prefix);
 #else
   is_restart = pm_init();

ext/threadcached/src/thread.cc

@@ -215,10 +215,8 @@ enum store_item_type store_item(item *item, int comm) {
 #endif
   ret = do_store_item(item, comm, hv).first;
 #ifdef MONTAGE
-  if(ret == STORED){
-    // Wentao: it is safe to persist item after do_store_item because of the lock
-    esys->register_alloc_pblk(item, epochs[_tid].ui);
-  }
+  // Wentao: PMALLOC is used to allocate new items, so we don't
+  // explicitly register allocated blocks anymore
   END_OP;
 #endif
   item_unlock(hv);

ext/threadcached/test/end.cc

@@ -8,7 +8,7 @@
 #include <string.h>
 
 int main(){
-  memcached_init();
+  memcached_init(1);
   memcached_end();
   memcached_close();
   return 0;

ext/threadcached/test/get.cc

@@ -9,7 +9,7 @@
 
 #define BUFF_LEN 32
 int main(){
-  memcached_init();
+  memcached_init(1);
   std::string name = "chris";
 
   char nbuff[BUFF_LEN];

ext/threadcached/test/insert.cc

@@ -53,7 +53,7 @@ char *keys[N_INSERT];
 char *dats[N_INSERT];
 
 int main(){
-  memcached_init();
+  memcached_init(1);
   std::string name = "chris";
   std::string quality = " tests memcached";
 

ext/threadcached/test/server.cc

@@ -4,6 +4,6 @@
 extern bool server_flag;
 int main(){
   server_flag = true;
-  memcached_init();
+  memcached_init(1);
   memcached_start_server();
 }

ext/threadcached/threadcached.diff

@@ -131,7 +131,7 @@ diff -r ../MemcachedProtectedLibrary/Makefile ./ext/threadcached/Makefile
 > obj/libthreadcached_nvm.a: $(PROT_OBJ_NVM) $(MONTAGE_PATH)/ext/ralloc/test/libralloc.a
 > 	ar -rcs $@ $^
 > 
-> obj/libthreadcached_montage.a: $(PROT_OBJ_MONTAGE) $(MONTAGE_PATH)/ext/ralloc/test/libralloc.a $(MONTAGE_PATH)/lib/libPDSHarness.a
+> obj/libthreadcached_montage.a: $(PROT_OBJ_MONTAGE) $(MONTAGE_PATH)/ext/ralloc/test/libralloc.a $(MONTAGE_PATH)/lib/libMontage.a
 > 	ar -rcs $@ $^
 61c80,86
 < 	$(CXX) -c $^ $(OPTS) -o $@

ext/ycsb-tcd/Makefile

@@ -6,7 +6,7 @@ CFLAGS += -I$(MONTAGE_PATH)/ext/tbb/include
 #CFLAGS=-std=c++11 -O0 -g -Wall -pthread -I./ -DTCD -g -m64
 LDFLAGS= -lpthread -levent \
 	 -L../threadcached/obj \
-	 -ldl -ljemalloc -lhwloc 
+	 -ldl -ljemalloc -lhwloc -lgomp
 SUBDIRS=core db cache_test
 SUBSRCS=$(wildcard core/*.cc) $(wildcard db/*.cc) $(wildcard cache_test/*.cc)
 OBJECTS=$(SUBSRCS:.cc=.o)
@@ -16,8 +16,9 @@ EXEC=ycsbc
 ifeq ($(OPT),montage)
 	CFLAGS += -DMONTAGE
 	CFLAGS += -DRALLOC
+	LDFLAGS += -latomic
 	LDFLAGS += -lthreadcached_montage
-	LDFLAGS += -L$(MONTAGE_PATH)/lib -lPDSHarness
+	LDFLAGS += -L$(MONTAGE_PATH)/lib -lMontage
 	LDFLAGS += -L$(MONTAGE_PATH)/ext/ralloc -lralloc
 endif
 
@@ -34,7 +35,7 @@ ifeq ($(OPT),dram)
 	LDFLAGS += -lthreadcached
 endif
 
-# PDSHarness
+# Montage
 
 # Ralloc
 

ext/ycsb-tcd/ycsbc.cc

@@ -64,6 +64,8 @@ exit(1);
 
 #ifdef MONTAGE
   pds::init_thread(tid);
+#elif defined(RALLOC)
+  RP_set_tid(tid);
 #endif
 
   db->Init(tid);
@@ -90,9 +92,9 @@ unsigned cache_test_item_size = 0;
 int main(const int argc, char *argv[]) {
   // TODO: include GlobalTestConfig, construct command line args, and initialize EpochSys
 
-  memcached_init();
   utils::Properties props;
   string file_name = ParseCommandLine(argc, argv, props);
+  const int num_threads = stoi(props.GetProperty("threadcount", "1"));
 #ifdef MONTAGE
   // Ralloc init and close are already handled by memcached_init() 
   // init epoch system with artificial gtc, only for passing t and d
@@ -131,8 +133,9 @@ int main(const int argc, char *argv[]) {
   hwloc_get_type_depth(gtc.topology, HWLOC_OBJ_PU));
   std::cout<<"initial affinity built"<<std::endl;
   gtc.buildAffinity(gtc.affinities);
-  // pds::init(&gtc);
+  pds::init(&gtc);
 #endif
+  memcached_init(num_threads);
   if (do_cache_test_flag){
     do_cache_test();
     memcached_close();
@@ -146,8 +149,6 @@ int main(const int argc, char *argv[]) {
   ycsbc::CoreWorkload wl;
   wl.Init(props);
 
-  const int num_threads = stoi(props.GetProperty("threadcount", "1"));
-
   // Loads data
   vector<future<int>> actual_ops;
   int total_ops = stoi(props[ycsbc::CoreWorkload::RECORD_COUNT_PROPERTY]);

src/utils/Persistent.hpp

@@ -67,7 +67,11 @@ class Persistent {
 	}
 
 	static void init_thread(GlobalTestConfig* gtc, LocalTestConfig* ltc){
-		Ralloc::set_tid(ltc->tid);
+		init_thread(ltc->tid);
+	}
+
+	static void init_thread(int tid_){
+		Ralloc::set_tid(tid_);
 	}
 };