normal state running ok

src/allreduce_robust.cc

@@ -26,12 +26,12 @@ void AllreduceRobust::Shutdown(void) {
   // need to sync the exec before we shutdown, do a pesudo check point
   // execute checkpoint, note: when checkpoint existing, load will not happen
   utils::Assert(RecoverExec(NULL, 0, ActionSummary::kCheckPoint, ActionSummary::kSpecialOp),
-                "check point must return true");
+                "Shutdown: check point must return true");
   // reset result buffer
   resbuf.Clear(); seq_counter = 0;  
   // execute check ack step, load happens here
   utils::Assert(RecoverExec(NULL, 0, ActionSummary::kCheckAck, ActionSummary::kSpecialOp),
-                "check ack must return true");    
+                "Shutdown: check ack must return true");    
   AllreduceBase::Shutdown();
 }
 /*!
@@ -201,9 +201,8 @@ void AllreduceRobust::CheckPoint(const utils::ISerializable *global_model,
       if (CheckAndRecover(TryCheckinLocalState(&local_rptr[new_version],
                                                &local_chkpt[new_version]))) break;
     }
-    // run the ack phase
-    utils::Assert(RecoverExec(NULL, 0, 0, ActionSummary::kLocalCheckAck),
-                  "check point must return true");
+    // run the ack phase, can be true or false
+    RecoverExec(NULL, 0, 0, ActionSummary::kLocalCheckAck);
     // switch pointer to new version
     local_chkpt_version = !local_chkpt_version;
   }
@@ -678,7 +677,7 @@ AllreduceRobust::ReturnType
 AllreduceRobust::TryGetResult(void *sendrecvbuf, size_t size, int seqno, bool requester) {
   // if minimum sequence requested is local check point ack,
   // this means all nodes have finished local check point, directly return
-  if (seqno == ActionSummary::kLocalCheckAck) return kSuccess;  
+  if (seqno == ActionSummary::kLocalCheckAck) return kSuccess;
   if (seqno == ActionSummary::kLocalCheckPoint) {
     // new version of local model
     int new_version = !local_chkpt_version;
@@ -972,8 +971,8 @@ AllreduceRobust::TryCheckinLocalState(std::vector<size_t> *p_local_rptr,
                      ring_prev, ring_next);
   if (succ != kSuccess) return succ;
   // update rptr
-  rptr.resize(n + 1);
-  for (int i = 1; i < n; ++i) {
+  rptr.resize(n + 2);
+  for (int i = 1; i <= n; ++i) {
     rptr[i + 1] = rptr[i] + sizes[i];
   }  
   chkpt.resize(rptr.back());
@@ -1013,7 +1012,7 @@ AllreduceRobust::RingPassing(void *sendrecvbuf_,
                              LinkRecord *read_link,
                              LinkRecord *write_link) {
   if (read_link == NULL || write_link == NULL || read_end == 0) return kSuccess;
-  utils::Assert(write_end <= read_end, "RingPassing: boundary check1");
+  utils::Assert(write_end <= read_end, "RingPassing: boundary check1, write_end=%lu, read_end=%lu", write_end, read_end);
   utils::Assert(read_ptr <= read_end, "RingPassing: boundary check2");
   utils::Assert(write_ptr <= write_end, "RingPassing: boundary check3");
   // take reference

src/mock.h

@@ -30,14 +30,16 @@ public:
     rabit::Allreduce<OP>(sendrecvbuf, count);
   }
 
-  inline bool LoadCheckPoint(utils::ISerializable *p_model) {
+inline int LoadCheckPoint(utils::ISerializable *global_model,
+                          utils::ISerializable *local_model) {
     utils::Assert(verify(loadCheckpoint), "[%d] error when loading checkpoint", rank);
-    return rabit::LoadCheckPoint(p_model);
+    return rabit::LoadCheckPoint(global_model, local_model);
   }
-
-  inline void CheckPoint(const utils::ISerializable &model) {
+  
+  inline void CheckPoint(const utils::ISerializable *global_model,
+                         const utils::ISerializable *local_model) {
     utils::Assert(verify(checkpoint), "[%d] error when checkpointing", rank);
-    rabit::CheckPoint(model);
+    rabit::CheckPoint(global_model, local_model);
   }
 
   inline void Broadcast(std::string *sendrecv_data, int root) {

src/rabit-inl.h

@@ -124,12 +124,14 @@ inline void Allreduce(DType *sendrecvbuf, size_t count) {
                      engine::mpi::GetType<DType>(), OP::kType);
 }
 // load latest check point
-inline int LoadCheckPoint(utils::ISerializable *p_model) {
-  return engine::GetEngine()->LoadCheckPoint(p_model);
+inline int LoadCheckPoint(utils::ISerializable *global_model,
+                          utils::ISerializable *local_model) {
+  return engine::GetEngine()->LoadCheckPoint(global_model, local_model);
 }
 // checkpoint the model, meaning we finished a stage of execution
-inline void CheckPoint(const utils::ISerializable &model) {
-  engine::GetEngine()->CheckPoint(&model);
+inline void CheckPoint(const utils::ISerializable *global_model,
+                       const utils::ISerializable *local_model) {
+  engine::GetEngine()->CheckPoint(global_model, local_model);
 }
 // return the version number of currently stored model
 inline int VersionNumber(void) {

src/rabit.h

@@ -84,7 +84,12 @@ template<typename OP, typename DType>
 inline void Allreduce(DType *sendrecvbuf, size_t count);
 /*!
  * \brief load latest check point
- * \param p_model pointer to the model
+ * \param global_model pointer to the globally shared model/state
+ *   when calling this function, the caller need to gauranttees that global_model
+ *   is the same in all nodes
+ * \param local_model pointer to local model, that is specific to current node/rank
+ *   this can be NULL when no local model is needed
+ * 
  * \return the version number of check point loaded
  *     if returned version == 0, this means no model has been CheckPointed
  *     the p_model is not touched, user should do necessary initialization by themselves
@@ -99,15 +104,24 @@ inline void Allreduce(DType *sendrecvbuf, size_t count);
  *
  * \sa CheckPoint, VersionNumber
  */
-inline int LoadCheckPoint(utils::ISerializable *p_model);
+inline int LoadCheckPoint(utils::ISerializable *global_model,
+                          utils::ISerializable *local_model = NULL);
 /*!
  * \brief checkpoint the model, meaning we finished a stage of execution
  *  every time we call check point, there is a version number which will increase by one
  * 
- * \param p_model pointer to the model
- * \sa LoadCheckPoint, VersionNumber
- */
-inline void CheckPoint(const utils::ISerializable &model);
+ * \param global_model pointer to the globally shared model/state
+ *   when calling this function, the caller need to gauranttees that global_model
+ *   is the same in all nodes
+ * \param local_model pointer to local model, that is specific to current node/rank
+ *   this can be NULL when no local state is needed
+   * NOTE: local_model requires explicit replication of the model for fault-tolerance, which will
+   *       bring replication cost in CheckPoint function. global_model do not need explicit replication.
+   *       So only CheckPoint with global_model if possible
+   * \sa LoadCheckPoint, VersionNumber
+   */
+inline void CheckPoint(const utils::ISerializable *global_model,
+                       const utils::ISerializable *local_model = NULL);
 /*!
  * \return version number of current stored model,
  *         which means how many calls to CheckPoint we made so far