[KVStore] Add fast-pull for kvstore (#1647)

* add fast-pull

* update

* add fast-pull

* update

* update

* update

* update test

* update

* update

* update

* update

* update

* update

* update

* update

* add omp

* update

* update

python/dgl/distributed/kvstore.py

@@ -1032,9 +1032,18 @@ class KVClient(object):
         tensor
             a data tensor with the same row size of id_tensor.
         """
-        #TODO(chao) : add C++ rpc interface and add fast pull
         assert len(name) > 0, 'name cannot be empty.'
         assert F.ndim(id_tensor) == 1, 'ID must be a vector.'
+        if self._pull_handlers[name] is default_pull_handler: # Use fast-pull
+            part_id = self._part_policy[name].to_partid(id_tensor)
+            return rpc.fast_pull(name, id_tensor, part_id, KVSTORE_PULL,
+                                 self._machine_count,
+                                 self._group_count,
+                                 self._machine_id,
+                                 self._client_id,
+                                 self._data_store[name],
+                                 self._part_policy[name])
+        else:
             # partition data
             machine_id = self._part_policy[name].to_partid(id_tensor)
             # sort index by machine id

python/dgl/distributed/rpc.py

@@ -14,7 +14,7 @@ __all__ = ['set_rank', 'get_rank', 'Request', 'Response', 'register_service', \
 'receiver_wait', 'add_receiver_addr', 'sender_connect', 'read_ip_config', \
 'get_num_machines', 'set_num_machines', 'get_machine_id', 'set_machine_id', \
 'send_request', 'recv_request', 'send_response', 'recv_response', 'remote_call', \
-'send_request_to_machine', 'remote_call_to_machine']
+'send_request_to_machine', 'remote_call_to_machine', 'fast_pull']
 
 REQUEST_CLASS_TO_SERVICE_ID = {}
 RESPONSE_CLASS_TO_SERVICE_ID = {}
@@ -844,6 +844,55 @@ def finalize_server():
     finalize_receiver()
     print("Server (%d) shutdown." % get_rank())
 
+def fast_pull(name, id_tensor, part_id, service_id,
+              machine_count, group_count, machine_id,
+              client_id, local_data, policy):
+    """Fast-pull api used by kvstore.
+
+    Parameters
+    ----------
+    name : str
+        data name
+    id_tensor : tensor
+        data ID
+    part_id : tensor
+        partition ID of id_tensor
+    service_id : int
+        service_id of pull request
+    machine_count : int
+        total number of machine
+    group_count : int
+        total number of server inside machine
+    machine_id : int
+        current machine ID
+    client_id : int
+        current client ID
+    local_data : tensor
+        local data tensor
+    policy : PartitionPolicy
+        store the partition information
+    """
+    msg_seq = incr_msg_seq()
+    pickle_data = bytearray(pickle.dumps(([0], [name])))
+    global_id = _CAPI_DGLRPCGetGlobalIDFromLocalPartition(F.zerocopy_to_dgl_ndarray(id_tensor),
+                                                          F.zerocopy_to_dgl_ndarray(part_id),
+                                                          machine_id)
+    global_id = F.zerocopy_from_dgl_ndarray(global_id)
+    g2l_id = policy.to_local(global_id)
+    res_tensor = _CAPI_DGLRPCFastPull(name,
+                                      int(machine_id),
+                                      int(machine_count),
+                                      int(group_count),
+                                      int(client_id),
+                                      int(service_id),
+                                      int(msg_seq),
+                                      pickle_data,
+                                      F.zerocopy_to_dgl_ndarray(id_tensor),
+                                      F.zerocopy_to_dgl_ndarray(part_id),
+                                      F.zerocopy_to_dgl_ndarray(g2l_id),
+                                      F.zerocopy_to_dgl_ndarray(local_data))
+    return F.zerocopy_from_dgl_ndarray(res_tensor)
+
 ############### Some basic services will be defined here #############
 
 CLIENT_REGISTER = 22451

python/dgl/distributed/rpc_server.py

@@ -76,6 +76,7 @@ def start_server(server_id, ip_config, num_clients, server_state, \
             rpc.send_response(client_id, register_res)
     # main service loop
     while True:
+        try:
             req, client_id = rpc.recv_request()
             res = req.process_request(server_state)
             if res is not None:
@@ -87,3 +88,12 @@ def start_server(server_id, ip_config, num_clients, server_state, \
                     break # break the loop and exit server
                 else:
                     rpc.send_response(client_id, res)
+        except KeyboardInterrupt:
+            print("Exit kvserver!")
+            rpc.finalize_sender()
+            rpc.finalize_receiver()
+        except:
+            print("Error on kvserver!")
+            rpc.finalize_sender()
+            rpc.finalize_receiver()
+            raise

src/rpc/rpc.cc

@@ -7,6 +7,8 @@
 
 #include <dgl/runtime/container.h>
 #include <dgl/packed_func_ext.h>
+#include <dgl/array.h>
+#include <dgl/random.h>
 #include <dgl/zerocopy_serializer.h>
 #include "../c_api_common.h"
 
@@ -34,6 +36,9 @@ RPCStatus SendRPCMessage(const RPCMessage& msg, const int32_t target_id) {
   for (auto ptr : zc_write_strm.buffer_list()) {
     network::Message ndarray_data_msg;
     ndarray_data_msg.data = reinterpret_cast<char*>(ptr.data);
+    if (ptr.size == 0) {
+      LOG(FATAL) << "Cannot send a empty NDArray.";
+    }
     ndarray_data_msg.size = ptr.size;
     NDArray tensor = ptr.tensor;
     ndarray_data_msg.deallocator = [tensor](network::Message*) {};
@@ -295,5 +300,130 @@ DGL_REGISTER_GLOBAL("distributed.server_state._CAPI_DGLRPCGetServerState")
   *rv = st;
 });
 
+//////////////////////////// KVStore ////////////////////////////
+
+DGL_REGISTER_GLOBAL("distributed.rpc._CAPI_DGLRPCGetGlobalIDFromLocalPartition")
+.set_body([] (DGLArgs args, DGLRetValue* rv) {
+  NDArray ID = args[0];
+  NDArray part_id = args[1];
+  int local_machine_id = args[2];
+  int64_t* ID_data = static_cast<int64_t*>(ID->data);
+  int64_t* part_id_data = static_cast<int64_t*>(part_id->data);
+  int64_t ID_size = ID.GetSize() / sizeof(int64_t);
+  std::vector<int64_t> global_id;
+  for (int64_t i = 0; i < ID_size; ++i) {
+    if (part_id_data[i] == local_machine_id) {
+      global_id.push_back(ID_data[i]);
+    }
+  }
+  NDArray res_tensor = dgl::aten::VecToIdArray<int64_t>(global_id);
+  *rv = res_tensor;
+});
+
+DGL_REGISTER_GLOBAL("distributed.rpc._CAPI_DGLRPCFastPull")
+.set_body([] (DGLArgs args, DGLRetValue* rv) {
+  // Input
+  std::string name = args[0];
+  int local_machine_id = args[1];
+  int machine_count = args[2];
+  int group_count = args[3];
+  int client_id = args[4];
+  int service_id = args[5];
+  int msg_seq = args[6];
+  std::string pickle_data = args[7];
+  NDArray ID = args[8];
+  NDArray part_id = args[9];
+  NDArray local_id = args[10];
+  NDArray local_data = args[11];
+  // Data
+  dgl_id_t ID_size = ID.GetSize() / sizeof(dgl_id_t);
+  dgl_id_t* ID_data = static_cast<dgl_id_t*>(ID->data);
+  dgl_id_t* part_id_data = static_cast<dgl_id_t*>(part_id->data);
+  dgl_id_t* local_id_data = static_cast<dgl_id_t*>(local_id->data);
+  char* local_data_char = static_cast<char*>(local_data->data);
+  std::vector<dgl_id_t> local_ids;
+  std::vector<dgl_id_t> local_ids_orginal;
+  std::vector<int64_t> local_data_shape;
+  std::vector<std::vector<dgl_id_t> > remote_ids(machine_count);
+  std::vector<std::vector<dgl_id_t> > remote_ids_original(machine_count);
+  // Get row size (in bytes)
+  int row_size = 1;
+  for (int i = 0; i < local_data->ndim; ++i) {
+    local_data_shape.push_back(local_data->shape[i]);
+    if (i != 0) {
+      row_size *= local_data->shape[i];
+    }
+  }
+  row_size *= (local_data->dtype.bits / 8);
+  size_t data_size = local_data.GetSize();
+  CHECK_GT(local_data_shape.size(), 0);
+  CHECK_EQ(row_size * local_data_shape[0], data_size);
+  // Get local id (used in local machine) and
+  // remote id (send to remote machine)
+  dgl_id_t idx = 0;
+  for (dgl_id_t i = 0; i < ID_size; ++i) {
+    dgl_id_t p_id = part_id_data[i];
+    if (p_id == local_machine_id) {
+      dgl_id_t l_id = local_id_data[idx++];
+      CHECK_LT(l_id, local_data_shape[0]);
+      CHECK_GE(l_id, 0);
+      local_ids.push_back(l_id);
+      local_ids_orginal.push_back(i);
+    } else {
+      CHECK_LT(p_id, machine_count) << "Invalid partition ID.";
+      dgl_id_t id = ID_data[i];
+      remote_ids[p_id].push_back(id);
+      remote_ids_original[p_id].push_back(i);
+    }
+  }
+  // Send remote id
+  int msg_count = 0;
+  for (int i = 0; i < remote_ids.size(); ++i) {
+    if (remote_ids[i].size() != 0) {
+      RPCMessage msg;
+      msg.service_id = service_id;
+      msg.msg_seq = msg_seq;
+      msg.client_id = client_id;
+      int lower = i*group_count;
+      int upper = (i+1)*group_count;
+      msg.server_id = dgl::RandomEngine::ThreadLocal()->RandInt(lower, upper);
+      msg.data = pickle_data;
+      NDArray tensor = dgl::aten::VecToIdArray<dgl_id_t>(remote_ids[i]);
+      msg.tensors.push_back(tensor);
+      SendRPCMessage(msg, msg.server_id);
+      msg_count++;
+    }
+  }
+  local_data_shape[0] = ID_size;
+  NDArray res_tensor = NDArray::Empty(local_data_shape,
+                                      local_data->dtype,
+                                      DLContext{kDLCPU, 0});
+  char* return_data = static_cast<char*>(res_tensor->data);
+  // Copy local data
+#pragma omp parallel for
+  for (int64_t i = 0; i < local_ids.size(); ++i) {
+    CHECK_GE(ID_size*row_size, local_ids_orginal[i]*row_size+row_size);
+    CHECK_GE(data_size, local_ids[i] * row_size + row_size);
+    CHECK_GE(local_ids[i], 0);
+    memcpy(return_data + local_ids_orginal[i] * row_size,
+           local_data_char + local_ids[i] * row_size,
+           row_size);
+  }
+  // Recv remote message
+  for (int i = 0; i < msg_count; ++i) {
+    RPCMessage msg;
+    RecvRPCMessage(&msg, 0);
+    int part_id = msg.server_id / group_count;
+    char* data_char = static_cast<char*>(msg.tensors[0]->data);
+    dgl_id_t id_size = remote_ids[part_id].size();
+    for (size_t n = 0; n < id_size; ++n) {
+      memcpy(return_data + remote_ids_original[part_id][n] * row_size,
+             data_char + n * row_size,
+             row_size);
+    }
+  }
+  *rv = res_tensor;
+});
+
 }  // namespace rpc
 }  // namespace dgl

tests/distributed/test_new_kvstore.py

@@ -88,7 +88,7 @@ def init_zero_func(shape, dtype):
     return F.zeros(shape, dtype, F.cpu())
 
 def udf_push(target, name, id_tensor, data_tensor):
-    target[name] = F.scatter_row(target[name], id_tensor, data_tensor*data_tensor)    
+    target[name][id_tensor] = data_tensor * data_tensor 
 
 @unittest.skipIf(os.name == 'nt' or os.getenv('DGLBACKEND') == 'tensorflow', reason='Do not support windows and TF yet')
 def test_partition_policy():