[core worker] Python core worker normal task submission (#5566)

This commit is contained in:
Edward Oakes
2019-09-14 13:02:53 -07:00
committed by Eric Liang
parent 5f88823c49
commit a5d7de6aaf
11 changed files with 194 additions and 80 deletions
+66 -5
View File
@@ -27,6 +27,8 @@ from ray.includes.common cimport (
CRayObject,
CRayStatus,
CGcsClientOptions,
CTaskArg,
CRayFunction,
LocalMemoryBuffer,
LANGUAGE_CPP,
LANGUAGE_JAVA,
@@ -46,7 +48,7 @@ from ray.includes.unique_ids cimport (
CObjectID,
CClientID,
)
from ray.includes.libcoreworker cimport CCoreWorker
from ray.includes.libcoreworker cimport CCoreWorker, CTaskOptions
from ray.includes.task cimport CTaskSpec
from ray.includes.ray_config cimport RayConfig
from ray.exceptions import RayletError, ObjectStoreFullError
@@ -96,6 +98,13 @@ cdef int check_status(const CRayStatus& status) nogil except -1:
raise RayletError(message)
cdef VectorToObjectIDs(const c_vector[CObjectID] &object_ids):
result = []
for i in range(object_ids.size()):
result.append(ObjectID(object_ids[i].Binary()))
return result
cdef c_vector[CObjectID] ObjectIDsToVector(object_ids):
"""A helper function that converts a Python list of object IDs to a vector.
@@ -244,6 +253,16 @@ cdef unordered_map[c_string, double] resource_map_from_dict(resource_map):
return out
cdef c_vector[c_string] string_vector_from_list(list string_list):
cdef:
c_vector[c_string] out
for s in string_list:
if not isinstance(s, bytes):
raise TypeError("string_list elements must be bytes")
out.push_back(s)
return out
cdef class RayletClient:
cdef CRayletClient* client
@@ -400,6 +419,10 @@ cdef class CoreWorker:
"outside _raylet. See __init__.py for "
"details.")
def disconnect(self):
with nogil:
self.core_worker.get().Disconnect()
def get_objects(self, object_ids, TaskID current_task_id):
cdef:
c_vector[shared_ptr[CRayObject]] results
@@ -538,6 +561,48 @@ cdef class CoreWorker:
with nogil:
self.core_worker.get().SetCurrentJobId(c_job_id)
def submit_task(self,
function_descriptor,
args,
int num_return_vals,
resources):
cdef:
unordered_map[c_string, double] c_resources
CTaskOptions task_options
CRayFunction ray_function
c_vector[CTaskArg] args_vector
c_vector[CObjectID] return_ids
c_string pickled_str
shared_ptr[CBuffer] arg_data
shared_ptr[CBuffer] arg_metadata
c_resources = resource_map_from_dict(resources)
task_options = CTaskOptions(num_return_vals, c_resources)
ray_function = CRayFunction(
LANGUAGE_PYTHON, string_vector_from_list(function_descriptor))
for arg in args:
if isinstance(arg, ObjectID):
args_vector.push_back(
CTaskArg.PassByReference((<ObjectID>arg).native()))
else:
pickled_str = pickle.dumps(
arg, protocol=pickle.HIGHEST_PROTOCOL)
arg_data = dynamic_pointer_cast[CBuffer, LocalMemoryBuffer](
make_shared[LocalMemoryBuffer](
<uint8_t*>(pickled_str.data()),
pickled_str.size(),
True))
args_vector.push_back(
CTaskArg.PassByValue(
make_shared[CRayObject](arg_data, arg_metadata)))
with nogil:
check_status(self.core_worker.get().Tasks().SubmitTask(
ray_function, args_vector, task_options, &return_ids))
return VectorToObjectIDs(return_ids)
def set_object_store_client_options(self, c_string client_name,
int64_t limit_bytes):
with nogil:
@@ -552,7 +617,3 @@ cdef class CoreWorker:
message = self.core_worker.get().Objects().MemoryUsageString()
return message.decode("utf-8")
def disconnect(self):
with nogil:
self.core_worker.get().Disconnect()
+50 -6
View File
@@ -2,11 +2,13 @@ from libcpp cimport bool as c_bool
from libcpp.memory cimport shared_ptr
from libcpp.string cimport string as c_string
from libc.stdint cimport uint8_t
from libc.stdint cimport uint8_t, uint64_t, int64_t
from libcpp.unordered_map cimport unordered_map
from libcpp.vector cimport vector as c_vector
from ray.includes.unique_ids cimport (
CActorID,
CActorHandleID,
CJobID,
CWorkerID,
CObjectID,
@@ -138,15 +140,57 @@ cdef extern from "ray/common/buffer.h" namespace "ray" nogil:
size_t Size() const
cdef cppclass LocalMemoryBuffer(CBuffer):
LocalMemoryBuffer(uint8_t *data, size_t size)
LocalMemoryBuffer(uint8_t *data, size_t size, c_bool copy_data)
cdef extern from "ray/core_worker/store_provider/store_provider.h" nogil:
cdef extern from "ray/common/ray_object.h" nogil:
cdef cppclass CRayObject "ray::RayObject":
const shared_ptr[CBuffer] &GetData()
const size_t DataSize() const
const shared_ptr[CBuffer] &GetMetadata() const
c_bool HasData() const
c_bool HasMetadata() const
const size_t DataSize() const
const shared_ptr[CBuffer] &GetData()
const shared_ptr[CBuffer] &GetMetadata() const
cdef extern from "ray/core_worker/common.h" nogil:
cdef cppclass CRayFunction "ray::RayFunction":
CRayFunction()
CRayFunction(CLanguage language,
const c_vector[c_string] function_descriptor)
CLanguage GetLanguage()
c_vector[c_string] GetFunctionDescriptor()
cdef cppclass CTaskArg "ray::TaskArg":
@staticmethod
CTaskArg PassByReference(const CObjectID &object_id)
@staticmethod
CTaskArg PassByValue(const shared_ptr[CRayObject] &data)
cdef extern from "ray/core_worker/task_interface.h" nogil:
cdef cppclass CTaskOptions "ray::TaskOptions":
CTaskOptions()
CTaskOptions(int num_returns,
unordered_map[c_string, double] &resources)
cdef cppclass CActorCreationOptions "ray::ActorCreationOptions":
CActorCreationOptions(uint64_t max_reconstructions,
const unordered_map[c_string, double] &resources)
cdef cppclass CActorHandle "ray::ActorHandle":
CActorHandle(
const CActorID &actor_id, const CActorHandleID &actor_handle_id,
const CLanguage actor_language,
const c_vector[c_string] &actor_creation_task_function_descriptor)
CActorHandle(const CActorHandle &other)
CActorID ActorID() const
CActorHandleID ActorHandleID() const
c_vector[c_string] ActorCreationTaskFunctionDescriptor() const
CObjectID ActorCursor() const
int64_t TaskCursor() const
int64_t NumForks() const
CActorHandle Fork()
void Serialize(c_string *output)
CActorHandle Deserialize(const c_string &data)
cdef extern from "ray/gcs/gcs_client_interface.h" nogil:
cdef cppclass CGcsClientOptions "ray::gcs::GcsClientOptions":
+23 -4
View File
@@ -1,6 +1,6 @@
from libc.stdint cimport int64_t
from libcpp cimport bool as c_bool
from libcpp.memory cimport shared_ptr
from libcpp.memory cimport shared_ptr, unique_ptr
from libcpp.string cimport string as c_string
from libcpp.vector cimport vector as c_vector
@@ -10,9 +10,14 @@ from ray.includes.unique_ids cimport (
CObjectID,
)
from ray.includes.common cimport (
CActorCreationOptions,
CActorHandle,
CBuffer,
CRayStatus,
CRayFunction,
CRayObject,
CRayStatus,
CTaskArg,
CTaskOptions,
CWorkerType,
CLanguage,
CGcsClientOptions,
@@ -20,6 +25,20 @@ from ray.includes.common cimport (
from ray.includes.libraylet cimport CRayletClient
cdef extern from "ray/core_worker/task_interface.h" namespace "ray" nogil:
cdef cppclass CTaskSubmissionInterface "CoreWorkerTaskInterface":
CRayStatus SubmitTask(
const CRayFunction &function, const c_vector[CTaskArg] &args,
const CTaskOptions &options, c_vector[CObjectID] *return_ids)
CRayStatus CreateActor(
const CRayFunction &function, const c_vector[CTaskArg] &args,
const CActorCreationOptions &options,
unique_ptr[CActorHandle] *handle)
CRayStatus SubmitActorTask(
CActorHandle &handle, const CRayFunction &function,
const c_vector[CTaskArg] &args, const CTaskOptions &options,
c_vector[CObjectID] *return_ids)
cdef extern from "ray/core_worker/object_interface.h" nogil:
cdef cppclass CObjectInterface "ray::CoreWorkerObjectInterface":
CRayStatus SetClientOptions(c_string client_name, int64_t limit)
@@ -50,13 +69,13 @@ cdef extern from "ray/core_worker/core_worker.h" nogil:
CWorkerType &GetWorkerType()
CLanguage &GetLanguage()
CObjectInterface &Objects()
# CTaskSubmissionInterface &Tasks()
CTaskSubmissionInterface &Tasks()
# CTaskExecutionInterface &Execution()
# TODO(edoakes): remove this once the raylet client is no longer used
# directly.
CRayletClient &GetRayletClient()
# TODO(edoakes): remove this once the Python core worker uses the task
# TODO(edoakes): remove these once the Python core worker uses the task
# interfaces
void SetCurrentJobId(const CJobID &job_id)
void SetCurrentTaskId(const CTaskID &task_id)
+7 -4
View File
@@ -707,11 +707,14 @@ class Worker(object):
self.current_job_id, self.current_task_id,
self.task_context.task_index, actor_id)
else:
# This is a normal task.
task_id = TaskID.for_normal_task(self.current_job_id,
self.current_task_id,
self.task_context.task_index)
# Normal tasks are submitted through the core worker (in the
# future, all tasks will be).
return self.core_worker.submit_task(function_descriptor_list,
args_for_raylet,
num_return_vals, resources)
# Actor creation tasks and actor tasks are submitted directly to
# the raylet.
task = ray._raylet.TaskSpec(
task_id,
job_id,