Use our own implementation of parallel_memcopy (#7254)

python/ray/__init__.py

@@ -13,12 +13,6 @@ if "pickle5" in sys.modules:
                       "requires a specific version of pickle5 (which is "
                       "packaged along with Ray).")
 
-if "OMP_NUM_THREADS" not in os.environ:
-    logger.debug("[ray] Forcing OMP_NUM_THREADS=1 to avoid performance "
-                 "degradation with many workers (issue #6998). You can "
-                 "override this by explicitly setting OMP_NUM_THREADS.")
-    os.environ["OMP_NUM_THREADS"] = "1"
-
 # Add the directory containing pickle5 to the Python path so that we find the
 # pickle5 version packaged with ray and not a pre-existing pickle5.
 pickle5_path = os.path.join(

python/ray/_raylet.pyx

@@ -112,7 +112,7 @@ include "includes/libcoreworker.pxi"
 
 logger = logging.getLogger(__name__)
 
-MEMCOPY_THREADS = 12
+MEMCOPY_THREADS = 6
 
 
 def set_internal_config(dict options):

python/ray/includes/serialization.pxi

@@ -10,7 +10,7 @@ DEF kMajorBufferSize = 2048
 DEF kMemcopyDefaultBlocksize = 64
 DEF kMemcopyDefaultThreshold = 1024 * 1024
 
-cdef extern from "arrow/util/memory.h" namespace "arrow::internal" nogil:
+cdef extern from "ray/util/memory.h" namespace "ray" nogil:
     void parallel_memcopy(uint8_t* dst, const uint8_t* src, int64_t nbytes,
                           uintptr_t block_size, int num_threads)
 

python/ray/worker.py

@@ -66,12 +66,6 @@ ERROR_KEY_PREFIX = b"Error:"
 # entry/init points.
 logger = logging.getLogger(__name__)
 
-# Whether we should warn about slow put performance.
-if os.environ.get("OMP_NUM_THREADS") == "1":
-    should_warn_of_slow_puts = True
-else:
-    should_warn_of_slow_puts = False
-
 
 class ActorCheckpointInfo:
     """Information used to maintain actor checkpoints."""
@@ -275,22 +269,10 @@ class Worker:
                 "do this, you can wrap the ray.ObjectID in a list and "
                 "call 'put' on it (or return it).")
 
-        global should_warn_of_slow_puts
-        if should_warn_of_slow_puts:
-            start = time.perf_counter()
-
         serialized_value = self.get_serialization_context().serialize(value)
-        result = self.core_worker.put_serialized_object(
+        return self.core_worker.put_serialized_object(
             serialized_value, object_id=object_id, pin_object=pin_object)
 
-        if should_warn_of_slow_puts:
-            delta = time.perf_counter() - start
-            if delta > 0.1:
-                logger.warning("OMP_NUM_THREADS=1 is set, this may slow down "
-                               "ray.put() for large objects (issue #6998).")
-                should_warn_of_slow_puts = False
-        return result
-
     def deserialize_objects(self, data_metadata_pairs, object_ids):
         context = self.get_serialization_context()
         return context.deserialize_objects(data_metadata_pairs, object_ids)