RPC: optimization by caching

This reduced the calls needed for socket send/recv.
This commit is contained in:
pca006132 2020-08-27 11:27:40 +08:00 committed by Sébastien Bourdeauducq
parent 69f0699ebd
commit b2572003ac
1 changed files with 56 additions and 31 deletions

View File

@ -66,16 +66,16 @@ def _receive_list(kernel, embedding_map):
tag = chr(kernel._read_int8())
if tag == "b":
buffer = kernel._read(length)
return numpy.ndarray((length, ), 'B', buffer).tolist()
return list(buffer)
elif tag == "i":
buffer = kernel._read(4 * length)
return numpy.ndarray((length, ), '>i4', buffer).tolist()
return list(struct.unpack(">%sl" % length, buffer))
elif tag == "I":
buffer = kernel._read(8 * length)
return numpy.ndarray((length, ), '>i8', buffer).tolist()
return list(struct.unpack(">%sq" % length, buffer))
elif tag == "f":
buffer = kernel._read(8 * length)
return numpy.ndarray((length, ), '>d', buffer).tolist()
return list(struct.unpack(">%sd" % length, buffer))
else:
fn = receivers[tag]
elems = []
@ -178,6 +178,17 @@ class CommKernel:
self._read_type = None
self.host = host
self.port = port
self.read_buffer = bytearray()
self.write_buffer = bytearray()
self.unpack_int32 = struct.Struct(">l").unpack
self.unpack_int64 = struct.Struct(">q").unpack
self.unpack_float64 = struct.Struct(">d").unpack
self.pack_header = struct.Struct(">lB").pack
self.pack_int32 = struct.Struct(">l").pack
self.pack_int64 = struct.Struct(">q").pack
self.pack_float64 = struct.Struct(">d").pack
def open(self):
if hasattr(self, "socket"):
@ -198,13 +209,18 @@ class CommKernel:
#
def _read(self, length):
r = bytes()
while len(r) < length:
rn = self.socket.recv(min(8192, length - len(r)))
if not rn:
raise ConnectionResetError("Connection closed")
r += rn
return r
# cache the reads to avoid frequent call to recv
while len(self.read_buffer) < length:
# the number is just the maximum amount
# when there is not much data, it would return earlier
diff = length - len(self.read_buffer)
flag = 0
if diff > 8192:
flag |= socket.MSG_WAITALL
self.read_buffer += self.socket.recv(8192, flag)
result = self.read_buffer[:length]
self.read_buffer = self.read_buffer[length:]
return result
def _read_header(self):
self.open()
@ -212,14 +228,14 @@ class CommKernel:
# Wait for a synchronization sequence, 5a 5a 5a 5a.
sync_count = 0
while sync_count < 4:
(sync_byte, ) = struct.unpack("B", self._read(1))
sync_byte = self._read(1)[0]
if sync_byte == 0x5a:
sync_count += 1
else:
sync_count = 0
# Read message header.
(raw_type, ) = struct.unpack("B", self._read(1))
raw_type = self._read(1)[0]
self._read_type = Reply(raw_type)
logger.debug("receiving message: type=%r",
@ -235,19 +251,18 @@ class CommKernel:
self._read_expect(ty)
def _read_int8(self):
(value, ) = struct.unpack("B", self._read(1))
return value
return self._read(1)[0]
def _read_int32(self):
(value, ) = struct.unpack(">l", self._read(4))
(value, ) = self.unpack_int32(self._read(4))
return value
def _read_int64(self):
(value, ) = struct.unpack(">q", self._read(8))
(value, ) = self.unpack_int64(self._read(8))
return value
def _read_float64(self):
(value, ) = struct.unpack(">d", self._read(8))
(value, ) = self.unpack_float64(self._read(8))
return value
def _read_bool(self):
@ -264,7 +279,15 @@ class CommKernel:
#
def _write(self, data):
self.socket.sendall(data)
self.write_buffer += data
# if the buffer is already pretty large, send it
# the block size is arbitrary, tuning it may improve performance
if len(self.write_buffer) > 4096:
self._flush()
def _flush(self):
self.socket.sendall(self.write_buffer)
self.write_buffer.clear()
def _write_header(self, ty):
self.open()
@ -272,7 +295,7 @@ class CommKernel:
logger.debug("sending message: type=%r", ty)
# Write synchronization sequence and header.
self._write(struct.pack(">lB", 0x5a5a5a5a, ty.value))
self._write(self.pack_header(0x5a5a5a5a, ty.value))
def _write_empty(self, ty):
self._write_header(ty)
@ -281,19 +304,19 @@ class CommKernel:
self._write(chunk)
def _write_int8(self, value):
self._write(struct.pack("B", value))
self._write(value)
def _write_int32(self, value):
self._write(struct.pack(">l", value))
self._write(self.pack_int32(value))
def _write_int64(self, value):
self._write(struct.pack(">q", value))
self._write(self.pack_int64(value))
def _write_float64(self, value):
self._write(struct.pack(">d", value))
self._write(self.pack_float64(value))
def _write_bool(self, value):
self._write(struct.pack("B", value))
self._write(1 if value == True else 0)
def _write_bytes(self, value):
self._write_int32(len(value))
@ -308,6 +331,7 @@ class CommKernel:
def check_system_info(self):
self._write_empty(Request.SystemInfo)
self._flush()
self._read_header()
self._read_expect(Reply.SystemInfo)
@ -332,6 +356,7 @@ class CommKernel:
def load(self, kernel_library):
self._write_header(Request.LoadKernel)
self._write_bytes(kernel_library)
self._flush()
self._read_header()
if self._read_type == Reply.LoadFailed:
@ -341,6 +366,7 @@ class CommKernel:
def run(self):
self._write_empty(Request.RunKernel)
self._flush()
logger.debug("running kernel")
_rpc_sentinel = object()
@ -441,14 +467,11 @@ class CommKernel:
if tag_element == "b":
self._write(bytes(value))
elif tag_element == "i":
array = numpy.array(value, '>i4')
self._write(array.tobytes())
self._write(struct.pack(">%sl" % len(value), *value))
elif tag_element == "I":
array = numpy.array(value, '>i8')
self._write(array.tobytes())
self._write(struct.pack(">%sq" % len(value), *value))
elif tag_element == "f":
array = numpy.array(value, '>d')
self._write(array.tobytes())
self._write(struct.pack(">%sd" % len(value), *value))
else:
for elt in value:
tags_copy = bytearray(tags)
@ -524,6 +547,7 @@ class CommKernel:
self._write_bytes(return_tags)
self._send_rpc_value(bytearray(return_tags),
result, result, service)
self._flush()
except RPCReturnValueError as exn:
raise
except Exception as exn:
@ -569,6 +593,7 @@ class CommKernel:
self._write_int32(line)
self._write_int32(-1) # column not known
self._write_string(function)
self._flush()
def _serve_exception(self, embedding_map, symbolizer, demangler):
name = self._read_string()