Chris Mellon wrote:
On Mon, Jan 12, 2009 at 1:13 PM, Laszlo Nagy <gand...@shopzeus.com> wrote:
Hi All,
To make the long story short, I have a toy version of an ORB being
developed, and the biggest problem is slow network speed over TCP/IP.
There is an object called 'endpoint' on both sides, with incoming and
outgoing message queues. This endpoint object has a socket assigned, with
nodelay:
conn.setsockopt(socket.IPPROTO_TCP,socket.TCP_NODELAY,1)
The endpoint is running two separate threads - those are dedicated for
reading/writing messages from/into the socket object, as shown below:
def _process_incoming(self):
try:
while not self.stop_requested.isSet():
data = self.read_str()
while not self.stop_requested.isSet():
try:
self.incoming.put(data,1)
break
except orb.util.smartqueue.Full:
pass
if not self.stop_requested.isSet():
if self.router:
self.router.on_message_arrived(self)
except Exception, e:
if self.router:
if not isinstance(e,TransportClosedError):
self.router.logger.error(dumpexc(e))
self.router.unregister_endpoint(self)
self.shutdown()
raise SystemExit(0)
def _process_outgoing(self):
try:
while not self.stop_requested.isSet():
data_ok = False
while not self.stop_requested.isSet():
try:
data = self.outgoing.get(1)
data_ok = True
break
except orb.util.smartqueue.Empty:
pass
if data_ok:
self.write_str(data)
except Exception, e:
if self.router:
if not isinstance(e,TransportClosedError):
self.router.logger.error(dumpexc(e))
self.router.unregister_endpoint(self)
self.shutdown()
raise SystemExit(0)
The main point is that the sender does not need to wait for the message to
be actually written into the socket (unless the outgoing queue becomes
full).
When I try to send a message and receive an answer for it, I can only get as
much as 130 request+response message pairs per second. Apparently, it is the
same from messages size =77bytes to message size=16 Kbytes.
However, if I send 100 outgoing messages first, then read back all answers
then the speed goes up to 1300 message pairs/sec. I suspect that this has
something to do with TCP/IP. Since this will be used for RPC/RMI, it would
be very important to lower the time needed to exchange messages. Is there
any way I can speed this up?
It is very likely that nodelay is actually hurting you here.
Using the select module and doing non-blocking IO will be faster than
using threads for this as well.
You might also want to replace those 'pass' statements when smartqueue
is empty or full with time.sleep() to avoid busy waiting.
--
http://mail.python.org/mailman/listinfo/python-list
