On 24/09/14 08:50, Alan Bateman wrote:
On 24/09/2014 02:19, Martin Sawicki (MS OPEN TECH) wrote:
Hello
We’re proposing an improvement to the OpenJDK which enables users to
take advantage of the TCP loopback fast path mechanism in Windows for
significantly higher performance of sockets whose both end points are
on the same machine. This is especially relevant in distributed
server-side/cloud scenarios, such as Hadoop.
We have the code figured out and tested internally. I’ve uploaded our
webrev package here (too big to send as an attachment):
https://openjdkcontrib.blob.core.windows.net/tcploopback/webrev-20140918.zip
The crux of the change lies in enabling the SIO_LOOPBACK_FAST_PATH
IOCTL flag on each socket creation call in Windows.
As for activating this optimization, we’re proposing the conservative
approach of keeping it off by default and adding a command line
argument to activate it:
-Dwindows.enableFastLocalTcpLoopback=true | false
(IMHO though, it’d be great to eventually just have it on by default).
We’d appreciate a review and acceptance of this improvement.
And, as we are new in the OpenJDK community and this may as well be
our first contribution to Java, I apologize for any steps in the
submission process that I may have missed here and would appreciate
guidance as needed.
Welcome, it's good to see Microsoft engineers on OpenJDK lists.
Yes, this looks like a useful contribution.
If I understand this ioctl correctly then it should only need to be
set once on a SOCKET. I'm curious about the listener oriented channels
(ServerSocketChannel etc.) where it is being set prior to each call to
accept. Is this needed? If not then I assume that we can just set it
when creating the SOCKET, sun.nio.ch.Net.socket0 of SocketChannel,
ServerSocketChannel, etc. You'll see that it already goes some Windows
specific setup and that might be the place to put it (and I think
should eliminate the need for most of the changes to the NIO code).
Right. The technet article suggests that it only needs to be set on the
listening socket (ie prior to calling accept()). So, the same
comment applies to the DualStackPlainSocketImpl change (could be moved
to socketListen() from socketAccept()).
The DualStack/TwoStacks code is being reorganised at the moment. So, I
think the patch will need to be rebased
against that work. But, the idea seems pretty reasonable to me.
- Michael
