Set send buffer size for the raw socket used by chromoting

Set send buffer size for the raw socket used by chromoting

On Windows the send buffer size of UDP socket has a much smaller value
than the PseudoTcp's send buffer size. This causes PseudoTcp to flood
the UDP socket and results in a EWOULDBLOCK, however the code doesn't
handle this properly but simply treat the packet as lost. This happens
very often and causes Windows host to be very janky.

BUG=91495
TEST=None

Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=95473

[Sergey Ulanov, 2011-08-03 19:12:08]

LGTM, but I don't think we need 256k buffer size for UDP send queue.

http://codereview.chromium.org/7508015/diff/1/remoting/protocol/jingle_stream...
File remoting/protocol/jingle_stream_connector.cc (right):

http://codereview.chromium.org/7508015/diff/1/remoting/protocol/jingle_stream...
remoting/protocol/jingle_stream_connector.cc:104:
socket->SetSendBufferSize(kTcpSendBufferSize);
I don't think it needs to be that big. Even 10k should be enough here (AFAIK
it's around 2k by default).

[Wez, 2011-08-03 20:13:58]

On 2011/08/03 19:12:08, sergeyu wrote:
> LGTM, but I don't think we need 256k buffer size for UDP send queue.
> 
>
http://codereview.chromium.org/7508015/diff/1/remoting/protocol/jingle_stream...
> File remoting/protocol/jingle_stream_connector.cc (right):
> 
>
http://codereview.chromium.org/7508015/diff/1/remoting/protocol/jingle_stream...
> remoting/protocol/jingle_stream_connector.cc:104:
> socket->SetSendBufferSize(kTcpSendBufferSize);
> I don't think it needs to be that big. Even 10k should be enough here (AFAIK
> it's around 2k by default).

Seconded.  We need a larger send buffer for stream sockets because buffer space
can't be re-used for new outgoing data until previously-sent data has been
acknowledged - there is no such requirement for unreliable sockets.

[Alpha Left Google, 2011-08-04 13:43:08]

Theoretically the maximum size we're going to send is 256k since we set out
PseudoTcp to be like so. The consequence of having UDP socket to return a
EWOULDBLOCK is too great (it takes several seconds to recover) that I would like
to gurantee it is set large enough to not suffer from it at all.

[Sergey Ulanov, 2011-08-04 18:10:28]

On 2011/08/04 13:43:08, Alpha wrote:
> Theoretically the maximum size we're going to send is 256k since we set out
> PseudoTcp to be like so. The consequence of having UDP socket to return a
> EWOULDBLOCK is too great (it takes several seconds to recover) that I would
like
> to gurantee it is set large enough to not suffer from it at all.

Ok, then, but it feels like we should really investigate it further to find out
what's the right buffer size. Can you please add TODO for that.