ssh_channel_write often takes 15+ ms on Windows

[d0 <public@xxxxxxxxx>]

Hi,

I noticed a curious phenomenon with an application, which works and
performs fine on Linux, but showed consistently lower (by a factor of
2-3x) throughput on Windows. It streams data through a channel in 64K
chunks - this works splendidly on Linux. After investigating I noticed
each individual call to ssh_channel_write blocking for 15-16 ms
despite only around 1-2 ms of latency between the hosts.

I dug into the libssh core and noticed that ssh_channel_write in
effect sends a bunch of SSH packets (up to 32K, due to the inner
packet headers we actually end up with two large and a small 20 byte
packet for a 64K write) and then calls ssh_channel_flush, which waits
for the send buffer to deplete.

It seems like ssh_channel_flush is subject to the low-resolution
Windows timer of 15.625 ms and will block for a multiple of that, and
it often enough ends up blocking for those 15 ms, reducing throughput.

I just increased the buffer size passed to ssh_channel_write, which
works around this, as each 15 ms delay is amortized over a much larger
buffer.

I don't really think a change to libssh is necessary per se (since
enabling high-resolution timers on windows still has some drawbacks as
far as I know, like worse power consumption), I just wanted to leave
the story here in case someone else stumbles over a similar issue.

Cheers, Marian