[BBLISA] Large file transfer in 65ms latency

[Daniel Hagerty]

"John Stoffel" <john at stoffel.org> writes:

> Absolutely not.  Read up on the Bandwidth Delay product.  If you crank
> up your TCP window size on both ends, you can mitigate this, but TCP
> was never designed with large latencies and large bandwidth at the
> core algorithm levels.  

    This.  The speed of light (information propogation) still matters.
For the link and latency in question, you have to have 8.3 megabytes in
flight (left the sender, not yet at the receiver) in order to fill the
end to end link.

    If I read the tea leaves of modern linux autotune defaults
correctly, the defaults don't allow for more than 4 megabytes in flight.
The things to search for, more specifically than tcp tuning (which will
doubtless get you here anyway), are SO_{SND,RCV}BUF.

> Nope, once TCP gets to a certain number of outstanding packets, it

    Well, bytes.  The receiver's kernel tells the sender how much space
is available in its receive buffer, and the protocol doesn't allow you
to have more in flight than this.