[c-nsp] [OT] Application Protocol Performance in low latency envrionments

[Brad Henshaw]

Ash Net wrote:

> The reason for performance degradation solely seems to be latency
related
> since there's tons of b/w available in the lab setup and over 10G
lanphy
> paths.

Generally it is latency, yes. Sadly in many cases those expensive WAN
acceleration devices are for the most part, munging TCP headers to
overcome shortcomings in performance caused by poor Layer 5-7 protocol
design.

If application developers took the issue of delay into account I'm sure
we wouldn't end up with business applications performing multiple small
TCP-based data transfers one after the other... or fifty transactions in
serial which might perform fine on a single host or a LAN with a few
microseconds of latency... but 50x20ms RTT WAN transactions ain't so
hot.

As an alternative to WAN acceleration devices, you can in many instances
look at tuning the TCP/IP stack of the host OS to support TCP window
scaling, different initial window sizes etc. to improve throughput. In
some situations this may benefit LAN performance also. This does
introduce other complex issues though, such as the fact that a stack
optimised for transmission of high-speed data across a 10GbE LAN has the
potential to hit low-speed WAN or Internet links with excessive initial
bursts of packets - a problem not generally introduced by the
aforementioned hardware devices. For some types of hosts (file servers)
this might not be an issue however.

I'll end my off-topic rant there.

Regards,
Brad