[Bloat] some market observations

[Srikanth Sundaresan]

On Fri, Mar 11, 2011 at 8:22 AM,  <BeckW at telekom.de> wrote:
> interesting discussions going on here..
>
> A consumer-oriented magazine complained that a VDSL access rated 50 mbps only delivered 31 mbps. It turned out that they used a single TCP session for their measurements. At those speeds, delay and packet loss determine the maximum throughput, independent of the available bandwidth. The formula in RfC 3819, Section 8.5 yielded an estimate of ~32 mbps for the given delay and loss, which matched nicely the actual measurements.
>

The same effects can be seen even on smaller pipes. I see about a
~0.5Mbits/s gap between multi-threaded and single threaded
measurements in my 3Mbits/s DSL line at home. (Loss ~0.7%)

> With increasing access bandwidths, ISPs are forced to look into the delay issues.
>
> DSL bit error rates go up when everybody switches to DSL. DSL standard bodies look into link layer mechanisms to mitigate this, but all solutions add delay. It's hard to find the right balance between delay and packet loss; both hurt TCP performance. As delay can't be controlled by a single ISP, the temptation is high to focus on the bit errors.
>

To a very large extent, it can be, as most of the delay is in the
local loop. For DSL, it is at least 8ms and can go up to 40-50 ms.
It's a significant fraction of end-to-end latency for most popular
websites. Error correction techniques make this worse.

- Srikanth