Introduction and Motivation

Hinely[very rough draft] describes a method of sending large and small packets to attempt to measure network bandwidth. His observations are that the network bandwidth as measured by the "size vs. rtt" algorithm is greater when we start to exceed the network MTU.

In the "size vs. rtt" algorithm, we send a large packet (size S_l bits) with round-trip time T_l, and a small packet (size S_s) and RTT T_s. We then compute the bandwidth as (S_l - S_s) / (T_l - T_s). This assumes that the difference between the times is due entirely to the additional time needed to transmit the additional (S_l - S_s) bits. This assumption is not true in practice, see below for further details.

One of the uncertainties is whether the responding host responds to the large packet (a) when the first fragment is received, or (b) when the last fragment is received. My guess, from the fairly linear slope we see in the graphs [http://www.ics.hawaii.edu/ phinely/measuring-the-internet/rtt1/rtt.html], is that the responding host responds when the last packet is received - otherwise, a very large packet (e.g. 20,000 bytes) would take the same time to respond as a packet just larger than the MTU (e.g. 1500 bytes). I am assuming that the original sending host performs the fragmentation. The description of the experiment confirms this - I assume that the receiving host will reassemble the UDP packet before doing UDP processing and discovering that the destination port is "bogus" (i.e. has no process listening on it) and sending a "port unreachable" reply.

Interesting questions include

1.

why is the measured bandwidth greater when packets are fragmented, and

2.

which is the actual link bandwidth, the one when packets are or are not fragmented?

Wang [verbal communication] has studied the data and preliminary indications are that indeed the two slopes apparent at a visual inspection are internally consistent and therefore truly measure two different things. We derive exact formulas for the measured time in the two cases, and try to answer (a) and (b).

In partial answer to question 1, what is really measured is the round-trip time. When packets are sent un-fragmented, the packet must be received in its entirety before it can be forwarded by a router. When packets are sent in fragments, a router can forward the first fragment before the second and subsequent fragments have been received. This is equivalent to the difference between store-and-forward routing, where a packet must be stored in its entirety before it can be forwarded, and wormhole routing, where a router can start to forward a packet before the router has received the end of the packet.