The following graphs show the result of a small study on the transfer speeds between the MIT Tier2 cluster and stargrid01/stargrid02.  For each point a 435 MB file was transferred with globus-url-copy 10 times, with the mean and RMS speeds chosen for display.  

Note the awkwardly large RMS values with larger stream multiplicities.  These are due to unknown overheads which prevent any transfer from occurring for a significant amount of time.  Hence a given sample with large nStreams will sometimes contain a few slow times and a few fast times.

Instantaneous speeds accelerated to a maximum then decelerated at the end of the transfer, so that the average speed is roughly 0.3-0.5 of the maximum instantaneous speed.  This  means  for nStreams > 10 or so the instantaneous speed reached 100 MB/s often.  Is this behavior to be expected?  Perhaps tuning of blocksize or similar parameters can avoid the slow down towards the end of the transfer?