On causes of GridFTP transfer throughput variance

In prior work, we analyzed the GridFTP usage logs collected by data transfer nodes (DTNs) located at national scientific computing centers, and found significant throughput variance even among transfers between the same two end hosts. The goal of this work is to quantify the impact of various factors on throughput variance. Our methodology consisted of executing experiments on a high-speed research testbed, running large-sized instrumented transfers between operational DTNs, and creating statistical models from collected measurements. A non-linear regression model for memory-to-memory transfer throughput as a function of CPU usage at the two DTNs and packet loss rate was created. The model is useful for determining concomitant resource allocations to use in scheduling requests. For example, if a whole NERSC DTN CPU core can be assigned to the GridFTP process executing a large memory-to-memory transfer to SLAC, then only 32% of a CPU core is required at the SLAC DTN for the corresponding GridFTP process due to a difference in the computing speeds of these two DTNs. With these CPU allocations, data can be moved at 6.3 Gbps, which sets the rate to request from the circuit scheduler.