November 20, 2004
Network bandwidth performance disparity across science applications
The network traffic on the HPWREN network originates in applications with drastically different bandwidth profiles. At one extreme, continuous low volume seismic sensors individually consume continuous resources, but often below 10 kilobits per second, as averaged during second-by-second calculations. Other applications, such as the transfer of optical images off the Palomar Observatory can consume all available bandwidth, but interspersed with no bandwidth needs at all while waiting for a new image, alongside significant night/day variations, with network resources predominantly used at night. Characteristics of various other applications generating significant traffic generally show a use of the space between such low and high margins.
An initial study to characterize the disparity across HPWREN applications in preparation for the QoS research of the new NSF HPWREN award is considering four applications and their bandwidth requirements. A continuation of such studies needs to set the ground work for the prioritization and re-prioritization based on actual bandwidth consumptions, real-time stringency, and appropriateness degree of a prioritization of specific traffic in the context of HPWREN research objectives. The four applications, measured via the HPWREN packet trace functionality across a 24 hour period on the November 19, 2004, comprise of:
The above graphic shows a time series graph of the four application during the 86,400 seconds of the day. Note that the Y axes are not normalized, they display the magnitude of the volume based on the maximum per-second data of the individual application. The bottom graph, the single seismic station, has a maximum of a few tens of kilobits per second, while the second graph from the bottom, the Palomar Observatory image traffic, exceeds 40 megabits per second.
This three orders of magnitude variation is more visible in the
This graphic shows the cumulative histograms of all four applications, with the X axis using a logarithmic scale. The disparities of network resource requirements is clearly visible, and will have to be well understood for traffic prioritization.
Another critical aspect for the prioritization is the degree of real-time requirements, alongside an understanding of a usable half-life of specific traffic, if not even in an extreme a consideration where traffic becomes useless if not delivered to the destination after a defined maximum latency. An example is a seismic application. A relatively low-bandwidth traffic set used for early warnings while outrunning a seismic shockwave at 5 kilometers a second will become useless after the shockwave arrived at the destination. On the other hand the same traffic may still be useful for other post-event analysis. The high-volume astronomy application tracking fast moving events may not have the sub-second stringency that the seismic application has, but still requires relatively quick delivery for real-time analysis, to support, for example, a followup to a supernova explosion by another telescope attached to a spectrometer in a location that may be a large distance away. Our Palomar collaborators have expressed interest in projects of even tighter time stringency, e.g., for followup to a gamma ray burst. As a third application, the SNMP network data collection is critical for network management and operations, and delays may hamper the detection of transient events.
back to top