Student use of the UF network has increased significantly over the last year.   This was primarily due to a decision by President John Lombardi that each UF student's electronic "rights" will include a permanent userID.   Students can register through the Web, have access to individual records on line via the Web, and have fifteen "free" hours of dial-up time per month to access the UF network and the Web.   As of the third week in February, 1998, just over 38,000 students had registered for these electronic rights by getting a GatorLink userID.  One measure of their increased use of the media was that about 9,500 of these students dialed into and through the UF network an average of 8.9 hours in January, 1998.   Two years earlier, 1,701 students used the UF's dial-up service.   This was a 5.6-fold increase in two years!  

     While no formal survey has been taken on this question, observation indicates almost all -- as of February, 1997 -- 4,038 UF faculty, 1,085 UF administrative and professional, and 5,562 "USPS" full-time-equivalent employees in office or information processing positions have micros on LANs used in their ongoing daily activities.   The UF will eventually have all its administrative and office procedures on 2-, 3-, or N-tier client/server applications.   The structure of a number of the current applications can be seen here.   These 10,685 employees are increasing their demands on the UF's network on the job days and at home nights.   The "at home nights" demand increases have also grown Significantly.   These users averaged 15.0 hours per month in January, 1996.   By January, 1998 these users averaged 20.2 hours of connect time per month.

     Demands on the UF network from users located other places in the State of Florida have also increased significantly. The State University System's (SUS) Florida Center for Library Automation ( FCLA) holds the "card catalogs" and an ever-growing inventory of on-line documents for all libraries in the SUS.   Users' terminal input/output on this system are growing at 96 million 2K bytes per month.   The UF network is a major site in the Florida Information Resource Network ( FIRN).  It, "...is an extensive telecommunications network accessible to all of Florida's public educators.   The fundamental goal of FIRN is to provide these educators with access to the computing resources which serve public education." (Department of Education, FIRN Yellow Pages, January 14, 1997, Page 2).   FIRN currently provides free dial access via 50 local calling areas throughout Florida and toll-free 800 access.  

     The UF has long supported national and international networking.   It was the 74th member of the original BITNET.   It is a member of the Internet 2 project.   The UF recently received a $350,000 grant the National Science Foundation (NSF) to support the UF's connection to the NSF's very high performance Backbone Network Service (vBNS); the connection site will be Georgia Tech.

     Many faculty and other users are awaiting the enhanced capacity Internet 2 will bring that will enable full-function real-time on-line audio and video services.   The College of Business at the UF will offer a "Flex" MBA program on the Web.   It will use multi-point on-line real-time audio and video connections to "hold class" when adequate network capacity is available to support this innovation in education.   The vBNS connection will enable UF scientists and engineers to collaborate with others across the country and share powerful computing and information resources via Internet 2.

     We believe the capacity of the existing FDDI ring-topology core has severely limited demand that would otherwise have taken place in the recent past.   This core is limited to a maximum of 100Mbps.   Utilization of the central FDDI ring is shown here, as is the utilization of the UF's 45 Mbps full-duplex link to the Internet via Jacksonville, FL.   We know of a number of cases where faculty and researchers did not attempt real-time audio and video applications, for example, because it was known, a priori, that performance would be so poor the result would not, in fact, be real-time response.

     Our estimate is that demand rates of up to 600 Mbps will occur with an alternative ATM 2-tier network hierarchy.   We are replacing the existing FDDI core with an ATM core using, initially, six IBM 8265 Nways ATM Switches.   The initial configuration of this core replacement, a partial mesh topology, is shown here.   It will give us OC12 throughput rates (622 Mbps) on the main paths that form the logical core replacement; they are shown in red. Some other major links on campus will be increased to give us OC3 rates (155 Mbps) where in the past we were limited to 100 Mbps.   The Internet 2 connection will give OC3 throughput.   The initial connection to vBNS will be at the OC12 level (NSF Advanced Networking Infrastructure and Research Division, NSF Approved 29 Connections to High-Performance Computer Network, Press Release, February 25, 1998).

     The upcoming core upgrade provides an excellent opportunity for research on

• communications technology in general and

• specific FDDI versus ATM core performance differences in a real production environment.

A. FDDI Tests:

1. A number (N > 30) of times during the normal weekday work periods will be randomly chosen by Dr. Newman, Dr. Elnicki, and Mr. Miller.   Total system traffic and performance measures will be taken at each of these time marks. The measures will be structured to permit analysis in performance models structured by Dr. Newman.  A typical work week will be chosen for these FDDI tests based on historical utilization data maintained by Dr. Elnicki.  Dr. Newman, Dr. Elnicki, and Mr. Miller will take the FDDI measures.

   Here (A.1), we will use standard volumetric models (bytes, packets) over various measurement interval sizes, and then determine the Hurst constant associated with the traffic seen (characterizes self-similar traffic).  In addition, if connection-based information is available, we will determine interarrival distribution, as well as volume (number of bytes, number of packets), length (in time), packet size, and packet interarrival time distributions for connections.

   We have found in the past that characterization of traffic at this higher level better represents behavior and explains what is seen at the strict volumetric level.  These characterizations will assist us in creating traffic generators that can simulate heavy loads, and load having different application mixes. In addition, measures that will be taken for A.2 will be taken during normal traffic to establish a baseline for them.

2. Dr. Newman and Mr. Miller will create a specific set of test traffic conditions to be run when the system is otherwise almost completely idle.   This has historically been early Sunday mornings.   This part of the research is referred to as A.2 below.

   The test traffic will permit performance measures that include delay and response time, as well as jitter. Jitter is important for asynchronous traffic. These will be measured between a wide sample of different points in the network, and will allow the characterization of delay attributable to the switches/routers and the intervening networks themselves.

3. Dr. Newman and Mr. Miller will create a set of traffic conditions designed to severely stress the system throughput.   These stress tests (A.3) will first use sheer volumetric overloading (traffic synthesis at peak rates for some number of hosts), using varying packet sizes (this allows us to distinguish byte volume from packet volume effects).   In addition, traffic adhering to more realistic models, such as self-similar traffic using the Hurst parameters ascertained in A.1 and connection-based traffic parameters also determined in A.1, will be generated to estimate the amount of traffic (bytes, packets, connections of various types) needed to cause the network to perform at various levels.

   Rather than tuning the traffic synthesis to produce a "desired level of inoperbility," several levels will be used to generate graphs depicting the response of the network to the varying loads and load types.  In this manner, "knees" and other sensitivities of the network to load may be discovered.

1. The same types of monitoring and parameter extraction will be done here (B.1) as was done for A.1, as described above.  By using a variety of measurement times, the samples may be compared statistically to determine if they are essentially the same.   We expect them to be the same, since we do not expect the change in core routing facilities to significantly change traffic in the short run.  

   Measures taken for A.2 as described above will again be taken during normal traffic after ATM installation for B.1.  The A.1 versus B.1 results will be compared with compensation for differences in traffic levels as necessary (in case we are wrong and do find significant differences in traffic by the time the network with the ATM core is brought to production mode).

2. The idle time performance measures taken with the network on the ATM core (B.2) will be the same as for A.2, as described above.  The two sets of results will then be compared.

3. The same set of stress tests will be run with the network on the ATM core (B.3).  But, we expect to have to extend the peak loads to a higher level in order to characterize fully the network's performance.   The performance measures will be graphed against load in order to determine critical load levels and sensitivities of the network under various types of traffic.   These will then be compared to those found for the network with the FDDI core, as described above in A.3.

• Performance measures taken on the production systems at the times during the typical work week days will be analyzed to determine whether significant differences exist and the extent to which those differences were due to varying traffic as compared to the change in the core technology.

• The performance differences with the specifically structured fixed set of traffic conditions run on the otherwise idle systems will provide a metric, albeit artificial, on relative performance.

• The relative differences in stress traffic levels of the FDDI system as compared to the ATM system will give a metric of the increase in the realistic capacity achieved by upgrading from a 100 Mbps FDDI core to a 622 Mbps ATM core.