In this report, we look at network Quality of Service issues related to a distributed application called the UNC Nanomanipulator system (Nano). Deployment of this application as a distributed system motivated research into network Quality of Service issues. These research issues are documented in detail in this report.
The nanoManipulator system is a virtual environment interface to an Atomic Force Microscope. The system simplifies the conduct of scientific experiments. By rebuilding this system as a distributed application across a network, we can provide scientists with a very powerful environment for scientific collaboration. This is a pioneering project that has been undertaken at UNC - Chapel Hill. Researchers at UNC - Chapel Hill are trying to build a campus wide working version of the distributed nanoManipulator system. Distributing the system across a network imposes quality of service constraints on the underlying network.
We are faced with many choices in trying to build the distributed nanoManipulator system. First of all, there are many options available for enforcing quality of service on network flows. Protection may be enforced by a quality-of-service-capable commercial bridge, switch or switch-router. Alternately, we may use a general purpose Unix based workstation configured as a router. Such a general purpose computer may use an active queue management scheme such as Class Based Thresholds, described in . We would like to have some concrete data in making this decision as to what scheme to choose. Closely related to this decision is the question how should the underlying network be configured. I have tried to help answer these questions with experimental data. These choices and the context in which we have made these decisions is explained in later sections of this chapter. As an example, we have found that tagging flows by layer 3 and layer 4 information gives more flexibility as compared to tagging flows by layer 2 information.
There are many issues involved in building a robust distributed Nanomanipulator system. At a high level, we need to answer the following questions:
The quality of service requirements of the distributed nanoManipulator system can be met in several ways. The Cabletron SSR switches provide sophisticated ways of enforcing quality of service guarantees on network flows. Many different levels of protection are supported and each level can be provided different bandwidth allocations. We are interested in knowing how to configure quality of service parameters on the Cabletron SSR to support nano flows. We are also interested in knowing how active queue management schemes implemented on Unix routers compare with hardware solutions to quality of service issues, such as that provided by the Cabletron SSR.
The other issue is how the campus wide network should be configured. Two possibilities arise. In the first case we may configure the entire network as a single broadcast domain. The entire network behaves as though it were a single shared bus Ethernet and each computer can access other computers on the network directly by the Ethernet address of the destination computer. There are no IP layer hops involved. We can do this using the ability to create Virtual Local Area Networks (VLANs), provided by the SSR 2000. This is described in more detail in the next chapter. The other option would be to configure the portion of the network local to each department as a VLAN and interconnect these VLANs using the routing facility provided by the SSR. In this case the computers local to each building would behave as though they were on the same Ethernet, but any packets exchanged between computers in different departments would have to be routed by one or more SSR units, and the packet would undergo at least one IP layer hop. Each of these options has its advantages and disadvantages. A related issue is whether to configure links interconnecting core switches and links between switches and end systems as full duplex or half duplex. As an example, a 10 megabit full duplex link would support 10 megabits of traffic, per second, in each direction. A similar half duplex link would support a total of 10 megabits per second summing up traffic in both directions. As we shall see, whether a link is configured as full duplex or half duplex has implications on quality of service guarantees.
I have designed experiments to answer the questions posed above. Chapter 2 discusses the strategy behind the experiments, provides an overview of the Cabletron SSRUs capabilities, explains the working of the CBT mechanism, and discusses the design of the experiments.
Chapter 3 lists all the experiments that were conducted, lists their outcomes and discusses what we learnt by running the experiment.
Chapter 4 provides the conclusion and summarizes the findings and provide recommendations for network configuration. This chapter also suggests further experiments that may be designed for further testing.
The appendix lists the source code for the data analysis tools and traffic generators.
Click here to get a PostScript copy of this thesis (1.5 MB uncompressed), or here to get a PDF copy of this thesis (1.0 MB uncompressed).
Last revised Sun Jul 23 17:50:34 EDT 2000 by jeffay at cs.unc.edu.