Biblio

The advancement of software-defined networking (SDN) technology is highly dependent on the successful transformations from in-house research ideas to real-life products. To enable such transformations, a testbed offering scalable and high fidelity networking environment for testing and evaluating new/existing designs is extremely valuable. Mininet, the most popular SDN emulator by far, is designed to achieve both accuracy and scalability by running unmodified code of network applications in lightweight Linux Containers. How- ever, Mininet cannot guarantee performance fidelity under high workloads, in particular when the number of concurrent active events is more than the number of parallel cores. In this project, we develop a lightweight virtual time system in Linux container and integrate the system with Mininet, so that all the containers have their own virtual clocks rather than using the physical system clock which reflects the se- rialized execution of multiple containers. With the notion of virtual time, all the containers perceive virtual time as if they run independently and concurrently. As a result, inter- actions between the containers and the physical system are artificially scaled, making a network appear to be ten times faster from the viewpoint of applications within the contain- ers than it actually is. We also design an adaptive virtual time scheduling subsystem in Mininet, which is responsible to balance the experiment speed and fidelity. Experimen- tal results demonstrate that embedding virtual time into Mininet significantly enhances its performance fidelity, and therefore, results in a useful platform for the SDN community to conduct scalable experiments with high fidelity.

Fat-tree topologies have been widely adopted as the communication network in data centers in the past decade. Nowa- days, high-performance computing (HPC) system designers are considering using fat-tree as the interconnection network for the next generation supercomputers. For extreme-scale computing systems like the data centers and supercomput- ers, the performance is highly dependent on the intercon- nection networks. In this paper, we present FatTreeSim, a PDES-based toolkit consisting of a highly scalable fat-tree network model, with the goal of better understanding the de- sign constraints of fat-tree networking architectures in data centers and HPC systems, as well as evaluating the applica- tions running on top of the network. FatTreeSim is designed to model and simulate large-scale fat-tree networks up to millions of nodes with protocol-level fidelity. We have con- ducted extensive experiments to validate and demonstrate the accuracy, scalability and usability of FatTreeSim. On Argonne Leadership Computing Facility’s Blue Gene/Q sys- tem, Mira, FatTreeSim is capable of achieving a peak event rate of 305 M/s for a 524,288-node fat-tree model with a total of 567 billion committed events. The strong scaling experiments use up to 32,768 cores and show a near linear scalability. Comparing with a small-scale physical system in Emulab, FatTreeSim can accurately model the latency in the same fat-tree network with less than 10% error rate for most cases. Finally, we demonstrate FatTreeSim’s usability through a case study in which FatTreeSim serves as the net- work module of the YARNsim system, and the error rates for all test cases are less than 13.7%.

Realistic and scalable testing systems are critical to evaluate network applications and protocols to ensure successful real system deployments. Container-based network emula- tion is attractive because of the combination of many desired features of network simulators and physical testbeds . The success of Mininet, a popular software- defined networking (SDN) emulation testbed, demonstrates the value of such approach that we can execute unmodified binary code on a large- scale emulated network with lightweight OS-level vir- tualization techniques. However, an ordinary network em- ulator uses the system clock across all the containers even if a container is not being scheduled to run. This leads to the issue of temporal fidelity, especially with high workloads. Virtual time sheds the light on the issue of preserving tem- poral fidelity for large-scale emulation. The key insight is to trade time with system resources via precisely scaling the time of interactions between containers and physical devices by a factor of n, hence, making an emulated network ap- pear to be n times faster from the viewpoints of applications in the container. In this paper, we develop a lightweight Linux-container-based virtual time system and integrate the system to Mininet for fidelity and scalability enhancement. We also design an adaptive time dilation scheduling mod- ule for balancing speed and accuracy. Experimental results demonstrate that (1) with virtual time, Mininet is able to accurately emulate a network n times larger in scale, where n is the scaling factor, with the system behaviors closely match data obtained from a physical testbed; and (2) with the adaptive time dilation scheduling, we reduce the running time by 46% with little accuracy loss. Finally, we present a case study using the virtual-time-enabled Mininet to evalu- ate the limitations of equal-cost multi-path (ECMP) routing in a data center network.

It is critical to ensure that network policy remains consistent during state transitions. However, existing techniques impose a high cost in update delay, and/or FIB space. We propose the Customizable Consistency Generator (CCG), a fast and generic framework to support customizable consistency policies during network updates. CCG effectively reduces the task of synthesizing an update plan under the constraint of a given consistency policy to a verification problem, by checking whether an update can safely be installed in the network at a particular time, and greedily processing network state transitions to heuristically minimize transition delay. We show a large class of consistency policies are guaranteed by this greedy heuristic alone; in addition, CCG makes judicious use of existing heavier-weight network update mechanisms to provide guarantees when necessary. As such, CCG nearly achieves the “best of both worlds”: the efficiency of simply passing through updates in most cases, with the consistency guarantees of more heavyweight techniques. Mininet and physical testbed evaluations demonstrate CCG’s capability to achieve various types of consistency, such as path and bandwidth properties, with zero switch memory overhead and up to a 3× delay reduction compared to previous solutions.

Today’s quality of life is highly dependent on the successful operation of many large-scale industrial control systems. To enhance their protection against cyber-attacks and operational errors, we develop a simulation-based verification framework with cross-layer verification techniques that allow comprehensive analysis of the entire ICS-specific stack, including application, protocol, and network layers.

In SDN, the underlying infrastructure is usually abstracted for applications that can treat the network as a logical or virtual entity. Commonly, the “mappings” between virtual abstractions and their actual physical implementations are not one-to-one, e.g., a single “big switch” abstract object might be implemented using a distributed set of physical devices. A key question is, what abstractions could be mapped to multiple physical elements while faithfully preserving their native semantics? E.g., can an application developer always expect her abstract “big switch” to act exactly as a physical big switch, despite being implemented using multiple physical switches in reality? We show that the answer to that question is “no” for existing virtual-to-physical mapping techniques: behavior can differ between the virtual “big switch” and the physical network, providing incorrect application-level behavior.

We also show that that those incorrect behaviors occur despite the fact that the most pervasive correctness invariants, such as per-packet consistency, are preserved throughout. These examples demonstrate that for practical notions of correctness, new systems and a new analytical framework are needed. We take the first steps by defining end-to-end correctness, a correctness condition that focuses on applications only, and outline a research vision to obtain virtualization systems with correct virtual to physical mappings.