Amin Vahdat, Ken Yocum, Kevin Walsh, Priya Mahadevan,
Dejan Kostic, Jeff Chase, and David Becker,
Proceedings of 5th Symposium on Operating Systems Design
and Implementation (OSDI)
December 2002.
Presentation: PowerPoint, PDF
ModelNet is an Internet emulator that allows researchers to evaluate Internet-scale distributed systems. The environment supports:
The system is based on dummynet and extended to improve accuracy and include multi-hop and multi-core emulation. ModelNet has two logical components: the core, which models the Internet as a collection of nodes and links (pipes), and the edge nodes, on which the user-specified OS and applications are run. Each instance of the target application is a virtual edge node (VN) with a unique IP in ModelNet. There may be multiple VNs per physical edge node. All the traffic is routed through the core. The network topology specified by the user can be simplified (distilled) to reduce emulation cost. An algorithm is presented to preserve last-mile (walk-in) links and interior (walk-out) links. The links in the core are emulated as a set of pipes, and for each VN pair the shortest path is pre-computed during the initialization phase. The packets are buffered at the entry core node and move through the pipes by reference only.
The system's throughput capacity can be improved by increasing the number of physical core nodes. However, communication between physical core nodes can become a bottleneck. Thus the scalability depends on the application communication characteristics and ability to optimally partition the network.
In order to evaluate adaptive Internet systems, the user can directly incorporate generators for competing traffic. This gives accurate for emulation of "background" cross traffic, but consumes resources at edge nodes and bandwidth at core. Alternatively pipe parameters can be modified during emulation to add faults or inject cross traffic, which has low overhead and scales independently of traffic rate. This approach does, however, not capture all details of Internet packet dynamics (e.g. slow start, bursty traffic) and is not responsive to congestion.
The authors note the following scalability issues: The traffic through the ModelNet core is limited to the cluster's physical internal bandwidth. Also, ModelNet must buffer up to the full bandwidth-delay product of target network. Further, the target network is assumed to have a perfect routing protocol, as the traffic in ModelNet is routed along the shortest path.
The results from ModelNet and NS2 are compared and shown to agree well. Each packet-hop was accurately emulated to within the granularity of the hardware timer, which was set at 100μs. Several case-studies are presented to illustrate the systems generality.