MCC: A Predictable and Scalable Massive Client Load Generator

摘要

The network load generators are widely used by network researchers to analyze link bandwidth, evaluate network performance and test device capabilities. Data center and IoT networks are quickly evolving and we desire to get a load generator that can precisely generate flow-level workload with high-throughput. Often researchers choose software-based generators because of their flexibility and open-source nature. However, despite the emerging of different solutions, existing software-based flow-level generators have difficulty in generating millions of concurrent TCP connections or achieving one-microsecond precision of packet inter departure time (IDT) which can undermine the correctness of experiments. In this paper, we present a new network load generator, called Massive Client Connections (MCC). MCC is a client load generator which means it performs flow-level load simulation. We separate the control plane from the data plane and design a two-stage timer mechanism to get higher precision. To take full advantage of multicore processors, we utilize the shared-nothing multi-threaded model. Our evaluation demonstrates that MCC generates network load conforming to expected distribution with one-microsecond precision. Moreover, MCC shows definite scalability of throughput in multicore systems. And it is capable of generating more than three million concurrent TCP connections with ten CPU cores.