MCCBench: A C10M Benchmark Oriented to Interactive Network Services

With the explosive growth of IoT and other interactive network services, billions of devices are now connected, leading to highly fluctuating traffic and diverse QoS requirements for servers. This, coupled with the C10M problem, means benchmarks for interactive services should be able to handle millions of concurrency, bursty load and multiple QoS evaluation. However, existing general benchmarks for network services cannot fully meet these requirements. To address this issue, we propose MCCBench as a benchmark for high concurrent interactive network services. MCCBench includes a methodology for load generation, service framework, and service performance evaluation, allowing for the measurement of over 10 million concurrent connections, bursty loads, and labeling of requests with different service qualities. The performance evaluation metrics include tail latency measured on the server side, and long-lived concurrent connections. To implement MCCBench, we have developed an open-source toolset called MCCBench-IoT, which includes a load generator, an IoT service system based on a user-space network stack, and an accurate monitor for measuring tail latency. We verified MCCBench by building a testbed with MCCBench-IoT to emulate a typical IoT service, successfully testing tail latency under a concurrency of 10.2 million on a single server node. The testbed was scaled to 300 million concurrency with cluster configuration. By providing a comprehensive benchmark for high-concurrent interactive network services, MCCBench can help improve the quality of service for such services and enable better decision-making for network infrastructure design and optimization.