Maximum Sustainable Throughput Prediction for Large-Scale Data Streaming Systems

In cloud-based stream processing services, the maximum sustainable throughput (MST) is defined as the maximum throughput that a system composed of a fixed number of virtual machines (VMs) can ingest indefinitely. If the incoming data rate exceeds the system’s MST, unprocessed data accumulates, eventually making the system inoperable. Thus, it is important for the service provider to keep the MST always larger than the incoming data rate by allocating a sufficient number of VMs. In this paper, we propose a cost-effective framework to predict MST values for a given number of VMs for stream processing applications with various scalability characteristics. Since it may be difficult to find one prediction model that works well for various stream processing applications, we first train several models using linear regression for each application. We then select the best-fitting model for the target application through the evaluation of extra MST samples. To save cost and time to collect MST samples while achieving high prediction accuracy, we statistically determine the most effective set of VMs within a budget. For evaluation, we use Intel’s Storm benchmarks running on Amazon EC2 cloud. Using up to 128 VMs, experiments show that the models trained by our framework predict MST values with up to 15.8% average prediction error. Further, we evaluate our prediction models with simulation-based elastic VM scheduling for a realistic data streaming workload. Simulation results show that with 20% over-provisioning, our framework is able to achieve less than 0.1% SLA violations for the majority of test applications. We save 36% cost compared to a static VM scheduling that covers the peak workload to achieve the same level of SLA violations.