To conduct real-time analytics computations, big data stream processing engines are required to process unbounded data streams at millions of events per second. However, current streaming engines exhibit low throughput and high tuple processing latency. Performance engineering is complicated by the fact that streaming engines constitute complex distributed systems consisting of multiple nodes in the cloud. A profiling technique is required that is capable of measuring time durations at high accuracy across nodes. Standard clock synchronization techniques such as the network time protocol (NTP) are limited to millisecond accuracy, and hence cannot be used. We propose a profiling technique that relates the time-stamp counters (TSCs) of nodes to measure the duration of events in a streaming framework. The precision of the TSC relation determines the accuracy of the measured duration. The TSC relation is conducted in quiescent periods of the network to achieve accuracy in the tens of microseconds. We propose a throughput-controlled data generator to reliably determine the sustainable throughput of a streaming engine. To facilitate high-throughput data ingestion, we propose a concurrent object factory that moves the deserialization overhead of incoming data tuples off the critical path of the streaming framework. The evaluation of the proposed techniques within the Apache Storm streaming framework on the Google Compute Engine public cloud shows that data ingestion increases from $700$ $\text{k}$ to $4.68$ $\text{M}$ tuples per second, and that time durations can be profiled at a measurement accuracy of $92$ $\mu\text{s}$, which is three orders of magnitude higher than the accuracy of NTP, and one order of magnitude higher than prior work.

翻译：暂无翻译