Jarvis: Large-scale Server Monitoring with Adaptive Near-data Processing

Abstract

Rapid detection of performance and reliability problems is critical to satisfy availability requirements of a datacenter, which typically consists of a large numbers of servers. To address this, we propose Jarvis, a new stream monitoring system that adaptively partitions monitoring queries for execution across data source and stream processor. Jarvis employs two novel schemes to meet the requirements of large-scale server monitoring scenarios. First, Jarvis processes a subset of the input records on each operator through fine-grained data-level partitioning, allowing even resource-intensive operators to execute under resource constraints on each data source and contribute to reducing the amount of data to transmit over the network to the stream processor. Second, Jarvis makes fast and fully decentralized near-data query refinement decisions guided by an approach combining model-based and model-free heuristics, enabling quick adaptation to dynamic resource conditions on each data source. We evaluate the effectiveness of Jarvis on a diverse set of monitoring queries and scales. As compared to the existing schemes, Jarvis handles up to 75% more data source nodes while improving throughput in resource-constrained scenarios by 1.2–4.4×. Moreover, Jarvis achieves these improvements while being able to converge to a stable query partition within a few seconds of a resource change that occurs on data source.