Reducing Query Latencies in Web Search Using Fine-Grained Parallelism

Semantic Web search is a new application of recent advances in information retrieval (IR), natural language processing, artificial intelligence, and other fields. The Powerset group in Microsoft develops a semantic search engine that aims to answer queries not only by matching keywords, but by actually matching meaning in queries to meaning in Web documents. Compared to typical keyword search, semantic search can pose additional engineering challenges for the back-end and infrastructure designs. Of these, the main challenge addressed in this paper is how to lower query latencies to acceptable, interactive levels. Index-based semantic search requires more data processing, such as numerous synonyms, hypernyms, multiple linguistic readings, and other semantic information, both on queries and in the index. In addition, some of the algorithms can be super-linear, such as matching co-references across a document. Consequently, many semantic queries can run significantly slower than the same keyword query. Users, however, have grown to expect Web search engines to provide near-instantaneous results, and a slow search engine could be deemed unusable even if it provides highly relevant results. It is therefore imperative for any search engine to meet its users’ interactivity expectations, or risk losing them. Our approach to tackle this challenge is to exploit data parallelism in slow search queries to reduce their latency in multi-core systems. Although all search engines are designed to exploit parallelism, at the single-node level this usually translates to throughput-oriented task parallelism. This paper focuses on the engineering of two latency-oriented approaches (coarse- and fine-grained) and compares them to the task-parallel approach. We use Powerset’s deployed search engine to evaluate the various factors that affect parallel performance: workload, overhead, load balancing, and resource contention. We also discuss heuristics to selectively control the degree of parallelism and consequent overhead on a query-by-query level. Our experimental results show that using fine-grained parallelism with these dynamic heuristics can significantly reduce query latencies compared to fixed, coarse-granularity parallelization schemes. Although these results were obtained on, and optimized for, Powerset’s semantic search, they can be readily generalized to a wide class of inverted-index search engines.