Search

Similarity search finds the most similar matches in an object collection for a given query; making it an important problem across a wide range of disciplines such as web search, image recognition and protein sequencing. Practical implementations of High Dimensional Similarity Search (HDSS) search across billions of possible solutions for multiple queries in real time, making its performance and efficiency a significant challenge. Existing clusters and datacenters use commercial multicore hardware to perform search, which may not provide the optimal performance and performance per Watt. This work explores the performance, power and cost benefits of using throughput accelerators like GPUs to perform similarity search for query cohorts even under tight deadlines. We propose optimized implementations of similarity search for both the host and the accelerator. Augmenting existing Xeon servers with accelerators results in a 3× improvement in throughput per machine, resulting in a more than 2.5× reduction in cost of ownership, even for discounted Xeon servers. Replacing a Xeon based cluster with an accelerator based cluster for similarity search reduces the total cost of ownership by more than 6× to 16× while consuming significantly less power than an ARM based cluster.

Contemporary frameworks for data analytics, such as Hadoop, Spark, and Flink seek to allow applications to scale performance flexibly by adding hardware nodes. However, we find that when the computation on each individual node is optimized, peripheral activities such as creating data partitions, messaging and synchronizing between nodes diminish the speedup obtainable from adding more hardware. We analyze workloads which distribute operations on correlated data—such as joins and aggregation found in SQL, text similarity searches, and image disparity computations. After optimizing computation on efficient, custom processors, we discover challenges in scaling the applications to hundreds of nodes on a high-bandwidth network. We then describe techniques to overcome these challenges towards prototyping a 512-node system which is able to execute SQL queries offloaded from a commercial database, and outperform SQL-on-hadoop and traditional parallel RDBMS executions by 173x and 7x respectively.