Thread cooperation in multicore architectures for frequency counting over multiple data streams

Abstract

Many real-world data stream analysis applications such as network monitoring, click stream analysis, and others require combining multiple streams of data arriving from multiple sources. This is referred to as multi-stream analysis. To deal with high stream arrival rates, it is desirable that such systems be capable of supporting very high processing throughput. The advent of multicore processors and powerful servers driven by these processors calls for efficient parallel designs that can effectively utilize the parallelism of the multicores, since performance improvement is possible only through effective parallelism. In this paper, we address the problem of parallelizing multi-stream analysis in the context of multicore processors. Specifically, we concentrate on parallelizing frequent elements, top-k, and frequency counting over multiple streams. We discuss the challenges in designing an efficient parallel system for multi-stream processing. Our evaluation and analysis reveals that traditional "contention" based locking results in excessive overhead and wait, which in turn leads to severe performance degradation in modern multicore architectures. Based on our analysis, we propose a "cooperation" based locking paradigm for efficient parallelization of frequency counting. The proposed "cooperation" based paradigm removes waits associated with synchronization, and allows replacing locks by much cheaper atomic synchronization primitives. Our implementation of the proposed paradigm to parallelize a well known frequency counting algorithm shows the benefits of the proposed "cooperation" based locking paradigm when compared to the traditional "contention" based locking paradigm. In our experiments, the proposed "cooperation" based design outperforms the traditional "contention" based design by a factor of 2 − 5.5X for synthetic zipfian data sets.

Count Me In: Viable Distributed Summary Statistics for Securing High-Speed Networks

“…We thus likewise choose a probabilistic version instead: Metwally et al's algorithm [17], which in addition also provides estimates on the number of times specific elements were seen. Just as HLL, the algorithm satisfies all our constraints, including composability (see [5]). …”

Scalable and dynamically balanced shared-everything OLTP with physiological partitioning

“…Where speculative lock inheritance allows the system to spread lock operations across multiple transactions to reduce contention, data-oriented systems replace the central lock manager with thread-local lock management. Reducing lock contention with data-oriented execution is also studied for data-streams' operators [12] by making threads delegate the work on some data to the thread that already holds the lock for that data and move to the next operation in their queues.…”

Parallelizing weighted frequency counting in high-speed network monitoring

“…For example, both Bandi et al [17] and Das et al [18] take advantage of the Content Addressable Memories (CAMs) to support the constant time lookup operation in hardware for finding frequent items in unweighted data streams. More closely related to our problem is the work of Das et al [12], which proposes a ''cooperation" based locking paradigm for parallelizing frequency counting in multiple unweighted data streams in the multi-core architecture. The major disadvantage is lacking scalability, as it only outperforms the ''contention" based design by a factor of 2-5.5X.…”

“…However, the frequency counting problem is not 0140-3664/$ -see front matter Ó 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.comcom.2010.04.026 ''embarrassingly parallel" as a result of the data dependencies and shared data structures in the stream processing [12]. Therefore, thoughtful and efficient designs are needed to parallelize the weighted frequency counting algorithms to improve the overall throughput in the multi-core architecture.…”