Trevor Brown

Rosina Kharal, Trevor Brown, Justus Henneberg et al.

GPU-based concurrent data structures (CDSs) achieve high throughput for read-only queries, but efficient support for dynamic updates on fully GPU-resident data remains challenging. Ordered CDSs (e.g., B-trees and LSM-trees) maintain an index layer that directs operations to a data layer (buckets or leaves), while hash tables avoid the cost of maintaining order but do not support range or successor queries. On GPUs, maintaining and traversing an index layer under frequent updates introduces contention and warp divergence. To tackle these problems, we flip the indexing paradigm on its head with FliX, a comparison-based, flipped indexing strategy for dynamic, fully GPU-resident CDSs. Traditional GPU CDSs typically take a batch of operations and assign each operation to a GPU thread or warp. FliX, however, assigns compute (e.g., a warp) to each bucket in the data layer, and each bucket then locates operations it is responsible for in the batch. FliX can replace many index layer traversals with a single binary search on the batch, reducing redundant work and warp divergence. Further, FliX simplifies updates as no index layer must be maintained. In our experiments, FliX achieves 6.5x reduced query latency compared to a leading GPU B-tree and 1.5x compared to a leading GPU LSM-tree, while delivering 4x higher throughput per memory footprint than ordered competitors. Despite maintaining order, FliX also surpasses state-of-the-art unordered GPU hash tables in query and deletion performance, and is highly competitive in insertion performance. In update-heavy workloads, it outperforms the closest fully dynamic ordered baseline by over 8x in insertion throughput while supporting dynamic memory reclamation. These results suggest that eliminating the index layer and adopting a compute-to-bucket mapping can enable practical, fully dynamic GPU indexing without sacrificing query performance.

Gaetano Coccimiglio, Trevor Brown, Srivatsan Ravi

Software transactional memory (STM) allows programmers to easily implement concurrent data structures. STMs simplify atomicity. Recent STMs can achieve good performance for some workloads but they have some limitations. In particular, STMs typically cannot support long-running reads which access a large number of addresses that are frequently updated. Multiversioning is a common approach used to support this type of workload. However, multiversioning is often expensive and can reduce the performance of transactions where versioning is not necessary. In this work we present Multiverse, a new STM that combines the best of both unversioned TM and multiversioning. Multiverse features versioned and unversioned transactions which can execute concurrently. A main goal of Multiverse is to ensure that unversioned transactions achieve performance comparable to the state of the art unversioned STM while still supporting fast versioned transactions needed to enable long running reads. We implement Multiverse and compare it against several STMs. Our experiments demonstrate that Multiverse achieves comparable or better performance for common case workloads where there are no long running reads. For workloads with long running reads and frequent updates Multiverse significantly outperforms existing STMS. In several cases for these workloads the throughput of Multiverse is several orders of magnitude faster than other STMs.

Federico Mazzone, Trevor Brown, Florian Kerschbaum et al.

Clustering is a fundamental data processing task used for grouping records based on one or more features. In the vertically partitioned setting, data is distributed among entities, with each holding only a subset of those features. A key challenge in this scenario is that computing distances between records requires access to all distributed features, which may be privacy-sensitive and cannot be directly shared with other parties. The goal is to compute the joint clusters while preserving the privacy of each entity's dataset. Existing solutions using secret sharing or garbled circuits implement privacy-preserving variants of Lloyd's algorithm but incur high communication costs, scaling as O(nkt), where n is the number of data points, k the number of clusters, and t the number of rounds. These methods become impractical for large datasets or several parties, limiting their use to LAN settings only. On the other hand, a different line of solutions rely on differential privacy (DP) to outsource the local features of the parties to a central server. However, they often significantly degrade the utility of the clustering outcome due to excessive noise. In this work, we propose a novel solution based on homomorphic encryption and DP, reducing communication complexity to O(n+kt). In our method, parties securely outsource their features once, allowing a computing party to perform clustering operations under encryption. DP is applied only to the clusters' centroids, ensuring privacy with minimal impact on utility. Our solution clusters 100,000 two-dimensional points into five clusters using only 73MB of communication, compared to 101GB for existing works, and completes in just under 3 minutes on a 100Mbps network, whereas existing works take over 1 day. This makes our solution practical even for WAN deployments, all while maintaining accuracy comparable to plaintext k-means algorithms.