We consider the problem of maintaining a hierarchical agglomerative clustering (HAC) in the dynamic setting, when the input is subject to point insertions and deletions. We introduce DynHac - the first dynamic HAC algorithm for the popular average-linkage version of the problem which can maintain a 1 + ε approximate solution. Our approach leverages recent structural results on 1 + ε-approximate HAC [1] to carefully identify the part of the clustering dendrogram that needs to be updated in order to produce a solution that is consistent with what a full recomputation from scratch would have output.

We consider the fundamental problem of decomposing a large-scale approximate nearest neighbor search (ANNS) problem into smaller sub-problems. The goal is to partition the input points into neighborhood-preserving shards , so that the nearest neighbors of any point are contained in only a few shards. When a query arrives, a routing algorithm is used to identify the shards which should be searched for its nearest neighbors. This approach forms the backbone of distributed ANNS, where the dataset is so large that it must be split across multiple machines. In this paper, we design simple and highly efficient routing methods based on clustering and locality-sensitive hashing. We prove strong theoretical guarantees for the LSH-based method, whereas the clustering-based method exhibits better empirical performance. A crucial characteristic of our routing algorithms is that they are inherently modular, and can be used with any partitioning method. This addresses a key drawback of prior approaches, where the routing algorithms are inextricably linked to their associated partitioning method. In particular, due to their modular structure, our routing methods enable the use of balanced graph partitioning , which is a high-quality partitioning method without a naturally associated routing algorithm. Prior routing methods compatible with graph partitioning are too slow to train on large-scale data. We provide the first routing methods that are simultaneously compatible with graph partitioning, fast to train, admit low latency, and achieve high recall. In a comprehensive evaluation of our partitioning and routing on billion-scale datasets, we show that our methods outperform existing scalable partitioning methods by significant margins, achieving up to 1.72× higher QPS at 90% 10-recall than the best competitor and 1.27× in the geometric mean. Through fast and modular routing we establish graph partitioning as the new method of choice for partitioning large-scale ANNS datasets.

Abstract The performance of today’s in-memory indexes is bottlenecked by the memory latency/bandwidth wall. Processing-in-memory (PIM) is an emerging approach that potentially mitigates this bottleneck by enabling low-latency memory access whose aggregate memory bandwidth scales with the number of PIM nodes. There is an inherent tension, however, between minimizing inter-node communication and achieving load balance in PIM systems, in the presence of workload skew. This paper presents PIM-tree , an ordered index for PIM systems that achieves both low communication and high load balance, regardless of the degree of skew in data/queries. Our skew-resistant index is based on a novel division of labor between the multi-core host CPU and the PIM nodes, which leverages the strengths of each. We introduce push-pull search , which dynamically decides whether to push queries to a PIM-tree node (CPU $$\rightarrow $$ → PIM-node) or pull the node’s keys back to the CPU (PIM-node $$\rightarrow $$ → CPU) based on workload skew. Combined with other PIM-friendly optimizations ( shadow subtrees and chunking ), PIM-tree achieves high throughput, (guaranteed) low communication, and (guaranteed) high load balance, for batches of point queries, updates, and range scans. We implement the PIM-tree structure, in addition to prior proposed PIM indexes, on the latest PIM system from UPMEM, with 32 CPU cores and 2048 PIM nodes. On workloads with 500 million keys and batches of 1 million queries, the throughput using PIM-trees is up to $$69.7\times $$ 69.7 × and $$59.1\times $$ 59.1 × higher than the two best prior PIM-based methods. As far as we know these are the first implementations of ordered indexes on real PIM systems.

Nucleus decompositions have been shown to be a useful tool for finding dense subgraphs. The coreness value of a clique represents its density based on the number of other cliques it is adjacent to. One useful output of nucleus decomposition is to generate a hierarchy among dense subgraphs at different resolutions. However, existing parallel algorithms for nucleus decomposition do not generate this hierarchy, and only compute the coreness values. This paper presents a scalable parallel algorithm for hierarchy construction, with practical optimizations, such as interleaving the coreness computation with hierarchy construction and using a concurrent union-find data structure in an innovative way to generate the hierarchy. We also introduce a parallel approximation algorithm for nucleus decomposition, which achieves much lower span in theory and better performance in practice. We prove strong theoretical bounds on the work and span (parallel time) of our algorithms. On a 30-core machine with two-way hyper-threading, our parallel hierarchy construction algorithm achieves up to a 58.84x speedup over the state-of-the-art sequential hierarchy construction algorithm by Sariyuce et al. and up to a 30.96x self-relative parallel speedup. On the same machine, our approximation algorithm achieves a 3.3x speedup over our exact algorithm, while generating coreness estimates with a multiplicative error of 1.33x on average.

We study the problem of dynamically maintaining the connected components of an undirected graph subject to edge insertions and deletions. We give the first parallel algorithm for the problem that is work-efficient, supports batches of updates, runs in polylogarithmic depth, and uses only linear total space. The existing algorithms for the problem either use super-linear space, do not come with strong theoretical bounds, or are not parallel. On the empirical side, we provide the first implementation of the cluster forest algorithm , the first linear-space and polylogarithmic update time algorithm for dynamic connectivity. Experimentally, we find that our algorithm uses up to 19.7× less space and is up to 6.2× faster than the level-set algorithm of Holm, de Lichten-berg, and Thorup, arguably the most widely-implemented dynamic connectivity algorithm with strong theoretical guarantees.

A fundamental building block in any graph algorithm is a graph container - - a data structure used to represent the graph. Ideally, a graph container enables efficient access to the underlying graph, has low space usage, and supports updating the graph efficiently. In this paper, we conduct an extensive empirical evaluation of graph containers designed to support running algorithms on large graphs. To our knowledge, this is the first apples-to-apples comparison of graph containers rather than overall systems, which include confounding factors such as differences in algorithm implementations and infrastructure. We measure the running time of 10 highly-optimized algorithms across over 20 different containers and 10 graphs. Somewhat surprisingly, we find that the average algorithm running time does not differ much across containers, especially those that support dynamic updates. Specifically, a simple container based on an off-the-shelf B-tree is only 1.22× slower on average than a highly optimized static one. Moreover, we observe that simplifying a graph-container Application Programming Interface (API) to only a few simple functions incurs a mere 1.16× slowdown compared to a complete API. Finally, we also measure batch-insert throughput in dynamic-graph containers for a full picture of their performance. To perform the benchmarks, we introduce BYO, a unified framework that standardizes evaluations of graph-algorithm performance across different graph containers. BYO extends the Graph Based Benchmark Suite (Dhulipala et al. 18), a state-of-the-art graph algorithm benchmark, to easily plug into different dynamic graph containers and enable fair comparisons between them on a large suite of graph algorithms. While several graph algorithm benchmarks have been developed to date, to the best of our knowledge, BYO is the first system designed to benchmark graph containers.

We introduce the ParClusterers Benchmark Suite (PCBS)---a collection of highly scalable parallel graph clustering algorithms and benchmarking tools that streamline comparing different graph clustering algorithms and implementations. The benchmark includes clustering algorithms that target a wide range of modern clustering use cases, including community detection, classification, and dense subgraph mining. The benchmark toolkit makes it easy to run and evaluate multiple instances of different clustering algorithms with respect to both the running time and quality. We evaluate the PCBS algorithms empirically and find that they deliver both the state of the art quality and the running time. In terms of the running time, they are on average over 4x faster than the fastest library we compared to. In terms of quality, the correlation clustering algorithm [Shi et al., VLDB'21] optimizing for the LambdaCC objective, which does not have a direct counterpart in other libraries, delivers the highest quality in the majority of datasets that we used.

We introduce TeraHAC, a (1+ε)-approximate hierarchical agglomerative clustering (HAC) algorithm which scales to trillion-edge graphs. Our algorithm is based on a new approach to computing (1+ε)-approximate HAC, which is a novel combination of the nearest-neighbor chain algorithm and the notion of (1+ε)-approximate HAC. Our approach allows us to partition the graph among multiple machines and make significant progress in computing the clustering within each partition before any communication with other partitions is needed. We evaluate TeraHAC on a number of real-world and synthetic graphs of up to 8 trillion edges. We show that TeraHAC requires over 100x fewer rounds compared to previously known approaches for computing HAC. It is up to 8.3x faster than SCC, the state-of-the-art distributed algorithm for hierarchical clustering, while achieving 1.16x higher quality. In fact, TeraHAC essentially retains the quality of the celebrated HAC algorithm while significantly improving the running time.

We proposeLightNE 2.0, a cost-effective, scalable, automated, and high-quality network embedding system that scales to graphs with hundreds of billions of edges on a single machine. In contrast to the mainstream belief that distributed architecture and GPUs are needed for large-scale network embedding with good quality, we prove that we can achieve higher quality, better scalability, lower cost, and faster runtime with shared-memory, CPU-only architecture.LightNE 2.0combines two theoretically grounded embedding methods NetSMF and ProNE. We introduce the following techniques to network embedding for the first time: (1) a newly proposed downsampling method to reduce the sample complexity of NetSMF while preserving its theoretical advantages; (2) a high-performance parallel graph processing stack GBBS to achieve high memory efficiency and scalability; (3) sparse parallel hash table to aggregate and maintain the matrix sparsifier in memory; (4) a fast randomized singular value decomposition (SVD) enhanced by power iteration and fast orthonormalization to improve vanilla randomized SVD in terms of both efficiency and effectiveness; (5) Intel MKL for proposed fast randomized SVD and spectral propagation; and (6) a fast and lightweight AutoML library FLAML for automated hyperparameter tuning. Experimental results show thatLightNE 2.0can be up to 84× faster than GraphVite, 30× faster than PBG and 9× faster than NetSMF while delivering better performance.LightNE 2.0can embed very large graph with 1.7 billion nodes and 124 billion edges in half an hour on a CPU server, while other baselines cannot handle very large graphs of this scale.

The performance of today's in-memory indexes is bottlenecked by the memory latency/bandwidth wall. Processing-in-memory (PIM) is an emerging approach that potentially mitigates this bottleneck, by enabling low-latency memory access whose aggregate memory bandwidth scales with the number of PIM nodes. There is an inherent tension, however, between minimizing inter-node communication and achieving load balance in PIM systems, in the presence of workload skew. This paper presents PIM-tree , an ordered index for PIM systems that achieves both low communication and high load balance, regardless of the degree of skew in data and queries. Our skew-resistant index is based on a novel division of labor between the host CPU and PIM nodes, which leverages the strengths of each. We introduce push-pull search , which dynamically decides whether to push queries to a PIM-tree node or pull the node's keys back to the CPU based on workload skew. Combined with other PIM-friendly optimizations ( shadow subtrees and chunked skip lists ), our PIM-tree provides high-throughput, (guaranteed) low communication, and (guaranteed) high load balance, for batches of point queries, updates, and range scans. We implement PIM-tree, in addition to prior proposed PIM indexes, on the latest PIM system from UPMEM, with 32 CPU cores and 2048 PIM nodes. On workloads with 500 million keys and batches of 1 million queries, the throughput using PIM-trees is up to 69.7X and 59.1x higher than the two best prior PIM-based methods. As far as we know these are the first implementations of an ordered index on a real PIM system.

We propose LightNE, a cost-effective, scalable, and high quality network embedding system that scales to graphs with hundreds of billions of edges on a single machine. In contrast to the mainstream belief that distributed architecture and GPUs are needed for large-scale network embedding with good quality, we prove that we can achieve higher quality, better scalability, lower cost and faster runtime with shared-memory, CPU-only architecture. LightNE combines two theoretically grounded embedding methods NetSMF and ProNE. We introduce the following techniques to network embedding for the first time: (1) a newly proposed downsampling method to reduce the sample complexity of NetSMF while preserving its theoretical advantages; (2) a high-performance parallel graph processing stack GBBS to achieve high memory efficiency and scalability; (3) sparse parallel hash table to aggregate and maintain the matrix sparsifier in memory; and (4) Intel MKL for efficient randomized SVD and spectral propagation.

SCAN (Structural Clustering Algorithm for Networks) is a well-studied, widely used graph clustering algorithm. For large graphs, however, sequential SCAN variants are prohibitively slow, and parallel SCAN variants do not effectively share work among queries with different SCAN parameter settings. Since users of SCAN often explore many parameter settings to find good clusterings, it is worthwhile to precompute an index that speeds up queries.

Graph clustering and community detection are central problems in modern data mining. The increasing need for analyzing billion-scale data calls for faster and more scalable algorithms for these problems. There are certain trade-offs between the quality and speed of such clustering algorithms. In this paper, we design scalable algorithms that achieve high quality when evaluated based on ground truth. We develop a generalized sequential and shared-memory parallel framework based on the LAMBDACC objective (introduced by Veldt et al.), which encompasses modularity and correlation clustering. Our framework consists of highly-optimized implementations that scale to large data sets of billions of edges and that obtain high-quality clusters compared to ground-truth data, on both unweighted and weighted graphs. Our empirical evaluation shows that this framework improves the state-of-the-art trade-offs between speed and quality of scalable community detection. For example, on a 30-core machine with two-way hyper-threading, our implementations achieve orders of magnitude speedups over other correlation clustering baselines, and up to 28.44× speedups over our own sequential baselines while maintaining or improving quality.

This paper studies the nucleus decomposition problem, which has been shown to be useful in finding dense substructures in graphs. We present a novel parallel algorithm that is efficient both in theory and in practice. Our algorithm achieves a work complexity matching the best sequential algorithm while also having low depth (parallel running time), which significantly improves upon the only existing parallel nucleus decomposition algorithm (Sariyüce et al. , PVLDB 2018). The key to the theoretical efficiency of our algorithm is a new lemma that bounds the amount of work done when peeling cliques from the graph, combined with the use of a theoretically-efficient parallel algorithms for clique listing and bucketing. We introduce several new practical optimizations, including a new multi-level hash table structure to store information on cliques space-efficiently and a technique for traversing this structure cache-efficiently. On a 30-core machine with two-way hyper-threading on real-world graphs, we achieve up to a 55x speedup over the state-of-the-art parallel nucleus decomposition algorithm by Sariyüce et al. , and up to a 40x self-relative parallel speedup. We are able to efficiently compute larger nucleus decompositions than prior work on several million-scale graphs for the first time.

This paper studies the hierarchical clustering problem, where the goal is to produce a dendrogram that represents clusters at varying scales of a data set. We propose the ParChain framework for designing parallel hierarchical agglomerative clustering (HAC) algorithms, and using the framework we obtain novel parallel algorithms for the complete linkage, average linkage, and Ward's linkage criteria. Compared to most previous parallel HAC algorithms, which require quadratic memory, our new algorithms require only linear memory, and are scalable to large data sets. ParChain is based on our parallelization of the nearest-neighbor chain algorithm, and enables multiple clusters to be merged on every round. We introduce two key optimizations that are critical for efficiency: a range query optimization that reduces the number of distance computations required when finding nearest neighbors of clusters, and a caching optimization that stores a subset of previously computed distances, which are likely to be reused. Experimentally, we show that our highly-optimized implementations using 48 cores with two-way hyper-threading achieve 5.8--110.1x speedup over state-of-the-art parallel HAC algorithms and achieve 13.75--54.23x self-relative speedup. Compared to state-of-the-art algorithms, our algorithms require up to 237.3x less space. Our algorithms are able to scale to data set sizes with tens of millions of points, which existing algorithms are not able to handle.

We study fundamental graph problems such as graph connectivity, minimum spanning forest (MSF), and approximate maximum (weight) matching in a distributed setting. In particular, we focus on the Adaptive Massively Parallel Computation (AMPC) model, which is a theoretical model that captures MapReduce-like computation augmented with a distributed hash table. We show the first AMPC algorithms for all of the studied problems that run in a constant number of rounds and use only O ( n ϵ ) space per machine, where 0 < ϵ < 1. Our results improve both upon the previous results in the AMPC model, as well as the best-known results in the MPC model, which is the theoretical model underpinning many popular distributed computation frameworks, such as MapReduce, Hadoop, Beam, Pregel and Giraph. Finally, we provide an empirical comparison of the algorithms in the MPC and AMPC models in a fault-tolerant distributed computation environment. We empirically evaluate our algorithms on a set of large real-world graphs and show that our AMPC algorithms can achieve improvements in both running time and round-complexity over optimized MPC baselines.

Connected components is a fundamental kernel in graph applications. The fastest existing multicore algorithms for solving graph connectivity are based on some form of edge sampling and/or linking and compressing trees. However, many combinations of these design choices have been left unexplored. In this paper, we design the ConnectIt framework, which provides different sampling strategies as well as various tree linking and compression schemes. ConnectIt enables us to obtain several hundred new variants of connectivity algorithms, most of which extend to computing spanning forest. In addition to static graphs, we also extend ConnectIt to support mixes of insertions and connectivity queries in the concurrent setting. We present an experimental evaluation of ConnectIt on a 72-core machine, which we believe is the most comprehensive evaluation of parallel connectivity algorithms to date. Compared to a collection of state-of-the-art static multicore algorithms, we obtain an average speedup of 12.4x (2.36x average speedup over the fastest existing implementation for each graph). Using ConnectIt, we are able to compute connectivity on the largest publicly-available graph (with over 3.5 billion vertices and 128 billion edges) in under 10 seconds using a 72-core machine, providing a 3.1x speedup over the fastest existing connectivity result for this graph, in any computational setting. For our incremental algorithms, we show that our algorithms can ingest graph updates at up to several billion edges per second. To guide the user in selecting the best variants in ConnectIt for different situations, we provide a detailed analysis of the different strategies. Finally, we show how the techniques in ConnectIt can be used to speed up two important graph applications: approximate minimum spanning forest and SCAN clustering.

Non-volatile main memory (NVRAM) technologies provide an attractive set of features for large-scale graph analytics, including byte-addressability, low idle power, and improved memory-density. NVRAM systems today have an order of magnitude more NVRAM than traditional memory (DRAM). NVRAM systems could therefore potentially allow very large graph problems to be solved on a single machine, at a modest cost. However, a significant challenge in achieving high performance is in accounting for the fact that NVRAM writes can be much more expensive than NVRAM reads. In this paper, we propose an approach to parallel graph analytics using the Parallel Semi-Asymmetric Model (PSAM) , in which the graph is stored as a read-only data structure (in NVRAM), and the amount of mutable memory is kept proportional to the number of vertices. Similar to the popular semi-external and semi-streaming models for graph analytics, the PSAM approach assumes that the vertices of the graph fit in a fast read-write memory (DRAM), but the edges do not. In NVRAM systems, our approach eliminates writes to the NVRAM, among other benefits. To experimentally study this new setting, we develop Sage , a parallel semi-asymmetric graph engine with which we implement provably-efficient (and often work-optimal) PSAM algorithms for over a dozen fundamental graph problems. We experimentally study Sage using a 48--core machine on the largest publicly-available real-world graph (the Hyperlink Web graph with over 3.5 billion vertices and 128 billion edges) equipped with Optane DC Persistent Memory, and show that Sage outperforms the fastest prior systems designed for NVRAM. Importantly, we also show that Sage nearly matches the fastest prior systems running solely in DRAM, by effectively hiding the costs of repeatedly accessing NVRAM versus DRAM.

Graph reordering is a powerful technique to increase the locality of the representations of graphs, which can be helpful in several applications. We study how the technique can be used to improve compression of graphs and inverted indexes.

We study compression techniques for parallel in-memory graph algorithms, and show that we can achieve reduced space usage while obtaining competitive or improved performance compared to running the algorithms on uncompressed graphs. We integrate the compression techniques into Ligra, a recent shared-memory graph processing system. This system, which we call Ligra+, is able to represent graphs using about half of the space for the uncompressed graphs on average. Furthermore, Ligra+ is slightly faster than Ligra on average on a 40-core machine with hyper-threading. Our experimental study shows that Ligra+ is able to process graphs using less memory, while performing as well as or faster than Ligra.