arXiv subjects

Julian Shun

Publications and source records attributed to Julian Shun.

At least 55 records · Page 3Linked to original sources

arXiv2021

Theoretically-Efficient and Practical Parallel DBSCAN

The DBSCAN method for spatial clustering has received significant attention due to its applicability in a variety of data analysis tasks. There are fast sequential algorithms for DBSCAN in Euclidean space that take $O(n\log n)$ work for two dimensions, sub-quadratic work for three or more dimensions, and can be computed approximately in linear work for any constant number of dimensions. However, existing parallel DBSCAN algorithms require quadratic work in the worst case, making them inefficient for large datasets. This paper bridges the gap between theory and practice of parallel DBSCAN by presenting new parallel algorithms for Euclidean exact DBSCAN and approximate DBSCAN that match the work bounds of their sequential counterparts, and are highly parallel (polylogarithmic depth). We present implementations of our algorithms along with optimizations that improve their practical performance. We perform a comprehensive experimental evaluation of our algorithms on a variety of datasets and parameter settings. Our experiments on a 36-core machine with hyper-threading show that we outperform existing parallel DBSCAN implementations by up to several orders of magnitude, and achieve speedups by up to 33x over the best sequential algorithms.

cs.DS↗

arXiv2021

Compilation Techniques for Graph Algorithms on GPUs

The performance of graph programs depends highly on the algorithm, the size and structure of the input graphs, as well as the features of the underlying hardware. No single set of optimizations or one hardware platform works well across all settings. To achieve high performance, the programmer must carefully select which set of optimizations and hardware platforms to use. The GraphIt programming language makes it easy for the programmer to write the algorithm once and optimize it for different inputs using a scheduling language. However, GraphIt currently has no support for generating high performance code for GPUs. Programmers must resort to re-implementing the entire algorithm from scratch in a low-level language with an entirely different set of abstractions and optimizations in order to achieve high performance on GPUs. We propose GG, an extension to the GraphIt compiler framework, that achieves high performance on both CPUs and GPUs using the same algorithm specification. GG significantly expands the optimization space of GPU graph processing frameworks with a novel GPU scheduling language and compiler that enables combining graph optimizations for GPUs. GG also introduces two performance optimizations, Edge-based Thread Warps CTAs load balancing (ETWC) and EdgeBlocking, to expand the optimization space for GPUs. ETWC improves load balancing by dynamically partitioning the edges of each vertex into blocks that are assigned to threads, warps, and CTAs for execution. EdgeBlocking improves the locality of the program by reordering the edges and restricting random memory accesses to fit within the L2 cache. We evaluate GG on 5 algorithms and 9 input graphs on both Pascal and Volta generation NVIDIA GPUs, and show that it achieves up to 5.11x speedup over state-of-the-art GPU graph processing frameworks, and is the fastest on 66 out of the 90 experiments.

cs.DC↗

arXiv2020

Parallel Batch-Dynamic $k$-Clique Counting

In this paper, we study new batch-dynamic algorithms for the $k$-clique counting problem, which are dynamic algorithms where the updates are batches of edge insertions and deletions. We study this problem in the parallel setting, where the goal is to obtain algorithms with low (polylogarithmic) depth. Our first result is a new parallel batch-dynamic triangle counting algorithm with $O(Δ\sqrt{Δ+m})$ amortized work and $O(\log^* (Δ+m))$ depth with high probability, and $O(Δ+m)$ space for a batch of $Δ$ edge insertions or deletions. Our second result is an algebraic algorithm based on parallel fast matrix multiplication. Assuming that a parallel fast matrix multiplication algorithm exists with parallel matrix multiplication constant $ω_p$, the same algorithm solves dynamic $k$-clique counting with $O\left(\min\left(Δm^{\frac{(2k - 1)ω_p}{3(ω_p + 1)}}, (Δ+m)^{\frac{2(k + 1)ω_p}{3(ω_p + 1)}}\right)\right)$ amortized work and $O(\log (Δ+m))$ depth with high probability, and $O\left((Δ+m)^{\frac{2(k + 1)ω_p}{3(ω_p + 1)}}\right)$ space. Using a recently developed parallel $k$-clique counting algorithm, we also obtain a simple batch-dynamic algorithm for $k$-clique counting on graphs with arboricity $α$ running in $O(Δ(m+Δ)α^{k-4})$ expected work and $O(\log^{k-2} n)$ depth with high probability, and $O(m + Δ)$ space. Finally, we present a multicore CPU implementation of our parallel batch-dynamic triangle counting algorithm. On a 72-core machine with two-way hyper-threading, our implementation achieves 36.54--74.73x parallel speedup, and in certain cases achieves significant speedups over existing parallel algorithms for the problem, which are not theoretically-efficient.

cs.DS↗

arXiv2020

Exploring the Design Space of Static and Incremental Graph Connectivity Algorithms on GPUs

Connected components and spanning forest are fundamental graph algorithms due to their use in many important applications, such as graph clustering and image segmentation. GPUs are an ideal platform for graph algorithms due to their high peak performance and memory bandwidth. While there exist several GPU connectivity algorithms in the literature, many design choices have not yet been explored. In this paper, we explore various design choices in GPU connectivity algorithms, including sampling, linking, and tree compression, for both the static as well as the incremental setting. Our various design choices lead to over 300 new GPU implementations of connectivity, many of which outperform state-of-the-art. We present an experimental evaluation, and show that we achieve an average speedup of 2.47x speedup over existing static algorithms. In the incremental setting, we achieve a throughput of up to 48.23 billion edges per second. Compared to state-of-the-art CPU implementations on a 72-core machine, we achieve a speedup of 8.26--14.51x for static connectivity and 1.85--13.36x for incremental connectivity using a Tesla V100 GPU.

cs.DC↗

arXiv2020

Parallel Algorithms for Butterfly Computations

Butterflies are the smallest non-trivial subgraph in bipartite graphs, and therefore having efficient computations for analyzing them is crucial to improving the quality of certain applications on bipartite graphs. In this paper, we design a framework called ParButterfly that contains new parallel algorithms for the following problems on processing butterflies: global counting, per-vertex counting, per-edge counting, tip decomposition (vertex peeling), and wing decomposition (edge peeling). The main component of these algorithms is aggregating wedges incident on subsets of vertices, and our framework supports different methods for wedge aggregation, including sorting, hashing, histogramming, and batching. In addition, ParButterfly supports different ways of ranking the vertices to speed up counting, including side ordering, approximate and exact degree ordering, and approximate and exact complement coreness ordering. For counting, ParButterfly also supports both exact computation as well as approximate computation via graph sparsification. We prove strong theoretical guarantees on the work and span of the algorithms in ParButterfly. We perform a comprehensive evaluation of all of the algorithms in ParButterfly on a collection of real-world bipartite graphs using a 48-core machine. Our counting algorithms obtain significant parallel speedup, outperforming the fastest sequential algorithms by up to 13.6x with a self-relative speedup of up to 38.5x. Compared to general subgraph counting solutions, we are orders of magnitude faster. Our peeling algorithms achieve self-relative speedups of up to 10.7x and outperform the fastest sequential baseline by up to several orders of magnitude.

cs.DC↗

arXiv2020

Sage: Parallel Semi-Asymmetric Graph Algorithms for NVRAMs

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.

cs.DC↗

arXiv2020

Chiller: Contention-centric Transaction Execution and Data Partitioning for Modern Networks

Distributed transactions on high-overhead TCP/IP-based networks were conventionally considered to be prohibitively expensive and thus were avoided at all costs. To that end, the primary goal of almost any existing partitioning scheme is to minimize the number of cross-partition transactions. However, with the new generation of fast RDMA-enabled networks, this assumption is no longer valid. In fact, recent work has shown that distributed databases can scale even when the majority of transactions are cross-partition. In this paper, we first make the case that the new bottleneck which hinders truly scalable transaction processing in modern RDMA-enabled databases is data contention, and that optimizing for data contention leads to different partitioning layouts than optimizing for the number of distributed transactions. We then present Chiller, a new approach to data partitioning and transaction execution, which aims to minimize data contention for both local and distributed transactions. Finally, we evaluate Chiller using various workloads, and show that our partitioning and execution strategy outperforms traditional partitioning techniques which try to avoid distributed transactions, by up to a factor of 2.

cs.DB↗

arXiv2020

Optimizing Ordered Graph Algorithms with GraphIt

Many graph problems can be solved using ordered parallel graph algorithms that achieve significant speedup over their unordered counterparts by reducing redundant work. This paper introduces a new priority-based extension to GraphIt, a domain-specific language for writing graph applications, to simplify writing high-performance parallel ordered graph algorithms. The extension enables vertices to be processed in a dynamic order while hiding low-level implementation details from the user. We extend the compiler with new program analyses, transformations, and code generation to produce fast implementations of ordered parallel graph algorithms. We also introduce bucket fusion, a new performance optimization that fuses together different rounds of ordered algorithms to reduce synchronization overhead, resulting in $1.2\times$--3$\times$ speedup over the fastest existing ordered algorithm implementations on road networks with large diameters. With the extension, GraphIt achieves up to 3$\times$ speedup on six ordered graph algorithms over state-of-the-art frameworks and hand-optimized implementations (Julienne, Galois, and GAPBS) that support ordered algorithms.

cs.PL↗

arXiv2020

Improved Parallel Construction of Wavelet Trees and Rank/Select Structures

Existing parallel algorithms for wavelet tree construction have a work complexity of $O(n\logσ)$. This paper presents parallel algorithms for the problem with improved work complexity. Our first algorithm is based on parallel integer sorting and has either $O(n\log\log n\lceil\logσ/\sqrt{\log n\log\log n}\rceil)$ work and polylogarithmic depth, or $O(n\lceil\logσ/\sqrt{\log n}\rceil)$ work and sub-linear depth. We also describe another algorithm that has $O(n\lceil\logσ/\sqrt{\log n} \rceil)$ work and $O(σ+\log n)$ depth. We then show how to use similar ideas to construct variants of wavelet trees (arbitrary-shaped binary trees and multiary trees) as well as wavelet matrices in parallel with lower work complexity than prior algorithms. Finally, we show that the rank and select structures on binary sequences and multiary sequences, which are stored on wavelet tree nodes, can be constructed in parallel with improved work bounds, matching those of the best existing sequential algorithms for constructing rank and select structures.

cs.DC↗

arXiv2019

Theoretically Efficient Parallel Graph Algorithms Can Be Fast and Scalable

There has been significant recent interest in parallel graph processing due to the need to quickly analyze the large graphs available today. Many graph codes have been designed for distributed memory or external memory. However, today even the largest publicly-available real-world graph (the Hyperlink Web graph with over 3.5 billion vertices and 128 billion edges) can fit in the memory of a single commodity multicore server. Nevertheless, most experimental work in the literature report results on much smaller graphs, and the ones for the Hyperlink graph use distributed or external memory. Therefore, it is natural to ask whether we can efficiently solve a broad class of graph problems on this graph in memory. This paper shows that theoretically-efficient parallel graph algorithms can scale to the largest publicly-available graphs using a single machine with a terabyte of RAM, processing them in minutes. We give implementations of theoretically-efficient parallel algorithms for 20 important graph problems. We also present the optimizations and techniques that we used in our implementations, which were crucial in enabling us to process these large graphs quickly. We show that the running times of our implementations outperform existing state-of-the-art implementations on the largest real-world graphs. For many of the problems that we consider, this is the first time they have been solved on graphs at this scale. We have made the implementations developed in this work publicly-available as the Graph-Based Benchmark Suite (GBBS).

cs.DS↗

arXiv2019

Kaskade: Graph Views for Efficient Graph Analytics

Graphs are an increasingly popular way to model real-world entities and relationships between them, ranging from social networks to data lineage graphs and biological datasets. Queries over these large graphs often involve expensive subgraph traversals and complex analytical computations. These real-world graphs are often substantially more structured than a generic vertex-and-edge model would suggest, but this insight has remained mostly unexplored by existing graph engines for graph query optimization purposes. Therefore, in this work, we focus on leveraging structural properties of graphs and queries to automatically derive materialized graph views that can dramatically speed up query evaluation. We present KASKADE, the first graph query optimization framework to exploit materialized graph views for query optimization purposes. KASKADE employs a novel constraint-based view enumeration technique that mines constraints from query workloads and graph schemas, and injects them during view enumeration to significantly reduce the search space of views to be considered. Moreover, it introduces a graph view size estimator to pick the most beneficial views to materialize given a query set and to select the best query evaluation plan given a set of materialized views. We evaluate its performance over real-world graphs, including the provenance graph that we maintain at Microsoft to enable auditing, service analytics, and advanced system optimizations. Our results show that KASKADE substantially reduces the effective graph size and yields significant performance speedups (up to 50X), in some cases making otherwise intractable queries possible.

cs.DB↗

arXiv2019

Parallel Local Graph Clustering

Graph clustering has many important applications in computing, but due to growing sizes of graphs, even traditionally fast clustering methods such as spectral partitioning can be computationally expensive for real-world graphs of interest. Motivated partly by this, so-called local algorithms for graph clustering have received significant interest due to the fact that they can find good clusters in a graph with work proportional to the size of the cluster rather than that of the entire graph. This feature has proven to be crucial in making such graph clustering and many of its downstream applications efficient in practice. While local clustering algorithms are already faster than traditional algorithms that touch the entire graph, they are sequential and there is an opportunity to make them even more efficient via parallelization. In this paper, we show how to parallelize many of these algorithms in the shared-memory multicore setting, and we analyze the parallel complexity of these algorithms. We present comprehensive experiments on large-scale graphs showing that our parallel algorithms achieve good parallel speedups on a modern multicore machine, thus significantly speeding up the analysis of local graph clusters in the very large-scale setting.

cs.DC↗

arXiv2019

Low-Latency Graph Streaming Using Compressed Purely-Functional Trees

Due to the dynamic nature of real-world graphs, there has been a growing interest in the graph-streaming setting where a continuous stream of graph updates is mixed with arbitrary graph queries. In principle, purely-functional trees are an ideal choice for this setting due as they enable safe parallelism, lightweight snapshots, and strict serializability for queries. However, directly using them for graph processing would lead to significant space overhead and poor cache locality. This paper presents C-trees, a compressed purely-functional search tree data structure that significantly improves on the space usage and locality of purely-functional trees. The key idea is to use a chunking technique over trees in order to store multiple entries per tree-node. We design theoretically-efficient and practical algorithms for performing batch updates to C-trees, and also show that we can store massive dynamic real-world graphs using only a few bytes per edge, thereby achieving space usage close to that of the best static graph processing frameworks. To study the efficiency and applicability of our data structure, we designed Aspen, a graph-streaming framework that extends the interface of Ligra with operations for updating graphs. We show that Aspen is faster than two state-of-the-art graph-streaming systems, Stinger and LLAMA, while requiring less memory, and is competitive in performance with the state-of-the-art static graph frameworks, Galois, GAP, and Ligra+. With Aspen, we are able to efficiently process the largest publicly-available graph with over two hundred billion edges in the graph-streaming setting using a single commodity multicore server with 1TB of memory.

cs.DC↗

arXiv2018

GraphIt: A High-Performance DSL for Graph Analytics

The performance bottlenecks of graph applications depend not only on the algorithm and the underlying hardware, but also on the size and structure of the input graph. Programmers must try different combinations of a large set of techniques to develop the best implementation for a specific algorithm and type of graph. Existing graph frameworks lack flexibility, supporting only a limited set of optimizations. This paper introduces GraphIt, a new DSL for graph computations that generates fast implementations for algorithms with different performance characteristics running on graphs with different sizes and structures. GraphIt separates what is computed (algorithm) from how it is computed (schedule). Programmers specify the algorithm using an algorithm language, and performance optimizations are specified using a scheduling language. The algorithm language simplifies expressing the algorithms. We formulate graph optimizations, including edge traversal direction, data layout, parallelization, cache, NUMA, and kernel fusion optimizations, as tradeoffs among locality, parallelism, and work-efficiency. The scheduling language enables programmers to easily search through this complicated tradeoff space by composing together optimizations. We also built an autotuner to automatically find high-performance schedules. The compiler uses a new scheduling representation, the graph iteration space, to model, compose, and ensure the validity of the large number of optimizations. GraphIt outperforms the next fastest of six state-of-the-art shared-memory frameworks (Ligra, Green-Marl, GraphMat, Galois, Gemini, and Grazelle) on 24 out of 32 experiments by up to 4.8$\times$, and is never more than 43% slower than the fastest framework on the other experiments. GraphIt also reduces the lines of code by up to an order of magnitude compared to the next fastest framework.

cs.PL↗

arXiv2018

Parallelism in Randomized Incremental Algorithms

In this paper we show that many sequential randomized incremental algorithms are in fact parallel. We consider algorithms for several problems including Delaunay triangulation, linear programming, closest pair, smallest enclosing disk, least-element lists, and strongly connected components. We analyze the dependences between iterations in an algorithm, and show that the dependence structure is shallow with high probability, or that by violating some dependences the structure is shallow and the work is not increased significantly. We identify three types of algorithms based on their dependences and present a framework for analyzing each type. Using the framework gives work-efficient polylogarithmic-depth parallel algorithms for most of the problems that we study. This paper shows the first incremental Delaunay triangulation algorithm with optimal work and polylogarithmic depth, which is an open problem for over 30 years. This result is important since most implementations of parallel Delaunay triangulation use the incremental approach. Our results also improve bounds on strongly connected components and least-elements lists, and significantly simplify parallel algorithms for several problems.

cs.DS↗

arXiv2018

Parallel Write-Efficient Algorithms and Data Structures for Computational Geometry

In this paper, we design parallel write-efficient geometric algorithms that perform asymptotically fewer writes than standard algorithms for the same problem. This is motivated by emerging non-volatile memory technologies with read performance being close to that of random access memory but writes being significantly more expensive in terms of energy and latency. We design algorithms for planar Delaunay triangulation, $k$-d trees, and static and dynamic augmented trees. Our algorithms are designed in the recently introduced Asymmetric Nested-Parallel Model, which captures the parallel setting in which there is a small symmetric memory where reads and writes are unit cost as well as a large asymmetric memory where writes are $ω$ times more expensive than reads. In designing these algorithms, we introduce several techniques for obtaining write-efficiency, including DAG tracing, prefix doubling, reconstruction-based rebalancing and $α$-labeling, which we believe will be useful for designing other parallel write-efficient algorithms.

cs.DS↗

arXiv2018

The Parallel Persistent Memory Model

We consider a parallel computational model that consists of $P$ processors, each with a fast local ephemeral memory of limited size, and sharing a large persistent memory. The model allows for each processor to fault with bounded probability, and possibly restart. On faulting all processor state and local ephemeral memory are lost, but the persistent memory remains. This model is motivated by upcoming non-volatile memories that are as fast as existing random access memory, are accessible at the granularity of cache lines, and have the capability of surviving power outages. It is further motivated by the observation that in large parallel systems, failure of processors and their caches is not unusual. Within the model we develop a framework for developing locality efficient parallel algorithms that are resilient to failures. There are several challenges, including the need to recover from failures, the desire to do this in an asynchronous setting (i.e., not blocking other processors when one fails), and the need for synchronization primitives that are robust to failures. We describe approaches to solve these challenges based on breaking computations into what we call capsules, which have certain properties, and developing a work-stealing scheduler that functions properly within the context of failures. The scheduler guarantees a time bound of $O(W/P_A + D(P/P_A) \lceil\log_{1/f} W\rceil)$ in expectation, where $W$ and $D$ are the work and depth of the computation (in the absence of failures), $P_A$ is the average number of processors available during the computation, and $f \le 1/2$ is the probability that a capsule fails. Within the model and using the proposed methods, we develop efficient algorithms for parallel sorting and other primitives.

cs.DC↗

arXiv2017

Variational Perspective on Local Graph Clustering

Modern graph clustering applications require the analysis of large graphs and this can be computationally expensive. In this regard, local spectral graph clustering methods aim to identify well-connected clusters around a given "seed set" of reference nodes without accessing the entire graph. The celebrated Approximate Personalized PageRank (APPR) algorithm in the seminal paper by Andersen et al. is one such method. APPR was introduced and motivated purely from an algorithmic perspective. In other words, there is no a priori notion of objective function/optimality conditions that characterizes the steps taken by APPR. Here, we derive a novel variational formulation which makes explicit the actual optimization problem solved by APPR. In doing so, we draw connections between the local spectral algorithm of and an iterative shrinkage-thresholding algorithm (ISTA). In particular, we show that, appropriately initialized ISTA applied to our variational formulation can recover the sought-after local cluster in a time that only depends on the number of non-zeros of the optimal solution instead of the entire graph. In the process, we show that an optimization algorithm which apparently requires accessing the entire graph, can be made to behave in a completely local manner by accessing only a small number of nodes. This viewpoint builds a bridge across two seemingly disjoint fields of graph processing and numerical optimization, and it allows one to leverage well-studied, numerically robust, and efficient optimization algorithms for processing today's large graphs.

math.OC↗