arXiv subjects

Tahsin Reza

Publications and source records attributed to Tahsin Reza.

5 recordsLinked to original sources

arXiv2026

Scaling Vision Transformers: Evaluating DeepSpeed for Image-Centric Workloads

Vision Transformers (ViTs) have demonstrated remarkable potential in image processing tasks by utilizing self-attention mechanisms to capture global relationships within data. However, their scalability is hindered by significant computational and memory demands, especially for large-scale models with many parameters. This study aims to leverage DeepSpeed, a highly efficient distributed training framework that is commonly used for language models, to enhance the scalability and performance of ViTs. We evaluate intra- and inter-node training efficiency across multiple GPU configurations on various datasets like CIFAR-10 and CIFAR-100, exploring the impact of distributed data parallelism on training speed, communication overhead, and overall scalability (strong and weak scaling). By systematically varying software parameters, such as batch size and gradient accumulation, we identify key factors influencing performance of distributed training. The experiments in this study provide a foundational basis for applying DeepSpeed to image-related tasks. Future work will extend these investigations to deepen our understanding of DeepSpeed's limitations and explore strategies for optimizing distributed training pipelines for Vision Transformers.

cs.LG↗

arXiv2022

Towards Distributed 2-Approximation Steiner Minimal Trees in Billion-edge Graphs

Given an edge-weighted graph and a set of known seed vertices, a network scientist often desires to understand the graph relationships to explain connections between the seed vertices. When the seed set is 3 or larger Steiner minimal tree - min-weight acyclic connected subgraph (of the input graph) that contains all the seed vertices - is an attractive generalization of shortest weighted paths. In general, computing a Steiner minimal tree is NP-hard, but several polynomial-time algorithms have been designed and proven to yield Steiner trees whose total weight is bounded within 2 times the Steiner minimal tree. In this paper, we present a parallel 2-approximation Steiner minimal tree algorithm and its MPI-based distributed implementation. In place of distance computation between all pairs of seed vertices, an expensive phase in many algorithms, our solution exploits Voronoi cell computation. Also, this approach has higher parallel efficiency than others that involve minimum spanning tree computation on the entire graph. Furthermore, our distributed design exploits asynchronous processing and a message prioritization scheme to accelerate convergence of distance computation, and harnesses both vertex and edge centric processing to offer fast time-to-solution. We demonstrate scalability and performance of our solution using real-world graphs with up to 128 billion edges and 512 compute nodes (8K processes). We compare our solution with the state-of-the-art exact Steiner minimal tree solver, SCIP-Jack, and two serial algorithms. Our solution comfortably outperforms these related works on graphs with 10s million edges and offers decent strong scaling - up to 90% efficient. We empirically show that, on average, the total distance of the Steiner tree identified by our solution is 1.0527 times greater than the Steiner minimal tree - well within the theoretical bound of less than equal to 2.

cs.DC↗

arXiv2021

TriPoll: Computing Surveys of Triangles in Massive-Scale Temporal Graphs with Metadata

Understanding the higher-order interactions within network data is a key objective of network science. Surveys of metadata triangles (or patterned 3-cycles in metadata-enriched graphs) are often of interest in this pursuit. In this work, we develop TriPoll, a prototype distributed HPC system capable of surveying triangles in massive graphs containing metadata on their edges and vertices. We contrast our approach with much of the prior effort on triangle analysis, which often focuses on simple triangle counting, usually in simple graphs with no metadata. We assess the scalability of TriPoll when surveying triangles involving metadata on real and synthetic graphs with up to hundreds of billions of edges.We utilize communication-reducing optimizations to demonstrate a triangle counting task on a 224 billion edge web graph in approximately half of the time of competing approaches, while additionally supporting metadata-aware capabilities.

cs.DC↗

arXiv2020

Scalable Pattern Matching in Metadata Graphs via Constraint Checking

Pattern matching is a fundamental tool for answering complex graph queries. Unfortunately, existing solutions have limited capabilities: they do not scale to process large graphs and/or support only a restricted set of search templates or usage scenarios. We present an algorithmic pipeline that bases pattern matching on constraint checking. The key intuition is that each vertex or edge participating in a match has to meet a set of constrains implicitly specified by the search template. The pipeline we propose, generates these constraints and iterates over them to eliminate all the vertices and edges that do not participate in any match, and reduces the background graph to a subgraph which is the union of all matches. Additional analysis can be performed on this annotated, reduced graph, such as full match enumeration. Furthermore, a vertex-centric formulation for constraint checking algorithms exists, and this makes it possible to harness existing high-performance, vertex-centric graph processing frameworks. The key contribution of this work is a design following the constraint checking approach for exact matching and its experimental evaluation. We show that the proposed technique: (i) enables highly scalable pattern matching in labeled graphs, (ii) supports arbitrary patterns with 100% precision, (iii) always selects all vertices and edges that participate in matches, thus offering 100% recall, and (iv) supports a set of popular data analysis scenarios. We implement our approach on top of HavoqGT, an open-source asynchronous graph processing framework, and demonstrate its advantages through strong and weak scaling experiments on massive scale real-world (up to 257 billion edges) and synthetic (up to 4.4 trillion edges) labeled graphs respectively, and at scales (1,024 nodes / 36,864 cores), orders of magnitude larger than used in the past for similar problems.

cs.DC↗

arXiv2014

Efficient Large-Scale Graph Processing on Hybrid CPU and GPU Systems

The increasing scale and wealth of inter-connected data, such as those accrued by social network applications, demand the design of new techniques and platforms to efficiently derive actionable knowledge from large-scale graphs. However, real-world graphs are famously difficult to process efficiently. Not only they have a large memory footprint, but also most graph algorithms entail memory access patterns with poor locality, data-dependent parallelism and a low compute-to-memory access ratio. Moreover, most real-world graphs have a highly heterogeneous node degree distribution, hence partitioning these graphs for parallel processing and simultaneously achieving access locality and load-balancing is difficult. This work starts from the hypothesis that hybrid platforms (e.g., GPU-accelerated systems) have both the potential to cope with the heterogeneous structure of real graphs and to offer a cost-effective platform for high-performance graph processing. This work assesses this hypothesis and presents an extensive exploration of the opportunity to harness hybrid systems to process large-scale graphs efficiently. In particular, (i) we present a performance model that estimates the achievable performance on hybrid platforms; (ii) informed by the performance model, we design and develop TOTEM - a processing engine that provides a convenient environment to implement graph algorithms on hybrid platforms; (iii) we show that further performance gains can be extracted using partitioning strategies that aim to produce partitions that each matches the strengths of the processing element it is allocated to, finally, (iv) we demonstrate the performance advantages of the hybrid system through a comprehensive evaluation that uses real and synthetic workloads (as large as 16 billion edges), multiple graph algorithms that stress the system in various ways, and a variety of hardware configurations.

cs.DC↗