Compiler and System for Resilient Distributed Heterogeneous Graph Analytics PDF Download

Author: Gurbinder Singh Gill
Publisher:
ISBN:
Category :
Languages : en
Pages : 498

Get Book

Book Description
Graph analytics systems are used in a wide variety of applications including health care, electronic circuit design, machine learning, and cybersecurity. Graph analytics systems must handle very large graphs such as the Facebook friends graph, which has more than a billion nodes and 200 billion edges. Since machines have limited main memory, distributed-memory clusters with sufficient memory and computation power are required for processing of these graphs. In distributed graph analytics, the graph is partitioned among the machines in a cluster, and communication between partitions is implemented using a substrate like MPI. However, programming distributed-memory systems are not easy and the recent trend towards the processor heterogeneity has added to this complexity. To simplify the programming of graph applications on such platforms, this dissertation first presents a compiler called Abelian that translates shared-memory descriptions of graph algorithms written in the Galois programming model into efficient code for distributed-memory platforms with heterogeneous processors. An important runtime parameter to the compiler-generated distributed code is the partitioning policy. We present an experimental study of partitioning strategies for distributed work-efficient graph analytics applications on different CPU architecture clusters at large scale (up to 256 machines). Based on the study we present a simple rule of thumb to select among myriad policies. Another challenge of distributed graph analytics that we address in this dissertation is to deal with machine fail-stop failures, which is an important concern especially for long-running graph analytics applications on large clusters. We present a novel communication and synchronization substrate called Phoenix that leverages the algorithmic properties of graph analytics applications to recover from faults with zero overheads during fault-free execution and show that Phoenix is 24x faster than previous state-of-the-art systems. In this dissertation, we also look at the new opportunities for graph analytics on massive datasets brought by a new kind of byte-addressable memory technology with higher density and lower cost than DRAM such as intel Optane DC Persistent Memory. This enables the design of affordable systems that support up to 6TB of randomly accessible memory. In this dissertation, we present key runtime and algorithmic principles to consider when performing graph analytics on massive datasets on Optane DC Persistent Memory as well as highlight ideas that apply to graph analytics on all large-memory platforms. Finally, we show that our distributed graph analytics infrastructure can be used for a new domain of applications, in particular, embedding algorithms such as Word2Vec. Word2Vec trains the vector representations of words (also known as word embeddings) on large text corpus and resulting vector embeddings have been shown to capture semantic and syntactic relationships among words. Other examples include Node2Vec, Code2Vec, Sequence2Vec, etc (collectively known as Any2Vec) with a wide variety of uses. We formulate the training of such applications as a graph problem and present GraphAny2Vec, a distributed Any2Vec training framework that leverages the state-of-the-art distributed heterogeneous graph analytics infrastructure developed in this dissertation to scale Any2Vec training to large distributed clusters. GraphAny2Vec also demonstrates a novel way of combining model gradients during training, which allows it to scale without losing accuracy