Katana Graph’s unified intelligence platform is built from the ground up to drive innovation across a wide range of industries and use cases.
Post 3 of 3: Katana Graph Engine
Part two of the Katana Graph Intelligence Platform series described the Katana Graph Flow, which is a cloud 3.0 data flow strategy that enables the value of GraphOps on the Katana Graph Intelligence Platform. Katana Graph Flow gives complete command of the what, who, why, where and how of your data infrastructure. This post concludes our introduction to the Katana Graph Intelligence Platform. Here we will describe the Katana Graph Engine.
The Katana Graph Engine has been designed as the flexible compute core of the Katana Graph Intelligence Platform, supporting decentralized workloads for graph query, analytics, mining and paving the way for new applications providing insights previously impossible.
Katana Graph worked with the University of Texas at Austin on three of the main features of the platform: CuSP, Galois, and Gluon. Let’s break down and explain each of these key features.
CuSP is a fast streaming graph partitioner that provides users with a high level of abstraction to specify the desired partitioning policy, and quickly generates high-quality graph partitions. When dealing with large datasets on a cluster of machines the Katana Graph Engine can automatically distribute data across a compute cluster optimized on a per application basis.
Galois is a native graph computing system that provides a straightforward, parallel-computing model to permit application programmers to exploit amorphous data-parallelism in graph algorithms without having to write explicitly parallel code. This abstraction accelerates graph algorithm research, development and optimization. The Galois programming model is available in C++ and is accessible through the Katana Graph Intelligence Platform’s Python frontend.
Gluon is the scale-out, partition-aware communication substrate that enables the subset of Galois programming model to execute applications on graph data partitioned through CuSP across heterogeneous clusters (exploiting CPU, GPU, XPU and memory architectures) through a lightweight API.
The Katana Graph Engine uses Automatic Virtualization of Accelerators (AVA) to create virtual accelerator stacks that are compatible with disaggregated and heterogeneous accelerators. This configuration provides all the performance benefits without relying on pass-through techniques or dedicated hardware.
The Katana’s Graphs Neural Network (GNN) Engine provides access to natively optimized distributed graph AI models like GraphSage GCN, and GAT as well as programming abstractions compatible with DGL and Pytorch Geometric to utilize the latest deep learning models for heterogeneous graphs.
Data is at the heart of enterprise success, and the Katana Graph Intelligence Platform makes data processing flexible, responsive and seamless. Katana Graph understands change, and is dedicated to supporting the needs of its customers to extract the most from their data.
Further Reading
https://www.cs.utexas.edu/~roshan/CuSP.pdf
https://dl.acm.org/doi/10.1145/3192366.3192404
https://www.cs.utexas.edu/~roshan/Gluon.pdf
