Abstract
The NTT laboratories have been focusing on expanding the use and development of open source software (OSS), primarily through the NTT Software Innovation Center (NTT SIC). NTT's active contributions to OSS communities such as OpenStack and Apache Hadoop have yielded productive results with refinement of the company's own technologies and training of personnel. Meanwhile, the wave of OSS is steadily spreading to new fields, including virtualization and artificial intelligence. NTT SIC is also engaged in these efforts, and its contributions are beginning to bear fruit. In this article, NTT SIC's work in OSS is introduced from a technical perspective and its community activities are also described.
Introduction
As the Internet of Things (IoT) and Artificial Intelligence (AI) transition from conceptual frameworks to real-world applications, they encounter significant infrastructure hurdles. Traditional virtualization methods, primarily based on hypervisors, are often too resource-intensive and slow to accommodate the rapid scaling required by modern microservices. Furthermore, the computational demands of training deep learning models have escalated, often requiring days or weeks of processing time. The NTT Software Innovation Center (NTT SIC) addresses these challenges by leveraging and contributing to Open Source Software (OSS) to create more agile and performant systems.
Container-Based Virtualization and Infrastructure
A core pillar of NTT's strategy is the shift from traditional virtual machines (VMs) to container-based virtualization. Unlike VMs, which require a full guest operating system for every instance, containers share the host's OS kernel, significantly reducing overhead and improving deployment speed.
NTT focuses on optimizing Docker for efficient image generation and Kubernetes for large-scale container orchestration. Beyond simple usage, NTT engineers are active maintainers of etcd, the distributed key-value store that serves as the "brain" of Kubernetes. Their contributions focus on enhancing the Raft consensus protocol and improving authentication mechanisms to ensure system reliability.
Streamlining the Container Lifecycle
To manage the complexity of thousands of containers, NTT utilizes a standardized workflow for image definition and deployment. This involves using Dockerfiles to define environment specifications, which are then pushed to registries like Docker Hub before being composed and deployed to cloud environments.
To ensure the robustness of these distributed systems, NTT developed Namazu, a specialized testing tool designed to identify bugs in complex OSS projects like etcd by simulating various failure scenarios.
Accelerating Deep Learning through RDMA
In the realm of AI, the primary bottleneck has shifted from raw computation to communication. When training models across multiple GPUs and servers using "data parallelism," the time spent synchronizing learning parameters between nodes can stall the entire process.
To overcome this, NTT is implementing Remote Direct Memory Access (RDMA). By bypassing the standard TCP/IP stack and the CPU, RDMA allows one server to read or write directly to another's memory. This significantly lowers latency and allows researchers to approach the theoretical maximum data transfer speeds of high-speed interconnects like InfiniBand. NTT has successfully integrated RDMA support into frameworks like TensorFlow and is working on similar optimizations for ChainerMN, a distributed version of the Japanese-developed Chainer framework.
Key Insights and Community Impact
NTT emphasizes that true leadership in the OSS space requires more than just code contribution; it requires holding "maintainer" roles. At the time of this work, NTT SIC provided the only Docker/Moby maintainer from Japan and one of only seven global etcd maintainers. This influence allows NTT to steer the technical direction of global infrastructure projects.
In terms of performance, these optimizations have yielded tangible results. For instance, ChainerMN demonstrated speeds up to six times faster than TensorFlow in specific distributed learning benchmarks during early 2017.
Future Directions
While RDMA offers massive performance gains, its implementation is complex. NTT is currently working on abstracting these APIs to make high-speed communication more accessible to developers. Additionally, as no single deep learning framework has become the definitive standard, NTT continues to evaluate various architectures—including TensorFlow, MXNet, and PyTorch—to determine the optimal solutions for diverse real-world IoT and AI challenges.
Summary updated