GPU-to-GPU Communication Across Nodes: What Actually Works
A practical guide to NVLink, NVSwitch, InfiniBand, RoCE, and GPUDirect for multi-node GPU clusters. Cut through the jargon and understand what hardware and software you actually need for distributed training....
Understanding RX vs TX - Making Sense of Network Traffic Direction with Real Examples
Learn the fundamentals of RX (receive) and TX (transmit) metrics in network monitoring through real-world Mellanox ONYX switch examples, common traffic patterns, and practical monitoring scripts...
Centralised Control Planes for SaaS - Part 1: SaaS Business Models
For a couple of years, my journey has revolved around constructing control planes for data infrastructure startups. As an engineer I have been fortunate to gain invaluable insights into the challenges and intricacies of developing successful SaaS products. Drawing from my firsthand experiences, this series of blog post's delves into the lessons learned and shares my goals at BaaZ, where I am building a centralised control plane for SaaS infrastructure. Join me as we explore the critical aspects and key considerations of constructing effective control planes in the dynamic and competitive SaaS industry....
Centralised Control Planes for SaaS - Part 2: Stateless Async Event Handling
In our previous post, "Centralised Control Planes for SaaS - Part 1," we delved into the world of SaaS business models and the challenges that arise when constructing a centralised control plane. In this instalment, we will take a closer look at the crucial first step in building a control plane application developer versus infrastructure developer....
Centralised Control Planes for SaaS - Part 3: Logical and Physical Models
In our previous post, "Centralised Control Planes for SaaS - Part 2," we discussed about the approaches to build a stateless control plane and how to handle async communication b/w the data planes and cloud provider. In this instalment we take dive into the physical and logical models in a SaaS control planes....
State Management for Infra Products - Part 1: State Machines vs Observed State
The goal of this series is to provide an in-depth understanding of building state-driven infrastructure products, as opposed to API-driven ones. Much of the content is drawn from my experience building control planes for infrastructure. Before Kubernetes, I worked as an infrastructure developer on platforms like BOSH, Cloud Foundry, and Mesos. However, Kubernetes fundamentally changed my approach to state management and the construction of states....
Understanding Power Management in GPU via PCIe
Learn how modern GPUs implement intelligent power management through PCIe generation scaling and why your high-end GPUs might be operating at Gen 1 speeds during idle periods...
Demystifying Helm and Operator pattern.
Discover insights and explanations about Helm and Operator Pattern in the DataInfra.io blog. Gain a clearer understanding of these tools and patterns used for efficient infrastructure management....