Emerging high-quality real-time applications require consistently low latency, which is often disrupted by latency spikes. We identify the reason as the mismatch between the abrupt bandwidth reallocation on routers and gradual sending rate reaction of congestion control. For example, when a burst of new flows arrives, queue schedulers such as fair queueing immediately reallocate the bandwidth for existing and new flows. However, the flow's sending rate, determined by the congestion control algorithm (CCA), needs several RTTs to converge to the new available bandwidth, during which severe stalls occur. This has been increasingly critical with the demand on consistent low latency. In this paper, we present Confucius, a practical queue management scheme that reallocate the bandwidth for flows following CCA's reaction. Confucius slows down bandwidth adjustment to match the reaction of congestion control, so that the end host can reduce the sending rate without overshooting the network. Confucius is designed for offering real-time flows with consistently low latency regardless of uncertain competition. Experiments show that Confucius reduces the stall duration by more than 50% against existing practical schemes, while competing flows also fairly enjoy on-par performance.Available at: https://github.com/hkust-spark/confucius-qdisc

The development of large language models (LLMs) poses new challenges in data center network topology design. To assist in exploring topology design, we propose ATOP, an Automated Topology Optimization Pipeline, which models network topology as a set of hyperparameters, enabling the discovery of potential topologies. With various optimization algorithms and customizable optimization objectives, ATOP achieves automated topology optimization on a scale of tens of thousands of GPUs. We apply ATOP on network topologies for 256, 1024, 4096, and 16384 GPUs, optimizing performance under LLMs training traffic patterns, collective communication performance, fault tolerance, and network cost. We also evaluate ATOP in different scenarios: building, optimizing, and expanding a data center. From ATOP's results, we discover a new topology — ZCube, which reaches the highest cost-effectiveness across various GPU scales. Simulation results show that ZCube, compared to the previous state-of-the-art topologies, including Rail-optimized Fat-tree (ROFT), Rail-only, and HPN, improves end-to-end LLM training speed by 3% to 7% and reduces network hardware costs by 26% to 46%. We also construct ZCube on a real-world testbed. Results show that ZCube reduces hardware costs by 25% compared to Rail-Optimized Topology while maintaining the same all-reduce and all-to-all performance.

The application of machine learning to large datasets has become a core component of many important and exciting software systems being built today. The extreme value in these trained systems is tempered, however, by the difficulty of constructing them. As shown by the experience of Google, Netflix, IBM, and many others, a critical problem in building trained systems is that of feature engineering. High-quality machine learning features are crucial for the system's performance but are difficult and time-consuming for engineers to develop. Data-centric developer tools that improve the productivity of feature engineers will thus likely have a large impact on an important area of work. We have built a demonstration integrated development environment for feature engineers. It accelerates one particular step in the feature engineering development cycle: evaluating the effectiveness of novel feature code. In particular, it uses an index and runtime execution planner to process raw data objects ( e.g. , Web pages) in order of descending likelihood that the data object will be relevant to the user's feature code. This demonstration IDE allows the user to write arbitrary feature code, evaluate its impact on learner quality, and observe exactly how much faster our technique performs compared to a baseline system.

Social media nowcasting--using online user activity to describe real-world phenomena--is an active area of research to supplement more traditional and costly data collection methods such as phone surveys. Given the potential impact of such research, we would expect general-purpose nowcasting systems to quickly become a standard tool among noncomputer scientists, yet it has largely remained a research topic. We believe a major obstacle to widespread adoption is the nowcasting feature selection problem. Typical nowcasting systems require the user to choose a handful of social media objects from a pool of billions of potential candidates, which can be a time-consuming and error-prone process. We have built RINGTAIL, a nowcasting system that helps the user by automatically suggesting high-quality signals. We demonstrate that RINGTALL can make nowcasting easier by suggesting relevant features for a range of topics. The user provides just a short topic query (e.g., unemployment) and a small conventional dataset in order for RINGTALL to quickly return a usable predictive nowcasting model.