TL;DR
This paper introduces UDT, a UDP-based transport protocol designed to efficiently utilize high-speed, high-bandwidth-delay product (BDP) networks, which TCP struggles with. UDT achieves this through a novel congestion control algorithm and a flexible architecture that allows for customizability, making it a practical solution for data-intensive applications.
What Problem Does This Paper Address?
The internet has seen a dramatic increase in bandwidth, with speeds reaching 10Gb/s and beyond. This has enabled new data-intensive applications, particularly in scientific research, that require moving massive datasets. However, the traditional Transmission Control Protocol (TCP), the backbone of internet communication, performs poorly on these high-speed, high-BDP networks. Specifically, TCP's congestion control mechanisms, designed for a different era of network conditions, lead to significant underutilization of available bandwidth. This gap between network capabilities and protocol performance hinders the progress of many cutting-edge applications. The authors aim to bridge this gap by developing a new transport protocol that can effectively leverage these high-speed networks.
Methods
The core of UDT's innovation lies in its UDP-based transport protocol combined with a sophisticated congestion control algorithm.
- UDP Foundation: UDT builds upon UDP, which provides a lower overhead and more flexibility than TCP. However, it re-implements reliability and congestion control on top of UDP.
- Rate-Based Congestion Control (DAIMD): UDT employs a Decaying Additive Increase Multiplicative Decrease (DAIMD) algorithm. Unlike TCP's ACK-driven increase, UDT uses a fixed time interval (SYN, 0.01 seconds) for rate adjustments. The "decaying" aspect refers to the additive increase parameter, which decreases as the sending rate increases, aiming for efficiency and stability.
- Bandwidth Estimation: UDT uses receiver-based packet pairs to estimate the available link capacity. This involves sending pairs of packets with minimal inter-packet delay and analyzing their arrival times at the receiver.
- Composable UDT: To cater to diverse application needs and facilitate research, UDT offers a Composable UDT (CUDT) framework. This allows developers to easily plug in and test various congestion control algorithms (including TCP variants) and customize protocol behavior through a well-defined API.
- Connection Setup: UDT supports both traditional client/server and rendezvous connection setup, which is crucial for traversing firewalls.
- Packet Structures: UDT defines distinct packet structures for data and control information, including sequence numbers, message sequence numbers, and timestamps for efficient data handling.
Datasets
Not discussed in the paper. The paper focuses on protocol design and experimental evaluation using real-world network testbeds.
Key Insights and Ideas
- TCP's Limitations on High-BDP Networks: The paper clearly demonstrates that TCP's performance degrades significantly on high-speed, high-delay networks, highlighting the need for alternative solutions.
- Rate-Based Control for High-Speed Networks: UDT's shift to a rate-based congestion control mechanism, decoupled from individual ACK events, is a key insight for achieving high throughput on these networks.
- Composable Architecture for Flexibility: The Composable UDT framework is a significant contribution, enabling rapid prototyping and deployment of new congestion control algorithms without modifying the operating system kernel. This democratizes research and development in this area.
- UDT's TCP Friendliness: The paper shows that UDT can be designed to be TCP-friendly, especially in low-BDP environments, and can coexist with TCP without significantly starving it.
- User-Level Implementation Advantage: Implementing UDT at the user level, rather than in the kernel, simplifies deployment and allows for greater flexibility, even with a slight increase in CPU usage.
Key Results
- High Throughput: UDT consistently achieves high throughput, often reaching over 90% of the link capacity on 1Gb/s networks, significantly outperforming tuned TCP in many scenarios. For example, on a 1Gb/s link with 110ms RTT, UDT reached ~940Mbps, while TCP reached ~128Mbps.
- Good Fairness: Experiments show that UDT exhibits good intra-protocol fairness, with concurrent UDT flows sharing bandwidth equitably, even with varying bottleneck bandwidths and RTTs. Fairness indices were consistently close to 1.
- Stability: UDT demonstrates stable performance with low oscillations, as indicated by low stability index values.
- CPU Efficiency: While UDT's user-level implementation incurs a slight CPU overhead compared to kernel-level TCP, it is considered acceptable for high-speed data transfer. For memory-to-memory transfers at ~970Mb/s, UDT sending/receiving CPU usage was around 43%/52%, compared to TCP's 33%/35%.
- Composable UDT Effectiveness: The paper demonstrates that implementing various TCP algorithms (like NewReno, Vegas, FAST TCP) using Composable UDT requires significantly fewer lines of code than native kernel implementations, showcasing its expressiveness and simplicity.
Limitations and Future Work
- CPU Usage: While acceptable, the user-level implementation of UDT does incur a higher CPU cost than kernel-level TCP, which could be a bottleneck in extremely high-throughput scenarios with limited CPU resources.
- Deployment: Although UDT is designed for easier deployment than kernel-level modifications, widespread adoption still requires user and application awareness.
- Future Work: The authors suggest that Composable UDT can be used to implement a wider range of protocols and algorithms, and further research can explore more advanced congestion control strategies and optimizations. The paper also hints at the potential for UDT to be used in real-time streaming applications beyond bulk data transfer.
Summary updated