Improving Traffic Splitting in Datacenter Software Switches for Faster and Efficient Networking

What is it about?

Modern datacenters, where many servers work together, use multiple network paths to send data between machines. This helps achieve faster and more reliable communication needed for services like web search, data mining, and training deep learning models. To use these multiple paths effectively, datacenters rely on a method called traffic splitting. Traffic splitting ensures that network connections are divided across different paths based on assigned weights that reflect how capable or busy each path is. These weights control the amount of traffic a path should handle to balance the load well. Software switches are commonly used in datacenters to perform traffic splitting. They make quick decisions at the server level about which network path each data connection should take. Previous methods assumed these software switches could split traffic accurately and without much cost. However, experiments show these assumptions are wrong. Existing traffic splitting techniques often assign connections to paths inaccurately, causing some paths to be overloaded while others are underused. Also, these methods demand significant computing resources in the switch, which can slow down data processing. This research introduces a new traffic splitting technique named VALO. VALO uses a mathematical model called a score graph to estimate how traffic volumes should be split according to path weights more accurately. It also uses a new idea called VALO gravity, which cleverly adjusts how connections are distributed to better follow the desired traffic proportions. VALO is implemented in Open vSwitch, a widely used software switch, and tested with real-world datacenter workloads.

Featured Image

Photo by Adrien on Unsplash

Photo by Adrien on Unsplash

Why is it important?

Efficient and accurate traffic splitting in datacenter networks directly impacts how quickly data moves between servers, which affects the performance of many online services and cloud applications. When traffic is split poorly, some paths become congested, causing delays. This extends communication time, reducing the speed of important tasks such as data analysis and AI training. Additionally, inefficient traffic splitting wastes CPU and memory resources in switches, increasing operational costs and potentially limiting scalability. By improving traffic splitting accuracy and resource usage, VALO can reduce data transmission delays and CPU consumption significantly. This leads to faster completion of networking tasks and better use of hardware resources in datacenters. For organizations relying on cloud infrastructure and large-scale data services, these improvements can translate into faster services and lower costs. Key Takeaways: 1. Existing traffic splitting methods in software switches are often inaccurate, causing some network paths to be overloaded and slowing down communication. 2. Current techniques also use excessive CPU and memory in switches, creating inefficiencies and limiting network performance. VALO introduces new mathematical models that improve the accuracy of traffic distribution according to path weights. 3. VALO greatly reduces resource usage, making traffic splitting more efficient and faster to perform. 4. In tests with real datacenter workloads, VALO speeds up communication by 1.3 to 2.5 times compared to previous methods, benefiting cloud services and AI training.