Making Datacenter Software Switches Split Network Traffic More Accurately and Efficiently

What is it about?

Modern datacenters rely on many network paths between servers to keep cloud services fast and reliable. A software switch decides which path each network connection should use. This operation, called traffic splitting, sounds simple: if one path is given more weight than another, it should receive a larger share of the connections. In practice, however, this paper shows that traffic splitting in software switches is not as accurate or as lightweight as many systems assume. Existing techniques can send too many connections to some paths and too few to others, creating avoidable congestion. Some methods improve accuracy, but only by using more CPU cycles, memory, or per-packet processing time inside the switch. We introduce VALO, a new traffic splitting technique for software switches. VALO is built on two ideas. First, a score graph models how existing score-based traffic splitting actually distributes connections. Second, VALO gravity adjusts the scoring process so that the final connection distribution better follows the intended path weights. VALO is implemented in Open vSwitch and evaluated with real-world traces and datacenter workloads, including web search, data mining, deep learning, and in-memory cache services. The main message is that traffic splitting should not be treated as a solved low-level detail. Its accuracy and overhead directly affect datacenter application performance, and improving it can make multipath networking faster and more efficient.

Featured Image

Photo by Adrien on Unsplash

Photo by Adrien on Unsplash

Why is it important?

Many datacenter networking studies and systems focus on how to choose path weights based on congestion, failures, or link capacity. This paper highlights a missing layer in that design: even if the weights are well chosen, the software switch must still realize those weights accurately and efficiently. When traffic splitting is inaccurate, some paths become overloaded while others are underused. This can increase flow completion time for real workloads. When traffic splitting is resource-heavy, the switch spends more CPU cycles and time just deciding where packets should go. Both problems become more serious as datacenters run communication-intensive services such as AI training, analytics, and large-scale cloud applications. VALO addresses both issues at the same time. In the paper’s evaluation, VALO achieves much higher traffic splitting accuracy and much better resource efficiency than existing techniques. It also reduces communication completion time for real datacenter workloads. These results suggest that future work on datacenter load balancing, multipath routing, software switches, and cloud networking should consider not only how path weights are computed, but also how accurately and efficiently those weights are enforced. Key takeaways: 1. Traffic splitting in software switches is a critical but often overlooked part of datacenter networking. 2. Existing traffic splitting techniques can be inaccurate, resource-inefficient, or both. 3. VALO uses score graph modeling and VALO gravity to match traffic distribution more closely to intended path weights. 4. VALO is implemented in Open vSwitch and evaluated with real traces and datacenter workloads. 5. Researchers building on multipath routing or datacenter load balancing may need to account for traffic splitting accuracy, rather than assuming it is ideal.

Perspectives