Sketchlet: Partitioning Traffic for Sketch with Clustering on Multi-core Commodity Servers

Network traffic measurement using Sketch has become an efficient and reliable means of network management. However, limited by the structure of Sketch, achieving high throughput at 100 Gbps or above on commodity server platforms is challenging. The reason is that each counter in Sketch is associated with multiple flows, making it difficult to direct flows across multiple processor cores without sharing the same counter. To the end, we propose the Sketchlet abstraction. By analyzing the positions of counters associated with flows in memory, Sketchlet establishes clustering relationships between flows. Flows in the same cluster are directed to the same processing core, and the traffic is thus partitioned. Hence, each cluster’s flows to access only one sketch slice (i.e., a sketchlet) of the entire Sketch structure. We have implemented a prototype system for Sketchlet abstract and conducted rigorous experiments. The preliminary experimental results show that the packet processing throughput of Sketchlet is 2.4 × higher than Sketch with locking under both using four cores.