Publications

Control Channel Isolation in SDN Virtualization: A Machine Learning Approach

2023 IEEE/ACM 23rd International Symposium on Cluster, Cloud and Internet Computing (CCGrid), Acceptance rate < 20%, Bangalore, India, 2023, pp. 273-285, doi: 10.1109/CCGrid57682.2023.00034.

Yeonho Yoo, Gyeongsik Yang, Changyong Shin, Jeunghwan Lee, Chuck Yoo

Download PDF

Abstract

Performance isolation is an essential property that network virtualization must provide for clouds. This study addresses the performance isolation of the control plane in virtualized software-defined networking (SDN), which we call control channel isolation. First, we report that the control channel isolation is seriously broken in the existing network hypervisor in that the end-to-end control latency grows by up to 15 x as the number of virtual switches increases. This jeopardizes the key network operations, such as routing, in datacenters. To address this issue, we take a machine learning approach that learns from the past control traffic as time-series data. We propose a new network hypervisor, Meteor, that designs an LSTM autoencoder to predict the control traffic per virtual switch. Our evaluation results show that Meteor improves the processing latency per control message by up to 12.7x. Furthermore, Meteor reduces the end-to-end control latency by up to 73.7%, which makes it comparable to the non-virtualized SDN.

Paper summary

Isolate control traffic before bursts interfere

Predicting each virtual switch’s control traffic helps a shared network hypervisor allocate its processing capacity.

Virtual switches share the network hypervisor’s message-translation resources. Bursts from one control channel can delay another, so Meteor assigns each virtual switch a separate processing quota based on its predicted traffic.

Predict and enforce a quota per switch

A traffic meter collects recent control traffic and topology features. An LSTM autoencoder predicts the next time window, and a switch tap converts the prediction into a translation quota. Messages above the quota are buffered and processed when capacity becomes available.

Measured traffic feeds Meteor’s predictor, which supplies the quotas enforced during message translation.View full-size figure
Measured traffic feeds Meteor’s predictor, which supplies the quotas enforced during message translation.

Compare isolation and prediction windows

The implementation builds on Libera and compares prediction accuracy, control-channel isolation and processing overhead. Tests with 8–128 virtual switches compare Meteor with Libera and a reactive quota scheme for forwarding setup and network monitoring. A two-second window balances prediction accuracy and control-processing delay in these experiments.

How measured demand becomes an enforced control-channel allocation.
ComponentRole
Traffic meterCollect recent traffic and topology features
LSTM autoencoderPredict demand in the next window
Switch tapConvert demand into a translation quota; buffer excess messages

Key takeaways

  1. Control-channel isolation depends on anticipating unequal demand. Per-switch predictions turn a shared processing budget into differentiated allocations.
  2. Prediction alone cannot contain a burst. Buffering during translation is what turns the estimated allocation into enforced isolation.