Network Configuration Verification & Repair
Many networks employ complex, and hence error-prone, routing control
plane configurations. In many cases, the impact of errors manifests
only under failures, leading to devastating effects. Thus, it is important to
proactively verify and correct control plane behavior under arbitrary link
failures.
Abstract Representation for Control Planes
State-of-the-art control plane verifiers are either too slow or impractical to
use for verification tasks. We have created a new high level abstraction for
control planes, ARC, that supports fast control plane analyses under arbitrary
failures. ARC can check key invariants without generating the data
plane--which is the main reason for current tools' ineffectiveness. We have
developed algorithms to automatically derive a network's ARC from its
configuration files and formally formally proven ARC's usefulness in proactive
analysis under arbitrary failures. Our evaluation over 300+ production
data center networks has shown that ARC can verify key invariants in under one
second in most cases, which is orders-of-magnitude faster than the
state-of-the-art.
Control Plane Repair (CPR)
Inspired by recent work in automatic program repair, we introduce CPR, a
system that automatically computes correct, minimal repairs for network
control planes. CPR casts configuration repair as a MaxSMT problem whose
constraints are based on a digraph-based representation of a control plane's
semantics. Crucially, this representation captures the dependencies between
traffic classes arising from the cross-traffic-class nature of control plane
constructs. The MaxSMT formulation accounts for these dependencies whilst also
accounting for all policies and preferring repairs that minimize the size
(e.g., number of lines) of the configuration changes. Using configurations
from 96 production data center networks, we show that CPR produces repairs in
less than a minute for 98% of the networks, and these repairs require changing
the same or fewer lines of configuration than hand-written repairs in 79% of
cases.