A Precise and Expressive Lattice-theoretical Framework for Efficient Network Verification
Alex Horn,Ali Kheradmand,Mukul R. Prasad
DOI: https://doi.org/10.48550/arXiv.1908.09068
2019-08-24
Abstract:Network verification promises to detect errors, such as black holes and forwarding loops, by logically analyzing the control or data plane. To do so efficiently, the state-of-the-art (e.g., Veriflow) partitions packet headers with identical forwarding behavior into the same packet equivalence class (PEC).
Recently, Yang and Lam showed how to construct the minimal set of PECs, called atomic predicates. Their construction uses Binary Decision Diagrams (BDDs). However, BDDs have been shown to incur significant overhead per packet header bit, performing poorly when analyzing large-scale data centers. The overhead of atomic predicates prompted ddNF to devise a specialized data structure of Ternary Bit Vectors (TBV) instead.
However, TBVs are strictly less expressive than BDDs. Moreover, unlike atomic predicates, ddNF's set of PECs is not minimal. We show that ddNF's non-minimality is due to empty PECs. In addition, empty PECs are shown to trigger wrong analysis results. This reveals an inherent tension between precision, expressiveness and performance in formal network verification.
Our paper resolves this tension through a new lattice-theoretical PEC-construction algorithm, #PEC, that advances the field as follows: (i) #PEC can encode more kinds of forwarding rules (e.g., ip-tables) than ddNF and Veriflow, (ii) #PEC verifies a wider class of errors (e.g., shadowed rules) than ddNF, and (iii) on a broad range of real-world datasets, #PEC is 10X faster than atomic predicates. By achieving precision, expressiveness and performance, this paper answers a longstanding quest that has spanned three generations of formal network analysis techniques.
Networking and Internet Architecture,Logic in Computer Science