GLaDOS-Michigan / Sift

The code base for the Sift prototype, as described in the ATC '22 paper "Sift: Using Refinement-guided Automation to Verify Complex Distributed Systems"

Sift

Sift provides a novel two-tier methodology that combines the power of refinement with the ability to automate proofs. Sift decomposes the proofs of complex distributed implementations into a number of refinement steps, each of which is amenable to automation. This repo shows how we use Sift to prove the refinement of several complex systems with little manual effort.

The following instructions have been tested in Ubuntu 16.04.7 LTS. Libraries and commands for other OS may differ.

Dependencies

Install Python2.7 and pip.

Install IC3PO and Ivy by running bash build.sh.

After installing IC3PO, add $(ic3po_path)/pysmt to PYTHONPATH so that ic3po can use its customized pysmt for SMT solving.

Verifying the systems

1. Starting from the spec and the implementation, we can generate the transition system in VMT format so that IC3PO can check.

   python2 ic3po/ivy/ivy/ivy_to_vmt.py isolate=system leader/leader_system.ivy leader.vmt
   

   isolate=system specifies that we're translating the refinement between the system (implementation) layer and its upper layer. We only need to pass the lowest layer, as ivy can find this layer from the included files.

2. Use ic3po to generate invariants.

   python2 ic3po/ic3po.py -g univ -v 2 -n $(test_name) leader.vmt
   

   ic3po would ask for the initial size of the instance. You can check our choice in run.sh under each experiment.

   • When ic3po finds a counter-example, it would report a violation and shows the execution trace at the end of output/$(test_name)/$(test_name).log.
   • When ic3po fins an inductive invariant for this vmt file, output/$(test_name)/$(test_name).inv shows it.
   • Some detailed information can be found in output/$(test_name)/$(test_name).result. sz-invariant shows the total number invariants in this inductive invariant; wall-time-sec and memory-mb shows the time and memory usage.

3. Remove given properties and convert operator and constants back in inv.txt.

   python2 ../converter.py output/$(test_name)/$(test_name).inv > leader.inv
   

4. Paste the leader.inv back into the isolate with other invariants. (e.g. Line 12 in leader_system.ivy).

   After adding the inductive invariants for all refinements, run ivy_check leader_system.ivy to verify the whole system.

5. bash tests.sh generates invariants for all examples.

Play with the system locally

1. You can use ivyc leader_system.ivy to generate the C++ implementation for this system, and compile the binary code.

2. To run the system, use ./leader_system $(num_of_nodes) $(my_idn) $(my_node_index). For example, we can run a system with 2 nodes by ./leader_system 2 5 0; ./leader_system 2 8 1. You should be able to see the debug log of all sending and receiving messages, and node1 is elected.

Performance evaluation

See perf_eval for more details about building/running the systems.