SMT-COMP

The International Satisfiability Modulo Theories (SMT) Competition.

GitHub

Home
Introduction
Benchmark Submission
Publications
SMT-LIB
Previous Editions

SMT-COMP 2021

Rules
Benchmarks
Tools
Specs
Parallel & Cloud Tracks
Participants
Results
Slides

SMT-COMP Cloud and Parallel Tracks

05 Feb 2021

This year we are planning new, experimental, cloud and parallel tracks for SMT-COMP. Similar tracks were organised in the SAT competition 2020 and to our understanding the competition had a positive impact on the development of parallel SAT solvers (see https://satcompetition.github.io/2020/).

The Concept

The goal in both the cloud and the parallel tracks will be to measure the success of a solver in solving a single, hard instance. This will be done by giving solvers instances one at a time, similar to the SMT-COMP single-query track. The participating solvers will be scored based on the number of instances that a solver solves within the per-instance wall-clock time limit and the total run time, similar to the single-query track’s parallel score.

For these tracks we need to choose in total 400 benchmarks from the single-query track logics, and we are specifically reaching out to you, the community and especially the competitors, for suggesting suitable instances to be included in the tracks. In addition we will try to identify instances that are considered interesting.

The solver submission rules follow those of the rest of the tracks. However, on this track we do accept portfolio solvers, as defined in the rules, as competitors. We encourage submission of non-portfolio solvers and reserve the right to give special mentions to non-portfolio solvers in reporting the results.

While the standard competition will be run in StarExec, the parallel and cloud tracks will be run on Amazon Web Services. AWS has kindly agreed to sponsor the participants in the testing phase.

Cloud track

The Cloud Track evaluates the effectiveness of parallel SMT solvers to run in a distributed manner. The solvers participating in this track will be executed with a wall-clock time limit of 20 minutes running on 100 m4.4xlarge machines in parallel. Each m4.4xlarge machine has 16 virtual CPUs and 64 GB memory. Communication between the machines is possible using MPI and SSH.

Participants of this track are required to submit their solver via a GitHub repository (which can be private). The repository should contain a docker file that compiles the solver. As an example, scripts for account configuration and instructions to run HordeSAT in the default configuration are available at https://github.com/aws-samples/aws-batch-comp-infrastructure-sample.

Parallel track

The solvers participating in this track will be executed with a wall-clock time limit of 20 minutes, thus similar to the Single Query Track. Each solver will be run on a single AWS machine of the type m4.16xlarge, which has 64 virtual cores and 256GB of memory. More details about m4.16xlarge nodes can be found here.

Similar to the Cloud Track, participants of this track are required to submit their solver via a GitHub repository (which can be private). The repository should contain a docker file that compiles the solver.

Solver Submission to Parallel and Cloud Tracks

In order to participate in the Cloud or Parallel Track please send an email to aws-smt-comp-2021@googlegroups.com containing the following:

  1. name of the solver and a list of the authors
  2. your AWS account number
  3. the URL of the GitHub repository including the branch
  4. the full, 40-character SHA-1 hash of the commit

Sincerely,

The organizing team