Title: Open Seismic imaging benchmarking platform for data inversion methods
Speaker: Gerard Gorman, Devito Codes and Imperial College London.
- Forewarning - imperfections ahead.
- If you see something broken...please file a PR/issue
https://github.com/devitocodes/SeismicImagingBenchmarkingSuite
Co-conspirators: F. Luporini (Devito Codes), A. St-Cyr (Shell), A. Loddoch, J. Washbourne (Chevron), A. Souza (Petrobras), K. Hester (NVidia), P. Witte (Microsoft), F. Dupros (ARM), M. Araya (Total), C. Leader (SLB), J. Tilly (BP)
Please let me know if I have missed your name (PRs welcome...)
Vision
Bring community together to:
- Standardize benchmarking of seismic imaging kernels.
- Provide open-source reference implementations for a range of architectures.
- Publish reproducible benchmark experiments.
Community/stakeholders
- Energy companies.
- O&G service companies.
- Processor manufacturers, OEMs, supercomputer/cluster manufacturers.
- Cloud providers.
- Independent software vendors.
- Academic research community.
Business case for open domain specific benchmarks
- Making informed purchasing decisions for on-prem/Cloud.
- Market value of HPC market in O&G ~1b USD.
- Software roadmaps:
- Performance-critical software must be written in vendor-native programming model.
- Unicorn developers are a scarce and high value.
- Extensive code verification and code hardening.
- Enables collaboration and code/data reuse - stop burning our human capital.
Porting key benchmarks to a new architecture, including performance analysis and tuning, typically takes multiple months of unicorn developer effort from user and vendor.
Cost estimates range from 500k-1m(?) USD per customer-architecture pair. Does not include duplication of work in consortia etc.
<rant> Apples-to-applesmystry-fruit comparisons
Common issues benchmarking against other codes:
- Difference in the PDEs (e.g., different BCs, constant vs variable density).
- Difference in discretisation, e.g.,
- Time-order, space-order, boundary conditions, (dx, dy, dz, dt), (nx, ny, nz, nt).
- Have encountered ~10 different variations of TTI.
- ...
- ...
- Difference in algorithmic optimizations, e.g., expanding box, factorization.
- Runtime choices, e.g.,
- Pure OpenMP vs MPI-x, pinning.
- Are we even computing the same solution?
- Floating point addition is non-associative.
- Lossy compression, reduced precision...
- Reproducibility:
- Lack of access to closed source codes.
- Performance data often covered by NDA.
- Problematic to get all parties using the same or twin compute environments ("it works on my laptop").
Are we comparing methods, hardware or skill?
Presenter: By moving to GPU we get a x1000 speedup.
Chair: What was your reference?
Presenter: An in-house code written in Java.
Chair: What did your roofline model look like?
Presenter: What's a roofline model?
<\rant>
Objectives of standardisation
-
Standardize problem specification, includes:
- Correctness tests.
- Performance metrics.
-
Open-source reference implementation:
- Design in reproducibility, e.g., containers to set runtime environemnent.
- Starting point porting, optimizing in-codes.
- Sandbox for vendors, customers and researchers to collaborate.
-
Increase relevance of benchmarks:
- Gradient-type benchmarks are uncommon despite the fact that backpropagation is frequently the main performance bottleneck.
-
In general, the benchmarks themselves are closed-source. This not only means significant duplication of effort but frequently leads to a lack of robust and reproducible results.
Challenges: Algorithmic development
- Improved physics models (e.g., elastic models).
- Improved numerical discretization of the forward model and its adjoint.
- Novel algorithmic research, e.g., machine learning.
Challenges: Software development and optimization:
- Ongoing optimization of software implementation on the current generation of CPUs/GPUs.
- Optimization for different runtime environments for on-prem and Cloud environments.
- Porting to new architectures for evaluation and productionization.
- Parallel programming models, e.g., MPI, OpenMP, CUDA, HIP.
Scope of benchmarking
- Seismic imaging is dominated by the cost of forward modelling and the adjoint-state method.
- Many different PDEs and choices of discretization.
- Variability in OI, memory pressure and register pressure.
- Important to choose benchmarks that reflect real use case, e.g.:
- Gradient benchmarks should be done in addition to forward model benchmarks.
- Large offsets or high frequency will necessitate distributed memory parallelization.
Out of scope
- Storage IO (with the exception is serialization for gradient computation).
- Full FWI/RTM - compute cost is too high.
Roadmap
- Learn from numerous benchmarking suites from other disciplines
- Bring together expertise from industry and academia to:
- Standardize benchmarking of seismic imaging kernels.
- Publish reproducible benchmark data.
- Provide reference implementations and maximize software reuse.
Phase I: Bootstrap the creation of v0.1
Milestones:
- Announce call for participation and create initial stakeholder group - this presentation.
- Create template benchmark specifications - encodes best practices.
- Create examplar benchmark specifications based on isotropic acoustic and TTI.
Phase I: details
Considerations when defining a benchmark specification:
- Unambiguous problem specification:
- Enable benchmark problems to be set up using different codes.
- Problem dependencies, e.g., datasets will be open access.
- Define measures of success:
- Correctness test, e.g., error norms.
- Measure stencil throughput (grid-points-per-second).
- Measure parallel efficiency.
- Problem size:
- Problem sizes should be representative of production workloads.
- Reduced problem sizes or space-order may also be necessary:
- Training, aid in code porting, evaluation on new architectures with low memory.
Submission process
- Benchmark specifications are submitted as PRs to the GitHub repo.
- The file CONTRIBUTIONS.mb Developer's Certificate of Origin 1.1.
- PRs will be peer-reviewed and published on main after revisions have been completed.
Phase II
Milestones:
- Develop reference implementations of examplar benchmarks.
- Establish and demonstrate reproducibility standards.
- Initialize benchmark league table.
- Provide reference implementations of benchmark problems using the open-source platform, Devito.
- Devito is a Python package to implement optimized stencil computation (e.g., finite differences, image processing, machine learning) from high-level symbolic problem definitions.
- Supports MPI, OpenMP, OpenACC, CUDA (HIP upcoming) and Intel and AMD x86, NVidia and AMD GPUs, ARM.
- Open source: devitoproject.org
- Commercial: devitocodes.com
Submission process
- New submissions built on examplar templates.
- Submissions via as PRs to the GitHub repo.
- IP covered CONTRIBUTIONS.mb Developer's Certificate of Origin 1.1.
- GitHub actions used to automate testing/correctness and automate benchmarking on self-hosted runners (requires some gate keeping).
- All PRs will be peer-reviewed after GitHub Actions checks have been completed.
- Published on main after revisions have been completed.
Phase III
Milestones:
- Governance structure.
- Define code and sustainability model.
- Open call for submissions.
Phase IV
Milestones:
- Bounties on improvements to existing benchmarks.
- Kaggle-inspired competitions to guide activity.
- Partner with publisher to create incentives for students and academics.