Meeting 2026 10 05

Meeting - October 5, 2026

• Date: Monday, October 5, 2026
• Time: (Slot #2) 7 pm - 8 pm US Eastern Daylight Time (View Meeting Time in Your Timezone)
• Location: Virtual Meeting.
• Meeting Link: openqse.org/captcha

Upcoming meetings

• Next: Monday, October 12, 2026 @ (Slot #1) 12 pm - 1 pm US Eastern Daylight Time (View Meeting Time in Your Timezone)

Agenda

• "Dataflows and Computational Patterns for Hybrid Quantum-Classical Scientific Computing" - Ryan Landfield (ORNL)

Attendance

Person	Institution
Thomas Naughton	ORNL
Ryan Landfield	ORNL
Laura Schulz	ANL
Abdur Rahman Hatim	IISc
Mahmoud Elkarargy	Brightskies
Yoonho Park	IBM
Eun Kyung Lee	IBM
Andy Stone	HPE
Doug Oucharek	ORNL
Nick Wright	NVIDIA
Yogesh Simmhan	IISc
Michael Ferguson	HPE
Vyron Vasileiadis	Unknown
Alain Roy	IonQ
Yasuko Eckert	AMD
Masoud Mohseni	HPE
Peter Groszkowski	ORNL
Ashish Kashinath	Unknown
Claudio Siqueira de Carvalho	IBM
Neal Erickson	Quantinuum
Jeff Heckey	AWS
Ermal Rrapaj	NERSC/LBL
Smriti Bajaj	Dell Technologies
Namit Anand	HPE
Josh Moles	IonQ
Bigeng Wang	NERSC/LBL

Notes

• NOTICE: Notes generated using AI; please verify accuracy.

1. Opening

   • Thomas Naughton welcomed attendees and introduced Ryan Landfield of Oak Ridge National Laboratory.
   • Landfield presented work titled “Dataflows and Computational Patterns for Hybrid Quantum-Classical Scientific Computing,” based on a Quantum Week paper and talk.

2. Presentation Overview (Ryan Landfield)

   2.1. Motivation and application matrix

   • The work began as an effort to catalog hybrid quantum-classical applications by algorithm, scientific goal, workflow coupling, hardware requirements, and the balance between HPC and quantum resources.
   • Comparing applications by scientific domain proved less useful than comparing how data and control move between the classical host and QPU.
   • The resulting abstraction focuses on where computation occurs, what crosses the host–QPU boundary, synchronization frequency, dependency between quantum calls, and the QPU’s computational role.

   2.2. Five recurring computational patterns

   • QOOL: quantum optimization outer-loop workflows.
   • QASP: quantum-assisted subspace projection, using the QPU to construct subspaces or projected representations.
   • QSMC: quantum sampling Monte Carlo, using the QPU primarily as a sampling engine.
   • QTES: quantum time-evolution simulation, keeping evolution on the QPU before measurement.
   • QLAO: quantum linear-algebra oracle workflows, with classical setup and post-processing around quantum linear-algebra primitives.
   • These are execution-structure categories rather than scientific-domain categories; one domain may contain multiple patterns.

   2.3. Quantum execution locality

   • Quantum execution locality describes how long useful computation remains resident on the QPU before host intervention or synchronization.
   • Low locality involves short quantum kernels separated by repeated host interaction; medium locality permits batching; high locality permits larger contiguous quantum-resident regions.
   • Locality is about the continuity and structure of quantum execution, not simply the fraction of runtime spent on the QPU.
   • The presentation gave approximate host–QPU crossing counts ranging from thousands for highly iterative optimization workflows to roughly one-to-ten for time-evolution and quantum linear-solver workflows. These were structural estimates, not measured performance results.

   2.4. Workload reformulation

   • The talk emphasized reformulating workflows to increase useful locality through batching, deferred measurement, and restructuring around subspace, sampling, time-evolution, or linear-algebra patterns.
   • A generative-model-based circuit-generation example was discussed as a reformulation of a strongly iterative optimization loop into a more batch-oriented ensemble evaluation.
   • The reported example used CUDA-Q and NVIDIA hardware; it was presented as evidence of workload reformulation, not as a direct measurement of QPU scheduling or host–QPU communication.

   2.5. Validation and benchmarking strategy

   • The framework is currently qualitative; the proposed validation strategy has two phases.
   • Phase 1 estimates structural characteristics and approximate stage timings from algorithm descriptions and published data.
   • Phase 2 implements representative QOOL, QASP, QSMC, QTES, and QLAO kernels in a common environment such as the QHET hybrid benchmarking toolkit.
   • The goal is to measure host–QPU crossings, timing distributions, synchronization costs, and pattern-specific system signatures across backends.

   2.6. QHET and QWAT

   • QHET provides a rationale for selecting benchmark workloads that span distinct hybrid execution structures.
   • QWAT (quantum workload analysis toolkit) uses execution metadata—such as job data, circuit runtime, utilization, and workload behavior—to characterize what users actually run.
   • QWAT addresses the gap between proposed allocation usage and executed workloads, supporting workload characterization, resource-demand forecasting, and validation of the pattern taxonomy.
   • The pattern is a useful unit of computational work for scheduling, checkpointing, performance, and cost analysis; completed execution still depends on the required shots or stopping criterion.

   2.7. QPASS pattern-aware scheduling

   • QPASS (quantum pattern-aware scheduling) is a prospective concept for using pattern identity and locality to select an appropriate backend and scheduling policy.
   • QPASS could combine declared application intent, inferred structural information such as dependencies, batching, iteration, and residency potential, and advertised device capabilities.
   • QPASS and QWAT are complementary rather than sequential: QPASS makes prospective placement decisions, while QWAT retrospectively measures actual execution and cost.
   • Unknown or low-confidence patterns could fall back to a general scheduling policy; unexpected telemetry could help identify new or refined patterns.

   2.8. Cost modeling and feedback loop

   • Cost is a second axis for the QPASS architecture because heterogeneous accounting models make backends difficult to compare directly.
   • QHET can measure pattern behavior by scale and backend, QWAT can measure actual resource use and cost in production, and QPASS can use those models to inform future scheduling decisions.
   • The long-term goal is to preserve vendor-native accounting while also comparing normalized measures such as cost per workflow or cost per valid result.

3. Discussion & Q&A

   3.1. Meaning and scope of locality

   • Masoud Mohseni asked whether locality was temporal rather than spatial and raised the possibility of describing locality across subsets of variables or subgraphs.
   • Landfield clarified that the current framing emphasizes when a workflow synchronizes with a classical resource and how long computation remains quantum-resident. He noted that mid-circuit measurement and other execution details may require refinement of the definition.
   • Namit Anand asked whether naturally separated quantum subcircuits could be considered local. Landfield said the current work first maps serial and iterative execution structures before extending the analysis to distributed quantum computing.

   3.2. Scheduling and software-stack integration

   • Claudio Siqueira de Carvalho asked how identified patterns might inform schedulers, quantum resource-management interfaces, and vendor software stacks.
   • Landfield said scheduler integration is not yet defined and that the immediate focus is retrospective analysis, benchmarking, profiling, and identifying the telemetry, timing, and cost data needed for deployment.
   • He also identified combinations of patterns in more complex applications as an open area for future work.

   3.3. System planning and premature optimization

   • Michael Ferguson asked whether observed patterns should influence system design now or whether the community risks designing around today’s workloads.
   • Landfield said the catalog could support resource, system, and integration planning, while acknowledging the need to account for fault-tolerant workloads and future implementations.
   • Ferguson cautioned against curtailing capabilities that future workloads may require. Landfield agreed that the field should preserve flexibility across hardware modalities and use current profiling as a baseline that can be repeated as systems evolve.
   • Jeff Heckey further asked about the risk that current use patterns could prematurely determine future device designs. Landfield emphasized keeping multiple modalities and approaches open rather than selecting a single path.

   3.4. Reproducibility, uncertainty, and dependency mapping

   • Laura Schulz emphasized that current algorithms and workflow observations should remain provisional because the field is changing rapidly.
   • Nick Wright asked about error bars and how changing assumptions would affect conclusions. Landfield described a staged process: map current hardware and timing first, then revisit the analysis as timing, coherence, integration, and other system characteristics improve.
   • Schulz connected the locality analysis to decomposing a workflow across software, hardware, networking, timing, and other dimensions, so that dependencies and downstream effects of changing one component can be identified.
   • Landfield agreed that quantifying these coupling terms and stresses would be valuable.

4. Key Takeaways

   • Hybrid quantum-classical applications can be compared by execution structure and host–QPU interaction, not only by scientific domain or algorithm name.
   • Quantum execution locality is a structural property describing continuity of useful QPU-resident computation.
   • Workload reformulation may improve system use by reducing synchronization and creating larger batches or contiguous execution regions.
   • Pattern catalogs can inform benchmarking and future resource planning, but should remain open to new algorithms, hardware, and workflow structures.
   • Understanding scheduler interfaces, telemetry, pattern combinations, and spatial or distributed locality remains future work.

5. Action Items

   • No explicit action items captured.

6. Closing

   • Thomas Naughton thanked Landfield and the attendees for the presentation and discussion.
   • The next meeting was announced as a talk by Graham Schelle (12-oct), and the meeting was adjourned.
   • Paolo Viali will speak on 26-oct.

Files

• Download Transcript