Feasibility-Correct Orchestration of Hybrid Quantum-Classical Workloads Under Quality Constraints: What Capacity-Based Placement Misses, and What Transpilation Cost Structure Implies for Cross-Backend Reuse
Schedulers for hybrid quantum-classical workloads usually admit a quantum processing unit (QPU) whenever it has enough qubits. We ask whether such placements return usable results, and what correct placement requires. We define feasibility as a quality predicate under which QPU and classical targets must certify the same bound on the total-variation distance of the output distribution; this yields a fidelity threshold for a bounded matrix-product-state (MPS) simulator and an estimated-success-probability (ESP) threshold for a QPU. Density-matrix simulation of 1,056 small routed instances shows that the conventional register-wide ESP is optimistic in 58% of them, whereas a per-qubit form is never optimistic under Pauli noise. In a simulation study with four modelled QPUs, two classical tiers and 400-sub-task workloads of 6 to 16 qubits over ten seeds, capacity-based placement attains a feasibility-correctness of 0.16 and a fidelity-aware baseline 0.48, against a coverage ceiling of 0.86. No sub-task is feasible on a QPU but infeasible on the bond-dimension-16 classical tier, and a quality-constrained policy that selects the cheapest predicted-feasible tier attains 0.85 at 1.2 s mean latency. The pattern persists under a SABRE-style router, other calibrations and topologies, and a shifted workload with unseen circuit families and 18- to 20-qubit circuits. Classical failure tracks entangling depth rather than width, and a seven-feature classifier reaches cross-validated ROC-AUC 0.979 in distribution but fails on one unseen family. Separately, controlled compilations show that cross-backend transpilation reuse saves a median of −0.4% in the configurations tested. All QPU results are model-based.
Schedulers for hybrid quantum-classical workloads usually admit a quantum processing unit (QPU) whenever it has enough qubits. We ask whether such placements return usable results, and what correct placement requires. We define feasibility as a quality predicate under which QPU and classical targets must certify the same bound on the total-variation distance of the output distribution; this yields a fidelity threshold for a bounded matrix-product-state (MPS) simulator and an estimated-success-probability (ESP) threshold for a QPU. Density-matrix simulation of 1,056 small routed instances shows that the conventional register-wide ESP is optimistic in 58% of them, whereas a per-qubit form is never optimistic under Pauli noise. In a simulation study with four modelled QPUs, two classical tiers and 400-sub-task workloads of 6 to 16 qubits over ten seeds, capacity-based placement attains a feasibility-correctness of 0.16 and a fidelity-aware baseline 0.48, against a coverage ceiling of 0.86. No sub-task is feasible on a QPU but infeasible on the bond-dimension-16 classical tier, and a quality-constrained policy that selects the cheapest predicted-feasible tier attains 0.85 at 1.2 s mean latency. The pattern persists under a SABRE-style router, other calibrations and topologies, and a shifted workload with unseen circuit families and 18- to 20-qubit circuits. Classical failure tracks entangling depth rather than width, and a seven-feature classifier reaches cross-validated ROC-AUC 0.979 in distribution but fails on one unseen family. Separately, controlled compilations show that cross-backend transpilation reuse saves a median of −0.4% in the configurations tested. All QPU results are model-based.
We present qScheduler, a prototype software system for integrating quantum processing units (QPUs) into Slurmmanaged HPC workflows. qScheduler implements a cooperative hybrid reservation-dispatch architecture in which Slurm remains the top-level scheduler for classical jobs, while qScheduler manages QPU reservation admission, fine-grained task dispatch, and cross-tier state synchronization. A Slurm job proceeds only after successfully acquiring a QPU reservation during the prolog phase, establishing a fail-closed binding between classical execution and quantum resource availability. Within each reservation, qScheduler uses an asyncio-based scheduler with credit-governed admission, priority dispatch, anti-starvation aging, and backfill to manage quantum tasks. To prevent allocation drift, the system combines persistent reservation heartbeats, restart-safe idle-reservation garbage collection, and epilog-driven cancellation semantics that promptly refund credits on job failure. We deploy qScheduler on the Param-Shavak cluster with three heterogeneous Qiskit Aer backends [1]: FakeTorino (133 qubits), FakeSherbrooke (127 qubits), and an ideal 20-qubit backend. Experimental results demonstrate correct resource binding, hard admission denial at configured capacity limits, proportional credit fairness across jobs, and sub-10-second credit refund after cancellation. A 55-test validation suite exercises all 17 targeted openQSE requirements together with adversarial robustness fixes for non-deterministic race conditions, executor-slot deadlocks, restart recovery, and ownership-safe state transitions.
Abstract Classical-to-quantum workload migration is not a direct code-translation task because theoretical speedup, reversible mapping, circuit resources, device noise, compilation, queueing, and classical--quantum orchestration must be considered together. This work presents CQ-Migrate, a structured decision framework for determining whether a workload should remain classical, be decomposed into a hybrid workflow, or be evaluated on a quantum processing unit. Grover search is used as an empirical case study. A true two-qubit Grover circuit was validated locally, executed on the Quantum Inspire QX emulator, and then evaluated on the 17-qubit Tuna-17 superconducting processor. Four marked states were tested with five independent hardware runs per state and 1024 shots per run, totaling 20,480 hardware shots. The emulator returned the marked state in 100% of 4096 shots. Tuna-17 produced 19,076 correct measurements, corresponding to an aggregate success probability of 93.14% (95% Wilson interval: 92.79--93.48%). State-wise mean success ranged from 91.82% to 94.30%; run-level differences across marked states were statistically detectable (one-way ANOVA (F=5.99), (p=0.0062); Kruskal--Wallis (H=10.88), (p=0.0124)). The study also reports complete provider job provenance and reinterprets earlier archived IBM executions as a backend-variability control rather than a Grover validation. The results support evidence-driven, hardware-aware migration decisions without claiming practical quantum speedup for the small instance studied.
Hybrid quantum-classical performance depends on how quantum processors share classical compute, controllers, memory, and communication links. Circuit simulators and provider measurements expose kernel and job behavior, but leave topology-level questions about controller bottlenecks, quantum capacity, and mixed-workload contention unanswered. We present HybridQC, a discrete-event simulator for configurable hybrid compute units (HCUs). Users specify resource graphs containing classical compute, memory, controllers, quantum annealing (QA) and digital quantum computing (DQC) devices, and links. Jobs become dependent stages for preparation, quantum execution, transfers, reconstruction, and postprocessing. Finite capacities and selectable scheduling policies determine where stages wait and overlap. Service profiles use measurements from D-Wave Advantage 1 and Advantage 2 and IBM Kingston, Marrakesh, and Fez, separating physical quantum-processor occupancy from cloud wall time. Mean absolute percentage errors are 3.92%-8.04% for D-Wave access time and 5.26%-19.01% for IBM quantum seconds. Balanced 10x HCU scaling improves makespan by 2.19x-3.42x, while policy choice changes makespan by 1.80x for a mixed workload with 20 jobs per workload family. Scaling depends on which inputs grow: 100x data growth yields median runtime near 306 s, while 100x joint growth in circuits, shots, and depth reaches 4.806 x 10^7 s on the fixed HCU. HybridQC enables designers to study topology, scheduling, and workload growth before tightly integrated quantum-classical machines exist physically.
High Performance Computing (HPC) centers are expanding to integrate quantum resources, enabling hybrid quantum-classical workflows for complex optimization. Integrating quantum processing units (QPUs) into workload managers poses an orchestration challenge: a remote QPU introduces a second queue - a "two-queue problem" - alongside the scheduler's own. We present Fluence, a Kubernetes scheduler plugin backed by the Fluxion graph-based scheduler, enabling gang-scheduled placement for quantum-classical workloads and custom resources. First, under contention, Fluence's atomic gang placement eliminates the node-time a default scheduler wastes on partially placed gangs. Second, a synchronization primitive gates consumers behind a single producer's shared quantum task, cutting worker idle time roughly 1.2-12x under short queues and orders of magnitude under long ones. Third, policy-aware backend selection cuts mean per-run cost roughly 72x and time-to-result from hours to under two minutes. Together, these results show that quantum-awareness can be added to a cloud-native scheduler without modifying user containers.
Quantum Computing is advancing from isolated prototypes toward heterogeneous, cloud-accessed quantum processors that are noisy, capacity-limited, and operationally diverse. In this landscape, the prevalent device-centric execution model (i.e., compiling a circuit for a single backend and running all shots monolithically) constrains scale, reliability, and cost-effectiveness. This thesis argues for and instantiates a system-centric paradigm of distributed execution, which treats execution as an engineered process that orchestrates spatial and temporal distribution across multiple devices and time scales, guided by explicit Quality-of-Service (QoS) objectives. Spatial distribution aggregates multiple, heterogeneous Quantum Processing Units and supporting classical services into a shared resource pool, placing circuits, subcircuits, and workloads where their algorithmic needs best align with hardware characteristics and availability. Temporal distribution adapts execution over time, modulating sampling depth, pacing, and scheduling in response to drift, queue dynamics, and evidence of statistical convergence. Together, these dimensions are governed by policy-driven orchestrations that reconciles fidelity, latency, and cost, and that elevates user intent above device idiosyncrasies. This thesis delivers five main contributions: (1) a shot-wise execution methodology that distributes measurement shots across heterogeneous QPUs under policy control, reducing output variance and matching or surpassing single-device averages; (2) Quantum Broker, a QoS-aware orchestration framework with the Quantum QoS Specification Language (QQSpec) that consistently finds feasible, high-quality allocations and outperforms single-backend and uniform-shot baselines under competing fidelity/latency/cost goals; (3) Cut&Shoot, an integrated circuit-cutting + shot-wise pipeline that executes large circuits across multiple devices with significantly lower error than monolithic, cutting-only, or shot-wise-only baselines, at negligible time overhead relative to standard cutting; (4) a hybrid orchestration architecture, QCRAFT Scheduler–Shotwise, which combines high-level circuit scheduling with shot-wise distribution to achieve up to 95% cost reduction and 92% fewer job submissions while improving robustness to hardware variability; and (5) IncrementalExecution, an adaptive, evidence-seeking shot manager that approximates optimal allocations with minimal overhead and high fidelity across unseen circuits and backends. Evaluations on simulated and real cloud backends demonstrate practicality across heterogeneous providers. Situated within Quantum Software Engineering, this thesis promotes abstractions and runtime mechanisms that make distribution a first-class program property rather than a backend option. The resulting perspective embraces hardware diversity as an optimization opportunity, improves portability across providers, and lays a path from today’s Noisy Intermediate-Scale Quantum (NISQ) platforms to future networked and fault-tolerant regimes. The overarching vision is that, by engineering execution as a principled, evidence-seeking process, heterogeneous quantum resources become the substrate for scalable, reliable, and efficient Quantum Computing.
Hybrid quantum datacentres are advancing rapidly, and yet most of the efforts are pointing to hardware development, while the software layer that orchestrates classical and quantum resources remains underdeveloped. This thesis designs and implements a state aware hybrid orchestrator that receives workload requests and places them into the designated cluster. The orchestrator is built as a set of REST services organised into NorthBound, allocation, resource and Southbound blocks. The classical cluster is deployed on an OpenStack environment using DevStack, while the quantum cluster is emulated by quantum agents. Two placement algorithms are implemented, First Fit and Load Balancing. Both operate over both clusters, they filter and select the corresponding node that suits the logic of the algorithm. Every allocation is persisted so the system recovers the status after a restart. A natural language entry layer was also implemented, combining retrieval augmentation generation with a large language model, converting the entry into a validated typed schema and routing to the appropriate tier, while a dashboard exposes the full pipeline and the live infrastructure state with role based visibility. The system was evaluated under identical 100 request campaigns using a mock backend that enables large scale testing without saturating the physical laboratory. Load Balancing increased the number of successful allocations by 19.15 per cent over First Fit, an advantage concentrated in the RAM constrained classical tier, and the natural language layer reached a routing accuracy of 96.8 per cent. The results confirm that a natural language interface can drive an orchestrator reliably. The main limitations are the emulated quantum cluster and a single classical laboratory deployment, future work includes integrating a real quantum backend, adding authentication to the dashboard, and introducing time aware resource release.