Telemetry-Based Server Selection in the Quantum Internet via Cross-Layer Runtime Estimation

TL;DR

An operating map based on requirements relating distance and entanglement rate requirements to protocol level counts, quantify how simple multiuser contention shifts the crossover, and use Sobol global sensitivity analysis to identify regime-dependent bottlenecks are derived.

Abstract

The Quantum Internet will allow clients to delegate quantum workloads to remote servers over heterogeneous networks, but choosing the server that minimizes end-to-end execution time is difficult because server processing, feedforward classical communication, and entanglement distribution can overlap in protocol-dependent ways and shift the runtime bottleneck. We propose $T_{\max}$, a lightweight runtime score that sums coarse telemetry from multiple layers to obtain a conservative ranking for online server selection without calibrating weights for each deployment. Using NetSquid discrete-event simulations of a modified parameter-blind VQE (PB-VQE) workload, we evaluate $T_{\max}$ on pools of 10,000 heterogeneous candidates (selecting among up to 100 per decision) across crossover and bottleneck-dominated regimes, including temporal jitter scenarios and jobs with multiple shots. $T_{\max}$ achieves single-digit mean regret normalized by the oracle (below 10%) in both regimes and remains in the single-digit range under classical communication latency jitter for multi-shot jobs, while performance degrades for single-shot jobs under severe jitter. To connect performance to deployment planning, we derive an operating map based on requirements relating distance and entanglement rate requirements to protocol level counts, quantify how simple multiuser contention shifts the crossover, and use Sobol global sensitivity analysis to identify regime-dependent bottlenecks. These findings suggest that simple cross-layer telemetry can enable practical server selection while providing actionable provisioning guidance for emerging Quantum Internet services.

Figures (11)

• Figure 1: Client–network–server schematic for the server selection problem, instantiated for the PB-VQE workflow. Two candidate servers (A and B) communicate with the client C over quantum channels through the shared Quantum Internet (QI). The quantum communication links (A–C and B–C) are annotated with the effective entanglement distribution rate $R$, while the classical feedforward communication latency $t_{\mathrm{cc}}$ depends on the physical channel length $d$ (propagation-limited model). Server nodes are parameterized by the server-side processing time $T_{\mathrm{srv}}$, and the client node by $T_{\mathrm{cli}}$. Parenthesized symbols indicate the telemetry components used in the runtime decomposition $(T_{\mathrm{cli}}, T_{\mathrm{srv}}, T_{\mathrm{cc}}, T_{\mathrm{ent}})$.
• Figure 2: Protocol diagram of a client--network--server workflow over the Quantum Internet. Blue denotes classical operations/communication, and orange denotes quantum ones. Latency parameters $T_{\mathrm{srv}}, T_{\mathrm{cli}}, T_{\mathrm{cc}}, T_{\mathrm{ent}}$ denote server processing, client processing, classical communication, and entanglement distribution, respectively.
• Figure 3: Bottleneck switching in $T_{\mathrm{exe}}$ illustrated by one-dimensional sweeps of $R$, $d$, and $\kappa$ around the light/heavy reference configurations in Table \ref{['tab:even']} ($S = 10$ shots). Plateaus and crossovers indicate transitions among $T_{\mathrm{ent}}$, $T_{\mathrm{cc}}$, and $T_{\mathrm{srv}}$.
• Figure 4: Mean oracle-normalized regret of server selection policies in the Even regime as a function of the number of candidate servers $M$. Shaded bands indicate 95% bootstrap confidence intervals.
• Figure 5: Oracle-normalized regret of server-selection policies in bottleneck-dominated regimes versus the number of candidate servers $M$. Shaded bands indicate 95% bootstrap confidence intervals. For visual clarity, we omit the Rank-sum baseline, whose regret is substantially larger and would compress the scale; the full curves including Rank-sum are reported in Appendix \ref{['app:ranksum_full']} (Fig. \ref{['fig:bn_full']}).