Residual Service Accounting Prevents False LEO Delivery Completions

LEO relay networks move finite objects: sensing tiles, telemetry blocks, model updates, checkpoints. For these workloads, partial delivery is not useful if the complete object misses its deadline. The paper targets a specific failure mode in contact-plan routing: when an object is split across candidate paths, evaluating each path privately can count the same contact service more than once and make an infeasible delivery look feasible.

The authors introduce a delivery layer that sits after contact-plan route generation and asks a narrower question: given the candidate paths already produced, can the whole object be delivered before its deadline when each edge has first-in-first-out residual service? The contribution is less a new route generator than a consistency check and scheduler for finite-object completion under shared-contact contention.

Core Contribution

The central idea is residual-service accounting. Instead of treating every candidate path as if it owned the full capacity of each contact, the method tracks already committed service on each edge and evaluates completion against what remains. That distinction matters whenever chunks of the same object, or controlled competing objects, share a contact. In that setting, path-private evaluation can under-count completion time because two planned transmissions silently consume the same service budget.

The paper separates two roles that are easy to conflate. Residual accounting is a correctness mechanism for shared contacts. Two-way striping is an optional search choice that can enlarge the feasible payload region when complementary contacts are edge-disjoint and service budgets satisfy the paper's sufficient condition. The former prevents false completion claims; the latter can improve deadline feasibility, but only when the contact geometry gives it something useful to combine.

Technical Approach

The delivery layer consumes candidate paths from contact-plan route generation. For each candidate, it evaluates per-edge service in FIFO order, subtracts committed transmissions, and checks whether all chunks of the finite object reach the ground before the deadline. The paper compares single-path delivery, bounded two-way striping, and a more expensive joint-scheduler variant, with a restricted exhaustive reference used as a check on the search procedures.

Figure 1 gives the paper's system view: route generation proposes intermittent RF/optical LEO contacts, while the delivery layer accounts for residual service before accepting a deadline-bound finite object.

The authors also analyze the edge-disjoint complementary-contact case. There, the residual-aware layer reduces to fixed-path service, and the paper derives a sufficient service-budget condition under which two-way striping strictly enlarges the feasible payload region. That result is useful because it identifies when striping is a real feasibility gain rather than a bookkeeping artifact.

Results and Analysis

The most direct result is a negative one for path-private evaluation. Under controlled shared-contact contention, it under-counts completion by up to 154 s and can report finite completion for a fixed plan where residual-service evaluation finds no completion. That is the paper's strongest practical claim: without residual accounting, a scheduler can accept an object that the contact plan cannot actually deliver.

Against the restricted exhaustive reference, bounded two-way search reduces mean and median completion gaps by about 40% compared with the same-model single-path baseline. The improvement does not extend to the tail: P90 and worst-case gaps remain unchanged. The paper's diagnosis is plausible and modest. The hard cases are not mainly caused by search-grid resolution; they come from whether the selected plan is single-path or two-way in the first place.

Runtime results are encouraging for the proposed delivery layer but less so for the joint scheduler. In the procedural Walker-like contact model, single-path and LTO sequential-booking remain in the low-second range across 20 to 180 satellites and stay below the plotted 60 s budget. The joint scheduler is much more sensitive: it reaches a greedy-fallback boundary at 80 satellites with candidate count 9. The authors therefore frame the completion gains as a candidate-path-budget effect, not as evidence that the joint scheduler is the right deployment path.

The event-grid sensitivity check is also concrete. Moving from a 20 s grid to a 5 s grid makes completion times about 1 s smaller on average, a small optimistic bias rather than a safety margin. Coarser 40–60 s grids keep the worst observed completion-time bias near 15 s, less than 0.25% of the corresponding completion times, while cutting per-evaluation cost by about 3.5×. That supports the paper's use of the event grid as a controlled characterization tool, not a deployment guarantee.

Caveats

The evaluation is still a procedural study rather than an operational validation. The authors explicitly omit detailed sensitivity figures for space and use retained runtime and event-grid results as controlled characterization. Future work named in the paper includes multi-object scheduling, node-level coupling, higher striping degrees, and SGP4/TLE-based CGR/SABR integration. For practitioners, the takeaway is narrow but useful: if finite-object deadlines matter and contacts are shared, residual service should be part of the acceptance test before any routing gain is trusted.