Haoyu Zhu
Coded matrix-vector multiplication (MVM) is often evaluated by recovery latency, but in storage-constrained edge settings the preferred operating point can change when pre-storage feasibility, materialized sparsity, amortized communication, decoding overhead, and failure modes are accounted for together. This paper studies a narrow resource-accounting question: when a per-worker storage budget cannot accommodate a fixed dense coded design, can systematic low-weight redundancy provide a feasible reliability-latency compromise for sparse MVM? The simulator separates offline encoded-block placement from online worker-level vector broadcast and reports storage feasibility, unassigned tasks, amortized communication, worker nonzero operations, source-accumulation encoding cost, materialized encoded nonzeros, decoding cost, timeout-capped latency, and failure reasons. In the configured simulator, dense coding remains latency-best when storage is unconstrained, but it requires 150.9 KiB of pre-stored blocks and about 3.07× 10 4 source-accumulation operations in the baseline setting. The low-weight design obtains 0.955 recovery success, 0.1632 s success-only mean latency, 0.2003 s timeout-capped mean latency, 68.5 KiB of pre-storage, and full feasibility after storage-constrained placement under an 8 KiB stress budget where the fixed dense and budgeted-dense placements cannot place all parity blocks. Workload, placement, support, and scaling diagnostics show that the low-weight point is useful but not optimal: placement is heuristic, support construction matters, and fixed w=3 with R/K≈1/3 weakens as K grows. The contribution is therefore a reproducible resource-accounting benchmark and lightweight reference baseline, not a new straggler-optimal or literature-faithful coded-computation scheme.