Explain the anomalous NASRDA replication censoring at 512 MiB

Determine why all three 512 MiB MinIO multi-site replication runs from NASRDA exceeded the 600-second measurement ceiling despite NASRDA’s substantially higher measured replication throughput for 64 MiB objects, thereby resolving whether the anomaly is specific to the largest object size or reflects another cause.

Background

The remote replication campaign measured MinIO bucket-to-bucket replication between partner-side instances and the AIR Data Centre for 4 MiB, 64 MiB, and 512 MiB objects. NASRDA achieved 2.51 MiB/s for 64 MiB objects, which was substantially above the median of cloud regions in the same round-trip-time band and projected to imply completion of a 512 MiB transfer within the 600-second polling limit. Nevertheless, all three NASRDA runs for 512 MiB objects were censored because they exceeded that ceiling.

The authors explicitly state that the measurements do not explain this discrepancy and reserve systematic attribution for future work. Resolving the cause would clarify whether object size interacts with replication behavior, endpoint conditions, or another component of the replication path.

References

NASRDA is a different case. It replicates 64. MiB COG tiles at 2.51. MiB/s — 2.67× the median of the cloud regions in its own RTT band, and the second-highest institutional ratio at that size — which projects to roughly 204 s for a 512. MiB object, comfortably inside the 600-second window. All three of its 512. MiB repetitions nevertheless exceeded the ceiling. The censoring is therefore specific to the largest object size and is not explained by the site’s sustained throughput at smaller ones. These measurements do not resolve why, and we do not speculate here; the effect is a candidate for the systematic attribution reserved for future work.

— Design and Empirical Evaluation of a Network-Centric, On-Premises Architecture for Earth Observation Data Access  (2608.20283 - Pinelo et al., 20 Aug 2026) in Section 7.2.1, Cross-site replication throughput