Optimizing throughput in the Spherical Storage tier
Six weeks after publishing the architectural overview of GZ-1 Orbis's spherical storage topology, the Spherical Storage Division is releasing its first performance audit. The initial deployment prioritized safety, distribution, and governance over raw throughput. The results were exactly as cautious as promised: a median custody-transfer latency of 840 ms between adjacent orbs, with tail latency reaching 2.3 s on the P99. Good for security. Unacceptable for inference.
The optimization pass focused on three axes: orbital scheduling, electromagnetic handshake reduction, and the introduction of a speculative pre-custody buffer. Each axis is described below with measured results. All benchmarks were run on the GZ-1 Orbis reference harness with the 14-Orbis production topology and a workload mirroring the current traffic distribution across the GZAI public endpoints.
1. Orbital scheduling
The original custody schedule rotated orbs on a fixed 120-second interval, regardless of incoming request volume. This meant that during low-traffic periods the system performed unnecessary rotations, and during burst periods it could not accelerate to keep transfers within the SLO. The new scheduler is rate-aware: it samples the request queue depth every 200 ms and adjusts the rotation interval dynamically between 60 s (peak) and 180 s (idle). The change is gated by a monotonic clock to prevent reordering anomalies.
Median latency dropped from 840 ms to 510 ms. P99 settled at 1.6 s. No anomalies in custody-chain integrity were detected over the 72-hour benchmark window.
2. Electromagnetic handshake reduction
Each orb-to-orb transfer requires an electromagnetic handshake to establish containment geometry. The original protocol was request-response by design: Orb A emits a probe, Orb B reflects, and a timing window is computed. This is correct and safe, but it costs two round-trips on every transfer.
The optimized protocol uses a continuous-wave carrier that both orbs broadcast on a shared frequency. Geometry is derived from phase-difference measurements taken over a sliding 50 ms window. The handshake is no longer a discrete exchange; it is an always-on negotiation. Because the carrier is lower-power than the previous probe-and-reflect burst, the change also reduced per-transfer energy consumption by 18%.
Median custody-transfer latency improved a further 190 ms (to 320 ms). P99 settled at 1.1 s. The carrier is classified as ZSL-2-compatible.
3. Speculative pre-custody buffer
This is the change that matters most to end-to-end throughput. In the original model the orb being passed into starts receiving data only after the previous orb has fully released and the new containment geometry is locked. There is a hard gap — approximately 140 ms — during which the data path is quiescent.
The speculative buffer eliminates this gap. The receiving orb opens a temporary read buffer 120 ms before the geometry lock is confirmed. Data flows in under a provisional containment that is rolled back atomically if the handshake fails. If the handshake succeeds (which it did in 99.7% of attempts over the benchmark window) the buffer is committed and zero-copy handed to the live storage path. The rollback path copies the buffer contents to a cold slot and logs an E-117 advisory; it never corrupts data.
Median custody-transfer latency dropped to 180 ms. P99 is 0.68 s. This is within the published SLO for GZ-1 inference and represents a 4.7× improvement over the original baseline.
| Metric | Baseline | Post-Optimization | Improvement |
|---|---|---|---|
| Median custody transfer | 840 ms | 180 ms | 4.7× |
| P99 tail latency | 2,300 ms | 680 ms | 3.4× |
| P50 throughput | 12.4 req/s | 58.1 req/s | 4.7× |
| Per-transfer energy | 1.00× | 0.82× | −18% |
| Rollback events (E-117) | 0/10,000 | 3/10,000 | 0.03% |
We did not make the balls faster. We made the hand-offs less wasteful. The balls themselves are exactly as heavy as they always were.
The speculative buffer is the only change classified at a higher safety boundary than the base architecture: ZSL-3, safe to deploy under continuous observation with an E-117 rollback budget of no more than 50 events per 10,000 custody cycles. The observed rate of 3/10,000 is comfortably within budget. The Spherical Storage Division will continue to monitor the buffer under the same cadence and will publish quarterly roll-rate reports.
A fourth optimization is already in design: a predictive pre-fetch that loads the next orb's data into the speculative buffer based on request-pattern heuristics rather than hard schedule triggers. Early simulations show median latency approaching 90 ms, but the safety case is not yet complete. No deployment date is committed.
The balls are not lighter. They just move through the hands a lot faster now.
