How long does it take to transfer a petabyte?
Transfer time is data size divided by the throughput you actually achieve — not the link’s nameplate speed. The table assumes a well-engineered path that sustains about 90% of the link (line rate).
| Data | 10 Gbps | 25 Gbps | 100 Gbps | 400 Gbps |
|---|---|---|---|---|
| 10 TB | 2.5 h | 59 min | 15 min | 4 min |
| 100 TB | 25 h | 9.9 h | 2.5 h | 37 min |
| 1 PB | 10 days | 4.1 days | 25 h | 6.2 h |
| 10 PB | 103 days | 41 days | 10 days | 2.6 days |
Formula: hours = data (TB) ÷ (link Gbps × 0.9 × 0.45 TB per hour per Gbps). If your tools realize only part of the link, scale accordingly — at 20% utilization, a petabyte on a 100 Gbps link takes about 4.6 days instead of one. Try your own numbers in the transfer-time calculator.
Why do big transfers run slower than your link?
Once a fast link is in place, the network is rarely the bottleneck. Throughput is set by the slowest part of the end-to-end path:
- Storage I/O, especially writes. Production file systems and object stores are built for capacity and concurrency, not for feeding one fast pipe. When the receiver can’t commit data fast enough, the network waits.
- Single-stream software. One TCP stream must keep the whole bandwidth-delay product in flight — about 1.25 GB on a 100 Gbps path with a 100 ms round trip. Single-stream tools can’t, so long, fast paths sit mostly idle.
- Host architecture. CPU, memory, PCIe, and NIC placement all cap what one server can push.
- Security and integrity overhead. Encryption and checksums are non-negotiable at this scale, so they have to be engineered into the path rather than bolted on.
That is why buying a bigger link alone rarely makes a transfer faster. The network isn’t the bottleneck — the whole path is.
What are your options?
| Approach | Good for | Where it falls short at petabyte scale |
|---|---|---|
| General-purpose tools (rsync, scp) | Everyday sync between two hosts | Single stream, so fast links stay mostly idle; scaling means hand-sharded scripts |
| Grid and managed-transfer tools (Globus, MFT) | Orchestration, sharing, governance | Throughput still depends on the endpoints and storage you tune yourself |
| Shipping drives or transfer devices | One-time moves with no usable network | Load, transit, and unload time; data leaves your custody; no help for recurring or continuous flows |
| Cloud providers’ standard transfer paths and CLIs | Convenience inside one cloud | Standard paths carry a performance penalty — in one measured case zx was 5.9× faster than aws-cli |
| A co-designed data-movement appliance | Petabyte-scale, recurring, multi-site movement | A capital purchase — worth it when data movement is on the critical path |
For a deeper comparison, see rsync vs Globus vs a data-movement appliance, or the head-to-heads with rsync, Globus, and Aspera.
A checklist for moving petabytes
- Measure the whole path, not just the link. Test storage read and write throughput at both ends, not only the network.
- Stage through fast storage. An NVMe burst buffer decouples slow, bursty storage I/O from the network.
- Move in parallel and scale out. Many streams across many nodes, so no single stream, file, or server caps throughput.
- Verify every byte. End-to-end checksums on every file or object, plus encryption in transit.
- Plan for distance. Use a mover that holds throughput on long, high-latency paths.
- Keep native formats. Avoid proprietary wrappers, so data lands ready to use and nothing is locked in.
- Size the cutover. For migrations, every extra day means running two environments — estimate the overlap cost.
What line rate looks like in practice
With SLAC National Accelerator Laboratory and ESnet (U.S. DOE), Zettar moved 1 PB in 29 hours over a 5,000-mile path, encrypted and checksummed, at 96% link utilization — while capped at 80 Gbps to share the network. On a full 100 Gbps link, that is a petabyte a day. In a separate test, zx moved a 1.2 TiB cryo-EM dataset from KEK in Japan to AWS N. Virginia in about 40 minutes, versus 235 minutes for aws-cli (Fang et al., arXiv:2512.15028).
The Zettar zx Appliance packages that engineering — storage, host, network, and software co-designed, with an NVMe burst buffer — as one turnkey, pretuned system, so you get line rate the day it is racked.
