You light up a fast link, start a transfer, and watch the throughput graph flatline well below line rate. The network isn't the problem. The disk is.
The storage can't keep up
Production storage — parallel file systems, NAS, object stores — is built for capacity and concurrency, not for feeding a single 100 Gb pipe at full tilt. Writes are especially punishing: the moment your data mover has to commit to slow storage, the network stalls and waits. Your expensive link idles while the disk catches its breath.
What a burst buffer does
A burst buffer is a fast NVMe staging tier between production storage and the network. Data lands on NVMe at full speed, and the mover streams from NVMe to the WAN — decoupling the transfer from the slow, bursty I/O of the underlying storage. The buffer absorbs the spikes so the network never has to wait.
Instead of running at the speed of your slowest disk operation, the transfer runs at the speed of the wire.
Built into the appliance
It isn't something you have to assemble yourself. The core Zettar zx Appliance — built on the HPE ProLiant DL380 Gen 11 — ships with roughly 51 TB of NVMe burst buffer engineered into the data path, with storage, host, NICs, and the data mover balanced as one system.
The payoff
This is a big part of how zx moved 1 PB in 29 hours at 96% link utilization with SLAC and ESnet — practically line rate, over 5,000 miles. The burst buffer kept the network full while the storage did its thing underneath. Buy more bandwidth and the disk still stalls it; design the whole path and the link finally runs flat-out.