About Capabilities Network News Contact
← Newsroom

Storage fabric passes one exabyte of cumulative transfer

Sep 19, 2025 JVNC Newsroom JVNC.NET

The JVNC distributed storage fabric has moved more than one exabyte of cumulative data since it entered production. That figure covers science datasets, checkpoint traffic and replicated archives across all member sites.

Peak sustained throughput during the last observing season reached 2.4 Tbps across the fabric, with a cache hit ratio of 71% on repeated dataset requests. Caching happens at the edge, so a popular dataset is usually served from a location within a few milliseconds of the requesting cluster.

Data is stored under a policy engine that understands the difference between checkpoint data, which can be evicted aggressively, and archival data, which must survive tape rotation cycles.

Members can request a private namespace with their own retention rules. There is no per-gigabyte charge inside the cooperative.

↳ More from the newsroom
01 About

JVNC is a non-profit cooperative running a low-latency research network, a distributed compute fabric and an identity layer for its member institutions.

We started in 1985 as the John von Neumann Computing Network — a handful of university sites sharing a single backbone. The ambition has not changed: make distant machines behave like nearby ones.

Today the cooperative carries science traffic between 380 institutions, operates its own optical layer, and publishes every measurement it makes.

Founded1985
Backbone capacity42 Tbps
Member institutions380
Measured uptime99.995 %
Median RTT, east corridor2.1 ms
Deterministic latency is what makes a distributed cluster behave like a single machine.
Network engineering group, JVNC
02 Capabilities

Everything a research network needs, run by the people who use it.

01 / 06

Optical Backbone Transit

Coherent 400G ZR+ lambdas on leased dark fibre, reconfigurable in under ninety seconds through a software-defined ROADM layer.

400G ZR+ROADMDark fibre
02 / 06

Supercomputing Fabric

Shared GPU partitions, large-memory nodes and a scheduler that understands topology, so jobs land close to their data.

GPUSlurmTopology-aware
03 / 06

Data Movement & Caching

Edge caches inside member campuses, namespace policies that separate checkpoints from archives, no per-gigabyte charge.

Edge cacheNamespaceTape
04 / 06

Research Identity Federation

Hardware-backed assurance levels, attribute-based authorisation, temporary rights across institutions without new credentials.

SSOABACStandards-track
05 / 06

Anycast DNS & Naming

Eleven points of presence, DNSSEC validation everywhere, free secondary service for member authoritative zones.

DNSSECAnycastIPv6-first
06 / 06

Observability & Telemetry

Hardware timestamping probes on every hop, one-way delay measurement, raw data published to members.

gNMIP4 probesOpen data
03 Network

Eleven points of presence. One flat address plan. No hidden queues.

01
Princeton
Primary peering
0.4 ms
02
Ashburn
Interchange
2.1 ms
03
Chicago
Midwest aggregation
7.8 ms
04
Amsterdam
EU interchange
38 ms
05
Frankfurt
DE-CIX presence
41 ms
06
London
LINX presence
40 ms
07
Tokyo
APAC anchor
102 ms
08
São Paulo
LATAM anchor
96 ms
09
Johannesburg
Africa gateway
168 ms
10
Singapore
South-east Asia
146 ms
11
Sydney
Oceania
178 ms
Latency is a design decisionOpen measurements Latency is a design decisionOpen measurements
04 Newsroom

Dispatches from the cooperative.

Engineering notes, allocation calls and field data — published as they happen, archived in full.

Connect your institution to the network.

Membership review takes three weeks. No fees for member institutions.