Octodec — flash throughput, hard-drive economics, one cluster

One rack against the alternatives.

All three at 8+3 erasure coding. GPFS is an estimate.

Usable PB per $1M

Octodec

11.8 PB

IBM Storage Scale

8.2 PB

Ceph

7.8 PB

Throughput per $1M

Octodec

~118 GB/s

IBM Storage Scale

~54 GB/s

Ceph

~25 GB/s

	Octodec rackSMR	Ceph	IBM Storage ScaleGPFS, estimate
Throughput	~200 GB/s *	~50 GB/s	~100 GB/s
Usable capacity / at practical fill	20.1 / 19.0 PB (95%)	15.4 / 13.1 PB (85%)	15.0 / 14.2 PB (95%)
Cost per TB at practical fill	$90–100	$151	$129
Drives / JBODs / servers	$1.39M / $160k / $150k	$1.49M / $160k / $330k	$1.49M / $160k / $195k
Total rack cost	$1.70M	$1.98M	$1.84M

* Peak, outer tracks — modelled from per-drive throughput measured on our own cluster. Competitor figures are from published specifications and list pricing; GPFS is an estimate. End-to-end S3 figures, TCO and per-workload models available on request.

Four ways Octodec lowers TCO

Bigger SSDs don’t mean more throughput.

An SSD is capped by its interface, so a 60 TB drive streams about as fast as a small one at many times the price. The same money buys enough hard drives to match that throughput and far exceed the capacity. So Octodec uses both — disk for the bulk of the data, flash for the working set a job is actively reading — and moves data between the two automatically.

01

Throughput maximum read speed from every drive

90%+

of cluster read throughput free for large objects

Parallel spindle reads. Each large object streams from many drives at once, so speed grows with size.
Small objects stay whole. One read is one disk seek, so small files never drag down large-object reads.
Full-dataset reads in a day, not a week. The same drives read naively take three to four times longer.

02

Capacity more usable TB per dollar

+2 PB

recovered per 20 PB rack. Append-only writes hold performance to 95% fill, where typical estates strand capacity above 80–85%.

Flexible erasure coding. Any layout from 6+2 (25% overhead) to 60+4 (6.25%).
Higher density. Denser drives, JBODs and racks with fewer servers and less CPU and RAM per PB.

03

Operations no storage team required

0

dedicated storage-ops headcount

Self-healing. Failed drives rebuild automatically from parity on 1 GB shards. No spare, no colo trip.
No centralised state. Add, move and remove drives live; swap failures in a scheduled batch.

04

Tiering hot tier at line rate

0 rewrites

on disk when a working set is promoted to flash

Line-rate promotion. The working set moves to flash at full read rate; writes made there fold back as changes only, not whole objects.
Lean flash tier. Flash holds only what the job needs now, with no application change.

Nothing above it has to change.

S3NFS9PNBD

Measured against a raw fio baseline

93%

Read retained

95%

Write retained

99.87%

NFS POSIX, pjdfstest

99.85%

9P POSIX

No application changes: S3 clients run untouched.

Real POSIX: Verified. No recompilation or driver work.

Hardware you already own: Commodity servers and JBODs, SMR and CMR mixed.

Adopt incrementally: Runs alongside existing systems, taking over drives.

Built for hundreds of petabytes that get read, not archived.

Not a fit for smaller clusters that must be read end to end at full speed — if that’s you, the deployment review will say so.

Neoclouds and training infrastructureMoving the capacity tier off all-flash while keeping GPUs fed, inside a jurisdiction customers can name.

Enterprise AI and ML platformsNo cloud egress, and dedicated, predictable throughput instead of a shared tier.

Physical AI and sensor fleetsContinuous video and telemetry kept accessible and re-read for training.

Genomics and life sciencesPetabytes a year, re-read on every reanalysis, with cheap multi-year retention.

Media and post-production4K, 8K and HDR footage streamed concurrently from an on-prem active archive.

Teams running Ceph or flash capacity tiersThe same capacity and throughput at a lower price and a fraction of the ops load.

Deployment review

Tell us what you’re running.

Octodec is available now as an assisted deployment — we size the cluster with you, not from a configurator. Share what you’re storing and someone here will connect with you.