cloudparts

DAOS

Distributed Asynchronous Object Storage: an all-flash storage system for supercomputers and AI clusters that runs entirely in user space and serves a parallel file system on top of its own key-value object store.

DAOS Foundation project
What it is
File storage
Type
Open source
Made by
DAOS Foundation (Linux Foundation), started at Intel
Foundation
DAOS Foundation, joined November 2023. Part of the Linux Foundation
License
BSD-2-Clause Plus Patent
Facts checked
28 October 2026
Website Source code

What it is

DAOS was built for the storage of supercomputers, where thousands of nodes write checkpoints and read training data at once and the locks and metadata servers of a POSIX file system become the bottleneck. Its engines run end to end in user space: they reach NVMe SSDs through SPDK and the network through libfabric (TCP or InfiniBand verbs, with UCX as an alternative), bypassing the kernel. Storage is carved into pools, a reservation of capacity across many servers, and each pool holds containers, separate object address spaces that are snapshotted and managed as a unit. Inside a container, objects are addressed by a distribution key, which decides the server, and an attribute key, under which sits a single value or an array. Each object’s class sets its protection: replication such as 1×4, or erasure coding such as 4+2. Pool metadata is replicated across a few nodes in separate fault domains and kept consistent by consensus, and lost redundancy is rebuilt automatically.

Applications rarely use the object API directly. A container created with --type=POSIX holds a file system, which programs reach through libdfs, through a FUSE mount (dfuse) that lets unmodified programs see ordinary files, or through interception libraries loaded with LD_PRELOAD that skip the kernel for reads, writes and, with libpil4dfs, metadata. MPI-IO, HDF5, a Hadoop and Spark connector, TensorFlow-IO and Python dictionaries sit on the same objects. DAOS was designed around Intel Optane persistent memory for metadata; after Intel cancelled Optane in 2022 it gained a Metadata-on-SSD mode for ordinary servers, which DAOS 2.8 (August 2026, the first release managed by the DAOS Foundation) declares production ready. It runs on x86-64 servers with EL 9 or SLES 15, and Argonne’s Aurora supercomputer uses it as its main storage.

Why teams pick it

  • Very high parallel throughput: Argonne’s Aurora system topped the IO500 production list in 2023 and again at SC24, where it scored 32,165 against 2,509 for the next system, also running DAOS
  • No distributed locks and no kernel in the storage servers’ data path: clients talk to user-space engines over RDMA, and interception libraries keep the client’s kernel out of reads and writes too
  • One store, several interfaces: POSIX files, MPI-IO, HDF5, Spark and key-value or array objects all live in the same pools
  • Replication or erasure coding chosen per object class, with automatic rebuild after a server or drive fails
  • BSD-licensed and governed by a neutral foundation whose members include Argonne, Google, HPE and Intel

Watch out for

  • Not fully POSIX: no flock, hard links or quotas, O_APPEND from several mounts can corrupt a file, and dfuse caches metadata and data by default, which can surprise parallel jobs
  • Narrow platform support: x86-64 only, EL 9 or SLES 15, and best with an RDMA network; ARM clients work but are not tested in CI
  • Missing features many enterprise file systems have: no encryption at rest or in flight, no compression or deduplication, and no rolling upgrades
  • Small community after Intel’s exit: Intel’s last release was 2.6 in July 2024, and development now rests mainly on HPE, Google and Enakta Labs

Alternatives

What all file storage have in common →Compare side by side →
CephOpen sourceA distributed storage cluster that serves S3 buckets, block devices and a file system from the same pool of disks.Amazon EFSAWSA managed NFS file system that grows on its own and mounts on many AWS machines at once.Apache HadoopOpen sourceThe project that made big-data clusters normal: a distributed file system (HDFS), a cluster scheduler (YARN) and the MapReduce engine.

How it fits with other technologies

Create a pool and a POSIX container, mount it on a compute node and use it · bash

# on an admin node: reserve 50 GB across the storage servers
dmg pool create --size 50GB tank

# a container holding a POSIX file system
daos cont create tank scratch --type=POSIX

# on every compute node: mount it with FUSE
dfuse -m /mnt/scratch tank scratch
ls /mnt/scratch

# bypass the kernel for reads, writes and metadata calls
LD_PRELOAD=/usr/lib64/libpil4dfs.so cp -r ./inputs /mnt/scratch/

fusermount3 -u /mnt/scratch

See it in a real stack

No architecture uses DAOS yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2012–2026

DAOS started in 2012 as research inside Intel’s high-performance storage group, which had come with the purchase of Whamcloud, the company behind the Lustre file system, and was funded in part by the US Department of Energy’s FastForward programme for future exascale machines. The idea was to stop forcing every workload through POSIX: build a lock-free, versioned object store on persistent memory and flash, and put file systems and scientific I/O libraries on top of it as layers.

The code was published on GitHub in 2016, and DAOS became the storage system of Aurora, the exascale supercomputer that Intel and HPE built for Argonne National Laboratory. Its design leaned on Intel Optane persistent memory for metadata, so when Intel cancelled Optane in 2022 the project re-architected the metadata path to log to NVMe SSDs instead.

In November 2023 Argonne, Enakta Labs, Google Cloud, HPE and Intel launched the DAOS Foundation at the Linux Foundation, and Intel donated the code. Intel’s last release was 2.6 in July 2024. Since then HPE, whose Johann Lombardi chairs the technical steering committee, has led development; Google runs its Parallelstore service on DAOS, and the foundation released 2.8 in August 2026.

Why it’s built this way

Objects first, POSIX as a layer
POSIX’s locking and shared metadata limit parallel I/O at supercomputer scale, so DAOS offers a key-value object model and builds the file system, MPI-IO and HDF5 on top of it.
Everything in user space
Flash and RDMA networks answer in microseconds, so DAOS reaches NVMe through SPDK and the network through libfabric and keeps the operating system out of every I/O.
Versioned writes instead of locks
Every update is tagged with an epoch and kept as a new version, so DAOS needs no distributed locks and a snapshot is just a reference to an epoch.
Metadata on SSD after Optane
With Optane discontinued, DAOS keeps metadata in DRAM and persists a write-ahead log to NVMe, so it runs on ordinary servers.

Timeline

  1. 2012

    Research starts at Intel under the DOE FastForward programme

  2. 2016

    Code published on GitHub

  3. 2022

    Intel cancels Optane; work starts on Metadata-on-SSD

  4. 2023

    DAOS Foundation launched on November 9; Aurora tops the IO500 production list

  5. 2024

    Version 2.6 in July, Intel’s last release

  6. 2026

    Version 2.8 in August, the first foundation release, with production-ready Metadata-on-SSD

Sources13 · facts checked 28 October 2026

Improve this page on GitHub