A practical guide to Distributed Asynchronous Object Storage: what it is, how it works under the hood, where teams are deploying it, and how the Enakta Platform turns it into a production-ready enterprise product.
What Is DAOS Storage?
DAOS stands for Distributed Asynchronous Object Storage. It's an open-source storage engine built from scratch for NVMe flash and persistent memory. Unlike traditional storage systems that were designed around spinning disks and then retrofitted for SSDs, DAOS was written with modern hardware in mind from day one.
The project started at Intel, ran for over a decade of R&D, and is now maintained by the DAOS Foundation under the Linux Foundation. We co-founded the Foundation alongside Argonne National Laboratory, Google Cloud, Hewlett Packard Enterprise, and Intel. The codebase is fully open source and available on GitHub.
A quick note on naming: if you've seen "DAO" in the context of cryptocurrency and blockchain, that's something entirely different. A DAO in crypto is a "Decentralized Autonomous Organization." DAOS in storage is a purpose-built storage engine for high-performance computing and AI workloads. No blockchain, no tokens, no governance votes. Just fast storage.
Why does it matter?
Most enterprise storage systems were designed when hard drives were the only option. They route I/O through the Linux kernel, use dedicated metadata servers that become bottlenecks, and add software layers that made sense for 200 IOPS spinning disks but waste the millions of IOPS that modern NVMe drives can deliver.
DAOS takes a different approach. It bypasses the kernel, distributes metadata across every node, and talks directly to NVMe drives through user-space I/O. The result is storage that actually uses the hardware you paid for.
How DAOS Works
At a high level, DAOS has three architectural properties that set it apart from conventional distributed storage.
Architecture Overview
Architecture Stack
Four layers from application to hardware. Click any layer to explore.
⬡
Applications
POSIX · PyTorch · Custom Apps
Applications access storage through familiar interfaces: POSIX for legacy workloads, PyTorch DataLoader for AI training, or libdaos directly for maximum performance.
›
◈
Enakta Platform
SMB · NFS · S3 Gateways · Monitoring · Management
Enterprise layer: protocol gateways (SMB/NFS/S3), automated provisioning, real-time monitoring, alerting, and engineering support from the team that builds and benchmarks DAOS deployments.
The core storage engine. Bypasses the kernel entirely with SPDK for NVMe and Mercury RPC for networking. Metadata is distributed across all nodes with consistent hashing, no dedicated metadata servers.
›
▣
Hardware
NVMe SSDs · x86 Servers · Ethernet / RDMA
Standard commodity hardware. No custom controllers or vendor-specific silicon. KIOXIA and Micron NVMe drives on Xeon 6 servers with 100–400 GbE networking.
›
Data flows ↕ through each layer
The Enakta Platform sits above the DAOS core. It's not part of the open-source project — it's the layer that provisions DAOS, connects your applications to it via SMB, NFS, and S3, monitors it, and automates recovery when things go wrong.
User-space I/O
Traditional file systems go through the Linux kernel for every read and write. That kernel path adds context switches, copies, and latency. DAOS bypasses all of it. Applications talk to storage through a user-space library (libdaos) that communicates directly with NVMe drives using SPDK, and with the network using Mercury RPC over RDMA or TCP. No kernel overhead, no unnecessary copies.
Why It Matters
I/O Path: Traditional vs DAOS
Traditional storage adds 5 kernel layers between your app and the SSD. DAOS eliminates all of them.
Traditional
Application
↓
KERNELSystem Call
↓
KERNELVFS Layer
↓
KERNELFile System
↓
KERNELBlock Layer
↓
KERNELDevice Driver
↓
NVMe SSD
5 software layers · High latency
DAOS
Application
↓
libdaos
↓
SPDK / Mercury RPC
↓
NVMe SSD
2 software layers · Minimal latency
⚡
No kernel. No copies. No context switches.
SPDK talks directly to NVMe hardware. Mercury RPC handles networking over RDMA or TCP.
Distributed metadata
Legacy parallel file systems like Lustre and GPFS use dedicated metadata servers. These work fine at small scale, but they become chokepoints as the cluster grows. Every file open, every directory listing, every stat call hits the same servers.
DAOS distributes metadata across all storage nodes using a scalable object store with consistent hashing. There are no dedicated metadata servers. As you add nodes, metadata capacity grows linearly with the rest of the system.
Distributed Metadata
Centralized vs Distributed
Legacy systems bottleneck on a single metadata server. DAOS distributes metadata across every node.
Traditional — Centralized MDS
DAOS — Distributed Metadata
Single point of contention
Every stat, open, readdir hits the same server
Linear metadata scaling
Add nodes → add metadata capacity automatically
Self-healing and data protection
DAOS uses erasure coding and replication for data protection, and it rebuilds automatically when drives or servers fail. Rebuild happens in the background without taking the system offline, and it's fast because every surviving node participates in the rebuild. No hot spares sitting idle, no multi-hour rebuild windows.
Data Protection
Self-Healing in Action
When a drive or server fails, DAOS rebuilds automatically using all surviving nodes. Watch the cycle.
No downtime · No manual intervention · No hot spares
Where DAOS Storage Is Used
DAOS shows up wherever storage speed is the constraint, not the network, not the compute, but the storage layer itself. Here are the workloads where it makes the biggest difference.
AI & ML Training
Keep GPUs fed. Native PyTorch DataLoader, sub-millisecond random reads for training pipelines. The Enakta Platform handles provisioning and PyTorch integration so you connect data pipelines without custom storage code.
Media & Post-Production
4K/8K editing, VFX rendering. SMB access via the Enakta Platform means Windows and macOS tools connect without changes.
HPC & Simulation
CFD, molecular dynamics, climate modelling. Checkpoint/restart at full NVMe speed, with self-healing automation to keep runs running.
Enterprise NAS
Consolidate NAS sprawl. SMB, NFS, and S3 on one pool with self-healing.
Sovereign & Secure
Open source, on-prem, commodity hardware. No vendor lock-in, full data sovereignty.
Performance
DAOS isn't theoretically fast. It's been benchmarked and validated at production scale, repeatedly.
#1 #2 #3
IO500 10-Node Production at SC25. All three top spots.
270+ GB/s
Sustained reads and writes to a single client with RDMA
Linear
Scaling from 4 nodes to 1,000+ nodes
The IO500 is the standard benchmark for high-performance storage. At Supercomputing 2025 (SC25), DAOS-based systems swept the top three positions in the 10-node production category. These are independently validated benchmarks submitted by us and by our partners.
Our collaboration with Core42 demonstrated IO500 #3 on TCP networking, proving that DAOS delivers production-grade performance even without RDMA.
Performance scales linearly. Double the nodes, double the throughput. That holds from small 4-node clusters up through deployments with over a thousand nodes. No metadata bottleneck, no single point of contention.
IO500 Production List · Performance Comparison
Per-Client-Node Performance
Best normalized IO500 score per client node across all IO500 production submissions (ISC22–SC25). Normalizing by client count puts systems of different scale on equal footing.
Core42 × Enakta — DAOS on TCP onlyNo RDMA, no InfiniBand — standard TCP/IP networking. Still outperforms every non-DAOS filesystem per client node.
Source: io500.org production list (ISC22–SC25). IO500 score = geometric mean of bandwidth and metadata IOPS. Normalized per client node. Best entry per filesystem shown.
Research and Publications
DAOS has been the subject of serious academic and engineering research since 2016. If you want to go deeper than marketing pages and product sheets, here are 20 papers worth reading. Click any title to expand.
Architecture and Design
›DAOS: A Scale-Out High Performance Storage Stack for Storage Class Memory2020
Zhen Liang, Johann Lombardi, Mohamad Chaarawi, Michael Hennecke. SCFA 2020.
The foundational DAOS architecture paper. Covers the key-value storage interface, transactional non-blocking I/O, versioned data model, global snapshots, and full OS bypass. If you read one paper, read this one. Read paper →
›DAOS and Friends: A Proposal for an Exascale Storage System2016
Jay Lofstead (Sandia), Ivo Jimenez, Carlos Maltzahn (UC Santa Cruz), Quincey Koziol (LBNL), John Bent, Eric Barton. SC16.
The early blueprint for DAOS as part of the DOE Extreme-Scale Technology Acceleration (Fast Forward) project. Examines each layer of the proposed exascale storage stack. Read paper →
›Distributed Transaction and Self-healing System of DAOS2020
Zhen Liang, Yong Fan, Di Wang, Johann Lombardi. SMC 2020.
How DAOS handles distributed transactions with an optimized two-phase commit protocol that doesn't block when a coordinator fails. A good read if you care about fault tolerance internals. Read paper →
›DAOS Beyond Persistent Memory: Architecture and Initial Performance Results2023
Michael Hennecke, Jeff Olivier, Tom Nabarro et al. ISC High Performance 2023.
What happened after Intel discontinued Optane PMem. Describes the new architecture where the Versioning Object Store lives in DRAM with a synchronous Write-Ahead-Log and asynchronous checkpointing to NVMe. This is the architecture most new deployments use today. Read paper →
›High Performance Fabric Support in DAOS2024
Michael Hennecke, Alexander Oganezov, Jerome Soumagne et al. ISC High Performance 2024.
The networking deep-dive. Covers running DAOS over 100/200/400 Gbps Ethernet, InfiniBand, Slingshot, and Omni-Path, plus TCP for simpler deployments. Read paper →
Investigates how DAOS scales with server hardware and client-side parallelism. Confirms that bandwidth is practically identical whether you use a single shared file or one file per process, which is unusual for a distributed storage system. Read paper →
›Performance Comparison of DAOS and Lustre for Object Data Storage Approaches2022
IEEE/ACM International Parallel Data Systems Workshop (PDSW), 2022.
Head-to-head comparison of DAOS and Lustre for object-based data storage, examining how the object store model addresses scalability issues in traditional POSIX file systems. Read paper →
›Exploring DAOS Interfaces and Performance2024
SC24 Workshops (PDSW), 2024.
Explores the performance impact of choosing different I/O interfaces in DAOS, from the native object API to the FUSE module for POSIX compatibility. Useful if you're deciding which interface fits your workload. Read paper →
›Evaluating DAOS Storage on ARM64 Clients2023
Michael Hennecke, Motohiko Matsuda, Masahiro Nakao. HPC Asia 2023.
Shows DAOS is viable on ARM64 platforms, not just x86. Relevant if you're running Arm-based compute nodes. Read paper →
Aurora Exascale Deployment
›Initial Experiences with DAOS Object Storage on Aurora2024
R. Latham, R. B. Ross, P. Carns, S. Snyder et al. (Argonne). SC24 Workshops (PDSW), 2024.
First-hand operational experience running DAOS on Aurora, the largest DAOS deployment in the world. 1,024 storage servers, ~230 PB, 31+ TB/s peak throughput. Read paper →
›Aurora: Architecting Argonne's First Exascale Supercomputer2025
William E. Allcock et al. (109 authors). arXiv, 2025.
The full Aurora architecture paper. 2,048 DAOS engines, 16x 15.3 TB NVMe drives per server, ~260 PB raw. Used for training AuroraGPT (up to 1T params, 20T+ tokens). Read paper →
Weather and Climate
›DAOS as HPC Storage: A View from Numerical Weather Prediction2023
Nicolau Manubens, Tiago Quintino, Simon D. Smart et al. (ECMWF/EPCC). ISC High Performance 2023.
ECMWF's initial assessment of DAOS for weather prediction workloads. Bandwidth scales linearly with additional server nodes, which is exactly what operational weather forecasting needs. Read paper →
›Reducing the Impact of I/O Contention in NWP Workflows at Scale Using DAOS2024
Nicolau Manubens, Simon D. Smart et al. (ECMWF/EPCC). PASC 2024.
Demonstrates DAOS outperforming Lustre under high I/O contention typical of forecast workflows. Server-side contention resolution and byte-granular NVM access prove decisive advantages. Read paper →
HDF5 and I/O Middleware
›Accelerating HDF5 I/O for Exascale Using DAOS2021
Jerome Soumagne, Jordan Henderson et al. (The HDF Group). IEEE TPDS, 2021.
The journal paper on making HDF5 work with DAOS. Object creation time was drastically reduced and a new HDF5 map feature was enabled through the DAOS Key-Value interface. Essential reading if your workloads use HDF5. Read paper →
›Enhancing Metadata Transfer Efficiency: Unlocking the Potential of DAOS in the ADIOS Context2023
SC23 Workshops (PDSW), 2023.
Examines how to store ADIOS-level metadata in DAOS, comparing KV and Array interfaces against POSIX-based metadata storage. Relevant for simulation workflows that use ADIOS. Read paper →
Specialized Applications
›An RDMA-First Object Storage System with SmartNIC Offload2025
SC25 Workshops, 2025.
Offloads the DAOS client to an NVIDIA BlueField-3 SmartNIC, keeping the host CPU entirely off the I/O fast path. SmartNIC-offloaded performance matches host-based performance. Read paper →
›Leveraging DAOS File System for Seismic Data Storage2021
EAGE, 2021.
Uses graph theory and object-based storage to redesign the SEG-Y seismic data format natively on DAOS, accelerating data access for geoscience workflows. Read paper →
›Evaluation of HPC Application I/O on Object Storage Systems2018
Jialin Liu, Quincey Koziol et al. SC18 Workshops (PDSW-DISCS), 2018.
Compares DAOS, Ceph RADOS, and OpenStack Swift with three HPC applications. Finds object stores have better scalability than POSIX in many cases. Read paper →
Surveys and Broader Context
›Survey of Storage Systems in High Performance Computing2025
CCF Transactions on High Performance Computing, 2025.
Comprehensive survey covering the evolution of HPC storage from the 1990s through DAOS, Lustre, GPFS, BeeGFS, and others. Good for understanding where DAOS fits in the landscape. Read paper →
The DAOS project also maintains a curated publications list at daos.io/publications.
The Enakta Storage Platform
DAOS is the engine. The Enakta Storage Platform is what makes it production-ready for organisations that need enterprise storage, not a research project.
Running DAOS in production means dealing with deployment, monitoring, protocol integration, firmware management, lifecycle operations, and support. That's what we build on top of the DAOS core:
Protocol access: SMB, NFS, and S3 gateways so applications connect without modification
Managed deployment: automated provisioning and configuration management
Monitoring and alerting: real-time health dashboards, capacity planning, and proactive alerts
Self-healing: automatic rebuild on drive or node failure, no manual intervention needed
Reference architecture: validated hardware configurations with KIOXIA and Micron NVMe drives, Xeon 6, and 400GbE networking
Engineering support: from the team that builds and benchmarks DAOS deployments
You can explore cluster configurations yourself with the Enakta Configurator, an interactive tool that lets you size deployments, compare drive options, and see live performance estimates.
None of this exists in the open-source DAOS project. It's the gap between a powerful engine and a production system.
Platform Architecture
From Storage Engine to Production Platform
DAOS is the engine. Enakta is everything you need to run it in production.
ENAKTAEnterprise layer
↔
Protocol Gateways
SMB, NFS, and S3 — applications connect without modification
Distributed Asynchronous Object Storage. It started as a research project at Intel over a decade ago and is now maintained by the DAOS Foundation under the Linux Foundation. The founding members are Argonne National Laboratory, Enakta Labs, Google Cloud, Hewlett Packard Enterprise, and Intel. The full codebase is open source and available on GitHub.
Is DAOS the same as cryptocurrency or blockchain storage?
No. DAOS in storage stands for Distributed Asynchronous Object Storage. It has no connection to DAOs (Decentralized Autonomous Organizations) used in blockchain. No tokens, no governance votes, no blockchain. Just a high-performance storage engine for NVMe drives.
How is DAOS different from Lustre, GPFS, and other parallel file systems?
Three fundamental differences. First, DAOS bypasses the Linux kernel entirely with user-space I/O, removing the context switches and copies that limit traditional systems. Second, metadata is distributed across every node using consistent hashing instead of funneling through dedicated metadata servers that become bottlenecks. Third, DAOS was designed from scratch for NVMe flash, not retrofitted from spinning-disk assumptions. The result is that it outperforms conventional parallel filesystems per-server on the same hardware.
What kind of performance can DAOS deliver?
DAOS holds the top three positions in the IO500 10-Node Production list at SC25. We've demonstrated over 270 GB/s of sustained reads and writes to a single client using RDMA on commodity hardware. Our collaboration with Core42 proved that DAOS delivers production-grade performance even on TCP networking without RDMA. Performance scales linearly: double the nodes, double the throughput.
Do I need special hardware to run DAOS?
No. DAOS runs on standard x86 servers with NVMe SSDs and standard Ethernet networking. No custom controllers, no custom ASICs, no special network cards. Our reference architecture uses KIOXIA and Micron NVMe drives on Xeon 6 servers with 100-400 GbE networking. DAOS works with RDMA for maximum performance or TCP for simpler deployments.
Where is DAOS deployed today?
The largest deployment is Aurora at Argonne National Laboratory: 1,024 storage servers, roughly 230 PB, with 31+ TB/s peak throughput. It's being used to train AuroraGPT (up to 1T parameters, 20T+ tokens). Beyond HPC, teams use DAOS for AI training pipelines, media and post-production (4K/8K editing, VFX rendering), enterprise NAS consolidation, and sovereign on-prem deployments where data residency matters.
Does DAOS support standard protocols like SMB, NFS, and S3?
The open-source DAOS project provides POSIX access (via DFuse) and a native PyTorch DataLoader. The Enakta Storage Platform adds SMB, NFS, and S3 gateways on top, so Windows, macOS, and Linux clients can connect over standard protocols without any changes to their applications.
What is the Enakta Storage Platform?
It's enterprise storage built on the DAOS engine. We built managed deployment, protocol gateways (SMB, NFS, S3), real-time monitoring, self-healing operations, and engineering support that don't exist in the open-source DAOS project. Think of DAOS as the engine and Enakta as everything you need to run it in production. You can explore cluster configurations with the Enakta Configurator.
Can I run DAOS without the Enakta Platform?
Yes. DAOS is fully open-source and you can run it yourself. What you won't have is SMB, NFS, or S3 access, automated provisioning, monitoring dashboards, self-healing orchestration, or someone to call when things go wrong. The Enakta Platform is the production layer that makes DAOS viable for teams without dedicated storage engineering staff.
Ready to talk about DAOS storage for your workload?
Whether you're evaluating storage for AI training, media production, or HPC, we can help you size a deployment and understand the performance you'd get.