Introduction
For technical architects designing high performance, high capacity storage, the data path is critical. Every network hop and processing step introduces latencies that compound under heavy concurrency.
With the release of the Nutanix Objects Storage 5.4, the engineering team here at Nutanix has introduced a major architectural paradigm shift for greenfield Nutanix Objects Storage deployments on the AHV hypervisor; this is known as the "Direct Path Architecture" architecture. By systematically excising the Layer 4 Load Balancer from the data path, we have fundamentally reimagined client-to-storage networking into something much more streamlined.
This is not merely an optimization; it is a structural teardown that resolves front-end bottlenecks, can substantially improve performance, and lays the vital groundwork for next-generation hardware offloading. Let's get into how the Direct Path Architecture topology works and what it unlocks for the future of enterprise object storage.
The Architectural Shift: Flattening the Data Path
In pre-5.4 architectures, Nutanix Objects Storage relied on integrated load balancer VMs to distribute S3 client traffic across the Nutanix Objects Storage virtual machines (known as "worker VMs"). While robust, this introduced the problem of multiple hops: client requests hit a load balancer and were then forwarded to a worker VM before hitting the backend storage nodes.
This architecture effectively capped front-end throughput and concurrent connections per object store, creating an artificial ceiling for high-throughput workloads. Furthermore, acting as an L4 proxy, the load balancers masked the true client IP from the object controllers running in the worker VMs, complicating security and auditing.
The Direct Path Architecture Implementation
The new architecture surgically removes this middle tier. Instead of pushing traffic through the load balancer stack, Nutanix utilizes Multus-CNI to attach a second, unmanaged network interface card directly to the core object storage services running on the worker VMs. This direct connection to the client network avoids the intermediate routing steps and processing overhead previously incurred by the dedicated load balancer tier.
The core object storage services are now fully autonomous in managing their network state, independently handling tasks such as NIC configuration, stateful firewalling, distributed failure detection, failover, and failback mechanisms. To allow these autonomous services to operate seamlessly, network configurations are pushed programmatically via APIs directly from the Nutanix management plane to the storage services. This approach allows new network settings and policies to be applied dynamically across the cluster, ensuring the environment remains agile and responsive to changing requirements in a fully seamless manner.
Quantifying the Impact: Performance and Resource Efficiency
By flattening the network topology, Nutanix reduces "network amplification", the ratio of internal network traffic generated per client request. For technical architects sizing clusters, this translates to immediate efficiency gains:
- Reclaiming Compute: Removing the load balancer frees up 2 CPU and 4 GiB of RAM (or up to 4 CPU and 6 GiB depending on the load balancer size) per physical node. At scale, this results in a 16.67% to 20% lower CPU footprint for the entire Nutanix Objects Storage cluster i.e., dropping a 12 CPU footprint down to 10 CPU, or in the case of high performance deployments, 20 CPUs down to 16.
- Reduced Network Amplification: By eliminating the extra network hop, bandwidth-constrained environments will see theoretical efficiency improvements of up to 25% for write workloads (in RF2 deployments), as the total network transits drop from four hops to three. Read workloads see up to a 33.3% improvement, as the read path is shortened from three hops down to just two.
- Real-World Throughput and Latency: The true value of the Direct Path architecture becomes apparent under heavy concurrency. Internal testing confirmed improvements in both throughput and latency compared to the previous architecture (with load balancers), delivering a more responsive storage fabric for demanding workloads. These performance tests were carried out on a Nutanix Objects Storage cluster consisting of 4 NX-8170-G9 nodes each with 12 NVMe drives. Connectivity was over 100GbE.
Non-Disruptive Operations and Scale-Out Agility
Because the core storage services now manage networking natively rather than relying on static external load balancer configurations, managing the cluster's network lifecycle becomes much more fluid. This architecture directly unblocks several capabilities for enterprise architects:
- Nondisruptive Network Changes: Administrators can now push updates directly to the cluster without requiring maintenance windows or traffic disruptions.
- Network Capacity Scale-Out: The architecture supports dynamic client network capacity scaling; every new worker VM deployed to the object store adds additional network bandwidth on the client side.
Architectural Note: With the dedicated load balancers removed, we leverage DNS round-robin with low Time-To-Live (TTL) values to efficiently distribute client traffic directly to the worker VMs.
Next-Generation Workloads
The true architectural beauty of the Direct Path Architecture isn't just what it fixes today, but what it enables tomorrow. By flattening the data path, Nutanix has laid a foundation for future enhancements. Stay tuned for our future releases.
Conclusion
The Direct Path Architecture in Nutanix Objects Storage 5.4 represents a fundamental maturation of the platform. By trusting the core storage services to manage their own direct network connections, Nutanix has bypassed traditional cluster network bottlenecks. For technical architects, this means delivering a storage fabric with lower overhead, increased concurrent throughput, and a data path fully prepared for the next generation of enterprise object storage.
©2026 Nutanix, Inc. All rights reserved. Nutanix, the Nutanix logo and all Nutanix product and service names mentioned are registered trademarks or trademarks of Nutanix, Inc. in the United States and other countries. All other brand names mentioned are for identification purposes only and may be the trademarks of their respective holder(s). This content reflects an experiment in a test environment. Results, benefits, savings, or other outcomes described depend on a variety of factors including use case, individual requirements, and operating environments, and this publication should not be construed as a promise or obligation to deliver specific outcomes.
