Apeiron Storage Architecture
Scaling the PCIe bus to thousands of NVMe SSDs is just not practical. Apeiron Data Fabric paired with economical Ethernet technology enables seamless scalability to thousands of NVMe SSDs.
The distributed storage architecture independently scales compute nodes and storage, virtualizing any host connection with any storage connection within the data store switch fabric. Because compute nodes handle storage processing, virtual storage is integrated into the fabric, solving the scaling problem.
➡ Independent scaling of compute and storage
➡ Small cluster configuration entry point
➡ Low incremental cost to add compute or storage
➡ Optimal power and performance density
➡ High availability and fault tolerance
➡ Storage virtualization
➡ Easily extensible manageability
Apeiron Data Fabric
Apeiron’s Scale-out NVMe solution accelerates real-time applications while significantly consolidating datacenter infrastructure. The Apeiron Data Fabric and virtualization software optimizes compute and storage performance density at scale.
Storage is virtualized at the block level interface with software managing the physical mapping via policy-based user configurations. Compute and storage scale completely independent of one another.
NVMe over Ethernet (NoE)
Apeiron Data Fabric employs lightweight layer 2 Ethernet as the foundation for the NoE transport protocol. This innovative hardware-accelerated data path—along with the compute node’s storage virtualization software—provides ultra-low latency interconnect to networked NVMe SSDs. The entire roundtrip protocol overhead is less than 3 microseconds.
Apeiron’s compact Input Output Module (IOM) provides sixteen 40Gbps Ethernet connections using QSFP+ connectors. Each ADS1000 Array employs two IOMs providing 32 ports in a small 2U form factor. With each IOM supporting twelve standard NVMe SSDs through a SFF-8639 connector mid-plane, the hardware supports intra-IOM communication for high-availability management functionality and full-speed access to 12 additional NVMe SSDs. Application compute nodes enable Apeiron Data Fabric connectivity through the Host Bus Adapter, providing dual 40Gbps Ethernet connections over a QSFP+ connector. Multiple HBAs can be utilized by the application compute node maximizing performance and bandwidth trade-offs.
Highly Reconfigurable Hardware and Firmware Capabilities
The Intel® (Altera) FPGA connecting the Ethernet switching fabric and PCIe interconnect enables the platform to implement Apeiron’s NoE technology. Intel’s programmable hardware allows for optimal tuning and performance enhancements to support today’s workloads and tomorrow’s new data-driven applications.
The Apeiron basic storage building block, the ADS1000 array, is two rack units (2U). It offers optimal performance density and ultra-low latency.
Apeiron implements storage software processing and storage management on the application compute nodes. Applications and open software cluster management tools use the policy-based storage management for allocating storage resources for each of the compute nodes.
Apeiron’s NVMe management software provides the ability to virtualize the physical NVMe storage volumes. These virtual volumes can be configured to support different application workloads for optimal performance and efficiency. Additionally, users can enable advanced features for instant compute node failover, automatic replication, and compute node to compute node direct communication messaging. These advanced features are enabled through the ADS1000 configuration administration tools.
Instant storage failover provides the ability for a hot standby compute node to take over responsibility for storage stranded by a server that has failed. The application managing the cluster coordinates appropriate metadata and configuration changes. Automatic ADS1000 replication provides huge improvements for real-time applications and front-end cache deployments. The user can select the number and locality of replicas depending on application requirements.
Compute to compute communications can be used for directly accessing memory locations in under one micro-second, accelerating applications that have critical SLAs. Applications can take advantage of the Apeiron Data Fabric storage network for optimizing overall cluster performance and accelerating applications.
Easy Extensible Manageability and REST API
The Apeiron Storage Manager (ASM) provides extensive storage platform statistics and monitors overall health. It also allows users to set storage policies and configurations.
The ASM Command Line Interface provides for detailed inspection of specific modules and components. A browser-based Graphical User Interface (GUI) is available to monitor appliance operations. The REST API supports OpenStack/CinderBrick[i], along with other data center management tools. Additionally, a YARN plugin is supported for dynamically allocating NVMe storage. The Apeiron Storage Manager can be run on any networked server with Ethernet access to the ADS1000 platform.
High Performance Compute
Apeiron’s application compute nodes contain two Intel(R) Xeon E5-v3 processors (2x 14 cores, 2x 64GB RAM) in a 1U form factor. Two 10G network connections are paired with two or four 40GbE Data Fabric ports.
ADS1000 Direct Scale-out Flash Platform
The ADS1000 is a completely new storage platform. Capable of delivering ultra-efficient high speed networking to multiple racks of NVMe SSD arrays. When compared to typical scale-out/hyperconverged architectures, the ADS1000 provides up to 150% better CPU efficiency and up to 80% smaller data center footprint. Apeiron’s technology provides the ability to leverage the optimum SSD profile for your workload regardless of NVMe supplier. This includes full support for Intel’s® next generation 3D XPointTM technology today. The ADS1000 platform is simply the most efficient solution for any scale-out environment both today and in to the future.