Clustered file storage is enterprise-ready

Four trends are accelerating end-user adoption of clustered storage and file virtualization.

By Steve Norall

—Clustered storage is an idea whose time has finally come. For the past decade, the concept of a "pay-as-you-grow" resource with unlimited scalability has been an elusive vision of end users and vendors alike. However, with the events of the past five years, clustered storage has matured as a crop of nimble, innovative start-ups have proven that they are capable of delivering on many of the promises of the initial concept.

Clustered storage architectures have wide sweeping implications for all types of storage systems, including both block-based (Fibre Channel and iSCSI) and file-based (NFS and CIFS). At present, clustered storage architectures have largely gained traction in two separate markets—iSCSI and NAS.

This article focuses primarily on clustered file storage systems, a market represented by vendors such as BlueArc, Exanet, Ibrix, Isilon, Hewlett-Packard (via PolyServe), Network Appliance (via Spinnaker Networks), ONStor, Panasas, and Red Hat (via Sistina Software) that are pioneering the concept of clustered file storage.

Problems with monolithic NAS
Traditionally, major NAS vendors have adopted a scale-up approach to delivering performance and capacity. For their main filer lines, both Network Appliance and EMC use a single-head design connected to a set of physical disks. This monolithic approach means that all I/O to a given file system must be processed by a single head that resides in the data path. To scale, NAS vendors sold bigger, more-expensive monolithic NAS systems that could handle higher I/O rates and capacity. However, this scale-up model creates three core problems:

It's difficult to scale the data path: A traditional single-head NAS system can deliver at best 250MBps of I/O throughput to a single file system. However, with consolidated workloads and new distributed applications that require more than 250MBps of throughput, traditional NAS systems hit a performance wall. To scale performance further, data must be manually distributed across multiple filers. However, by doing this, administrators lose a key benefit: the ability to manage all file data under a single namespace. Therefore, to get the required performance, administrators must sacrifice manageability of their overall unstructured data environment—a jagged little pill for most to swallow.

High-capacity systems are costly and unwieldy to manage: Each filer has limits on how much capacity can be accessible by a single system. Therefore, if data grows faster than expected, administrators may need to do a "forklift upgrade" and purchase a larger, more-expensive NAS system to meet increased capacity requirements. This forklift upgrade approach not only causes downtime as the system is upgraded and data is migrated, but it also requires large up-front capital expenditures, instead of small, incremental capacity purchases.

No economies of management across filers: NAS filers are generally easy to use and operate. However, no matter how easy a system is to manage, as the number of filers proliferates in a data center, the amount of administration increases proportionately. Each filer introduces a separate namespace and set of file data that must be provisioned, protected, replicated, and maintained individually. Hence, a common user refrain is: "I loved my first filer, but hated my tenth."

Enter clustered storage
Clustered file storage addresses the drawbacks associated with monolithic NAS systems. Through the use of advanced file systems, clustered storage systems aggregate the performance and capacity of a set of disparate controllers and physical disks to create a single, scalable, fault-tolerant NAS resource. An outgrowth of this approach is that the system presents a single shared namespace accessible via common file access protocols, such as NFS, CIFS, and HTTP. The clustered storage system's namespace provides a single file and directory structure for all unstructured data to be centrally managed.

Clustered storage offers four key advantages over traditional monolithic storage:

  • I/O scalability: The ability not only to support the most I/O-intensive applications, but also to consolidate multiple workloads onto a single scalable resource;
  • Single point of management: Superior management through virtualization of physical nodes and storage resources under a single shared namespace;
  • Fault tolerance: The ability to survive multiple node, disk and network failures, while continuing to provide access to the stored content; and
  • Investment protection: The ability to grow the storage infrastructure and capacity over time, without the need to make large upfront purchases.

Four trends
A number of trends point to a bright future for clustered storage:

New classes of clustered computing architectures require the scalability of clustered storage. One of the more interesting developments in the last five years is the rapid rise of clustered and distributed applications. Traditionally, to scale a system, IT would elect to scale up by purchasing a larger, more-powerful server to deliver higher levels of computing horsepower. However, monolithic servers have given way to scale out approaches for many applications in the data center.

Examples of clustered computing architectures include Web applications, consolidated and virtualized server environments, and technical computing applications.

With the rise of the Internet, Web-based applications have replaced client/server as the dominant application architecture. Moreover, the Internet enables a level of scalability and reliability previously unimaginable. Storage systems must be able to handle thousands of concurrent users, expand seamlessly to support hundreds of online terabytes, and meet 24x7 uptime requirements.

Second, through the use of server virtualization tools such as VMware, IT has consolidated underutilized servers and their applications onto a centrally managed pool of servers running multiple virtual operating environments. As the scale of these virtual environment increases, the performance, availability, and cost-efficiency of the storage infrastructure become crucial to enabling the next level of scalability. And virtual server environments require low-latency, high-performance storage systems.

Third, technical computing applications have crossed over from scientific research organizations to become core mission-critical applications responsible for key business functions. These clustered computing applications loosely couple potentially thousands of cost-effective x86 servers together to amass tremendous amounts of computational performance. Examples of these applications abound across a wide variety of industries—econometric and Monte Carlo simulation clusters in financial services, blast clusters in biosciences, seismic processing applications in oil and gas exploration, video rendering in movies and TV production, and CAD/CAM applications in manufacturing. Most technical computing applications are voracious consumers of storage capacity and I/O throughput, making monolithic NAS systems a poor fit.

The common theme among these clustered computing applications is that they must share data and that the performance of the storage is a key to the overall responsiveness and throughput of the application. In short, monolithic storage systems cannot meet the high-throughput demands, fault tolerance, and seamless capacity scaling capabilities that these new class of applications require. As these distributed and clustered applications have gained a foothold in corporate data centers, they have caused IT to rethink their storage architectures. Vendors such as Isilon have ridden this wave effectively. For example, Isilon initially focused on applications such as video rendering and video production before branching out to oil and gas exploration, large-scale Web and media streaming, and digital archiving.

Consolidating file servers onto clustered storage delivers proven ROI: One out of five Windows servers shipped today is dedicated to file and print serving, and more than 50% of all NAS devices shipped run Windows. These servers and devices are proliferating at the department and workgroup level of large enterprises. As a result, enterprises are waking up to the need to gain visibility and tighter controls over these systems and their file data. File server consolidation has emerged as a way to reduce overall management and gain greater control over scattered unstructured data resources. End users that we have spoken with have reported ROI of between 30% and 50% after completing a file server consolidation initiative.

Clustered storage systems are excellent consolidation solutions because they provide a modular, scalable, fault-tolerant platform for collapsing many file servers into a single management point. However, not all clustered systems are well-suited for consolidation. Typically, the solutions that stand out have strong integration with Microsoft technologies such as Active Directory and Windows authentication, as well as a strong CIFS implementation. ONStor is an example of a vendor that meets these criteria, while blending file virtualization technologies such as a global namespace and virtual servers to help end users consolidate disparate file servers. For example, using ONStor's virtual server capabilities, each physical file server to be consolidated can be represented as a virtual server. As a result, clients are unaware that there have been any changes to the underlying file serving infrastructure, and administrators avoid namespace collisions (e.g., when multiple file servers with the same root directory name are collapsed together).

Clustered storage has been validated by Tier-1 vendors: Initially, small innovative start-ups pioneered the concept of clustered storage. However, over the past three years, larger vendors such as Network Appliance and Hewlett-Packard embraced the concept of clustered storage through high-profile acquisitions and OEM deals. These moves have given credence to the concept of clustered storage and signaled to end users that the technology is ready for enterprise deployment.

In 2003, Network Appliance became the first Tier-1 storage vendor to acquire advanced file system, namespace, and clustering technologies with its acquisition of Spinnaker Networks—an early pioneer in the clustered storage space. After a long and arduous integration process with its core WAFL file system, advanced software capabilities, and FAS hardware platforms, NetApp began shipping Data ONTAP GX in mid-2006—a clustered version of its Data ONTAP operating system.

HP realized it needed a game-changing approach if it aspired to dethrone EMC or Network Appliance in the enterprise NAS market. To that end, HP acquired PolyServe in April 2007 and used PolyServe's cluster file system to create a scalable NAS offering using standard HP ProLiant servers and EVA and XP storage systems.

However, despite these acquisitions, no Tier-1 storage vendor has yet to fully capitalize on the full value proposition that clustered storage enables and establish market leadership. Network Appliance appears to be doing a "soft roll" of the Spinnaker technologies for fear of cannibalizing its main filer business, and as a result has focused the Data ONTAP GX offering primarily on specific technical computing markets. Meanwhile, HP is still in the process of integrating PolyServe's technology with its existing products. The result is that even though several Tier-1 vendors have embraced the concept, a clear winner—or winners—has yet to emerge. Meanwhile, several of the smaller, nimble vendors are accelerating revenues and customer wins as the larger vendors struggle with internal issues.

Clustered storage is closing the enterprise feature set gap: One of the main drivers for purchasing a traditional single-head NAS system was that it came with a rich set of integrated data protection, storage management, disaster recovery, and compliance capabilities. Advanced storage management capabilities such as snapshots, quotas, asynchronous mirroring, thin provisioning, and write-once, read-many (WORM) functionality have become "must-have" capabilities for many enterprises. However, many of the early incarnations of clustered storage products did not have these capabilities. Therefore, despite their obvious architectural advantages, clustered storage was often eliminated from consideration for core data-center deployments.

However, the picture is now changing. Several clustered storage vendors have added these "must-have" enterprise-level features. Vendors such as BlueArc, Isilon, ONStor, and others offer many of the core data protection (e.g., snapshots) and disaster-recovery (e.g., asynchronous and synchronous replication) features expected in an enterprise NAS product. As clustered storage offerings gain feature parity with traditional NAS systems, end-user adoption will ramp rapidly.

Given the rapid innovation and conflicting product claims in the marketplace, it can be difficult for end users to determine which clustered storage solution is best for their environment. Our advice to users evaluating these solutions is to assess vendors and products on three key dimensions.

First, users should not adopt a "one-size-fits-all" approach to clustered storage. Users should recognize that the application and I/O workload should drive the selection of the appropriate clustered storage offering. Administrators should create I/O and data-access profiles for their applications and match them to the capabilities of clustered storage candidates. Different architectures are optimized for different I/O access patterns, amount of files, and file sizes. For example, some vendors have tuned their systems for large block, sequential I/O access, and large file sizes, while other solutions are better-suited for small block, random I/O patterns and small files.

Second, users should evaluate protocol requirements. Some NAS systems are best-suited for NFS data access, while others excel in CIFS performance.

Finally, if pursuing a file server consolidation project, users should evaluate vendors on the breadth and depth of their file virtualization capabilities. Specifically, users should look for products that have global namespaces that can be extended to include heterogeneous file servers and virtual server capabilities that prevent namespace collisions in a consolidation project.

With proper evaluation, users should be able to match their specific requirements to the features and functions of clustered file storage systems.

Steve Norall is a senior analyst with the Taneja Group research and consulting firm.

This article was originally published on August 17, 2007