Clustering challenge

Clustered storage is emerging as a potential answer to performance, scalability and reliability issues, but it doesn’t come cheap and may prompt confusion. Eliot Beer reports.

  • E-Mail
By  Eliot Beer Published  October 23, 2005

|~|Jaguros,-Ch_m.jpg|~|“A lot of users think they don’t need DR if they have clustered storage, which is definitely not the case.” - Chris Jaguros, ASP Gulf.|~|When it comes to storage, companies have to balance capacity and security, performance and investment protection. Finding a solution which satisfies all of these requirements is impossible, but striking a better balance is getting easier thanks to developments in clustered storage. But clustered storage is not a clear technology; almost every vendor has its own interpretation of the term. There are, however, a few basic principles they do agree on, and these lie at the heart of what makes clustering storage devices so interesting. The idea originated after clustered servers became an accepted technique for improving the efficiency of networked applications, and the thoughts of engineers turned to applying the same principle to storage. Essentially, all varieties of clustered storage attempt to tackle two main problems: elimination of a single point of failure (SPOF) and enhancing performance when capacity is scaled up. They aim to achieve this by increasing the interconnectivity between storage devices, storage controllers and the network using high-bandwidth links. Beyond this, though, the various implementations are very different. The first variety of clustered storage is the most uncompromising, and as such is highly expensive and suitable only for situations where performance and scalability are beyond cost considerations. In this scenario, the cluster is made up of self-contained modules, with each module not only containing storage devices, but also a controller, a network connection, and a full complement of memory and processor power. This means the modules can share the load of user requests, both in terms of processing and bandwidth to the network, between each other. The cluster can also tolerate the failure of a single module very easily and can incorporate new modules, sometimes within only a few minutes. The key to this is very complex engineering at the module and infrastructure level, resulting in very simple operation from the end user’s point of view. This type of system is reportedly excellent for processing large, contiguous files. For instance, US magazine Sports Illustrated uses an Isilon cluster to handle large photo shoots such as the Olympics or the American football Superbowl competition. Because this system is essentially a very advanced network attached storage (NAS) device, it is more suitable for file system applications than database applications, according to vendors. Many in the Middle East are sceptical about the application for such systems in the region. Gerard Van Antwerpen, data management architect at Sun Microsystems, says companies here do not yet need this level of technology, both in terms of their storage requirements and the cost. “Users here still haven’t fully realised the potential of technologies such as SAN (storage area network), many are still using NAS or storage island solutions,” he says. “Storage clustering is the next level of storage, but the requirement isn’t here in the Middle East yet.” The second approach to storage clusters is that taken by NetApp. In many ways it represents a halfway-house between full-scale hardware clustering and software approaches to the issue. Whereas the clustering system described above does everything through hardware, NetApp uses a dedicated head unit, using standard processors and memory, to perform all the RAID and clustering functions through software. “When you look at hardware clusters, they are extremely complex internally,” says Gavin Keeler, regional director for MENA at NetApp. “We take the approach that software is less likely to fail than hardware, so we do as much as possible with software in the head unit. We can then use two head units and cluster them with two storage devices, so everything is redundant.” Keeler claims this has a big advantage over full hardware clustering. Because the system is less complex and reduces the number of stages information must pass through, NetApp’s clusters have no inherent disadvantage when dealing with databases or other applications which require frequent, small accesses to storage. However, he says this option still comes at a premium, in part simply because of the extra connections needed between the various components. One recent NetApp implementation is running at PWC Logistics in Kuwait. Muntaser Ayyash, IT operations manager at PWC, says the company’s NetApp FAS (fabric attached storage) 920c system is able to handle both database applications, file sharing and home directories, and can prioritise the high-availability applications. He also says PWC experienced significant improvements in performance with the new system. “It was like day and night, compared with what we had before. We noticed huge performance gains with Oracle, SQL and Exchange applications; with Exchange sometimes performance was 10 times better. The NetApp system will also let us expand very easily, and we are looking at making our business centres actually drive the storage — we can just switch it on as required,” says Ayyash. The final category of clustered storage is all-software clustering, as used in Red Hat Linux’s Global File System (GFS). This is very similar to virtualisation in many ways, but importantly it usually uses a dedicated set of storage devices utilising RAID systems, and so is suitable for high-availability applications such as core databases, as well as large-file based applications. Dusan Djordjevic, senior Linux consultant with Opennet MEA, a Red Hat master distributor, says, “GFS can be used for most types of application. For example, we’ve implemented GFS for both Emirates Media in Abu Dhabi and Dubai TV. Both of these companies use very big files, and they needed to be able to process them quickly — GFS allows this. In addition we’ve just finished a Red Hat implementation for the Dubai Department for Economic Development (DED) for their core Oracle 9i RAC database.” DED’s IT manager, Sultan Ali Lootah, explains the benefits of GFS for the organisation: “First, it is more secure and more reliable than our previous system, and the support costs are only around 20% of what we were paying before. The performance of the system is also better than before, and we don’t expect any problems with performance if we have to scale up the storage.” Lootah also says DED wanted to free itself from having to get support from a hardware vendor, as the vendor for its last system recently told DED the system would be obsolete within a year, after which there would be no spare parts and a lower level of support. He says he expects the Red Hat system to last 10 to 15 years, and much of the maintenance will be done in-house. This is something both the Red Hat and NetApp approaches have in common. Ayyash says a major factor in PWC choosing NetApp was the ability to do in-house maintenance, instead of having a vendor drive the support. And this is also an obvious factor when comparing the full-hardware clustering discussed above to the software-based approaches. But all of this can come at a high initial cost. Both DED and PWC had obsolete systems, and took the opportunity to replace storage hardware in favour of their new systems. But if companies are not in a position to do this, then introducing a clustered solution could prove both complex and expensive. Graham Porter, marketing manager for Sun, says, “At the moment, we are seeing customers with increasingly complex storage requirements, but not increasing budgets. Companies, especially financial institutions, are facing new compliance rules, information lifecycle issues, and increased storage requirements from growth. If a vendor were to say ‘I can make your storage super-reliable, but the cost will more than double’, I don’t think there’d be a market for that.” Chris Jaguros, general manager for ASP Gulf Dubai, agrees. His company is a storage service provider and claims to be technology-neutral, offering the best-fit solution for a particular customer. He says a lot of customers are initially attracted by the idea of clustered storage, but often misunderstand its application, and underestimate the cost. “Often, when we outline the various options at the consultation stage and explain the concept of clustered storage, customers will say ‘Oh, yes, that’s what we want’. Then when we come to outline the costs, the same customer will say ‘Maybe RAID is ok, we don’t need clustering’. Also, there is some confusion with clustered storage and disaster recovery (DR) — a lot of users think they don’t need DR if they have clustered storage, which is definitely not the case,” says Jaguros. Several vendors make this point, and it is important to emphasise. While there are clustered storage solutions which can work across multiple sites, these use highly expensive connections which must be physically laid between the sites, and are practical only for users such as major banks, governments and defence departments. A typical clustered storage implementation will be single-site, and if a user decides to adopt a DR strategy, clustered storage should be a separate consideration. “There’s no point having a storage cluster at one site, and expect it to cover DR as well,” says NetApp’s Keeler. “At the same time, it’s also not usually appropriate to use a cluster at a DR site, unless business continuance is worth the cost of running two clusters, one at each site. We usually recommend a single-head system for a DR site.” This example is indicative of the confusion which still surrounds clustered storage. When asked, many vendors were unclear as to what the term meant, some confusing it with clustered servers, and others talking about tiered storage. Part of this confusion is explained by the lack of appetite for clustered storage as promoted mainly by American firms such as LeftHand and Isilon, discussed above. But this may change, as demand for ever more powerful systems grows. David-Alan White, solutions architect for eLinear Solutions, sees clustered storage as being tied into clustered servers, and, ultimately, grid computing. He regards this as the future for high-performance computing and sees applications in a number of sectors, including the medical, media and oil & gas industries. “We’re seeing increasing demand for high-performance systems, and they’re becoming more deliverable with cheaper components and open-source software such as Linux. As far as clustered storage goes, I think this will grow, and there’ll be more use of the Linux Beowulf system [a similar system to Red Hat’s GFS]. I think the Middle East will start to take up this technology in the very near future, but I don’t think there are many applications for expensive hardware systems in the region,” says White. The message, then, from vendors and users alike, is there is some demand for clustered storage, particularly for software models based around Linux. But this is tempered by a degree of confusion about the technology and the high implementation costs it can potentially bring. But as companies in the Middle East continue to experience exponential growth, it is clear the market can only grow, especially as costs are almost certain to come down. Sun’s Van Antwerpen agrees, and says, “If you go back 10 years, the big debate was between direct-attached storage (DAS) and NAS. Then, NAS was massively more expensive, now it’s almost the same price. The same thing will happen with clustered storage, it will just take time.”||**||

Add a Comment

Your display name This field is mandatory

Your e-mail address This field is mandatory (Your e-mail address won't be published)

Security code