Thoughtput

I/O Bottlenecks In The Limelight

Forbes recently interviewed the CEO of Limelight Networks, a powerful player in the booming content delivery business. When faced with the "bottleneck" question about the greatest download pain points between host to consumer, Lunsford identifies disks within the data centers as the primary culprit for delays in the food chain.

From host to consumer, where's the biggest bottleneck on a download?

Disk input-output time at the origin servers, which are where the content is originally being fed from.....We have plenty of servers. We have plenty of bandwidth. It really gets down to pulling that thing off the disk.

- Jeffrey W. Lunsford, CEO, Limelight as quoted in Forbes

Content delivery, particularly the need to quickly send media to an ever increasing audience, is back now that consumers have more broadband coverage. Once dominated almost exclusively by Akamai, other players such as Limelight, Vitalstream (acquired by Internap Network Services) and Level 3 who recently purchased the content delivery network of Savvis, are showcasing the power of matching content with media hungry consumers.

Of particular interest in Lunsford's description is the focus on fewer more powerful points-of-presence (POP), compared to his competition. Why replicate data in hundreds of locations around the world if you can have a dozen or two locations with the bandwidth, capacity, and I/O horsepower to fill the pipes and saturate consumer demand.

This will be an exciting space to watch. Is it better to have more POPs with redundant data, or fewer POPs that can fill the same need? Consolidation trends favor fewer POPs with enough scalability and performance to meet the need. And nothing helps reduce the disk input-output bottleneck like a dose of high-speed, high-capacity caching!

Storage Performance Transcript

SearchStorageChannel.com has posted the transcript of our podcast with Tony Asaro of Enterprise Strategy Group. The podcast itself is worth a listen, but if you prefer to read the details, point your browser here:

Transcript of Storage Performance Podcast with Tony Asaro of Enterprise Strategy Group

Enjoy!

Balancing capacity and performance

The impact of the server-storage performance gap showing up in the Oracle database market. The Independent Oracle Users Group (IOUG) recently released a survey of its members titled Managing the Storage Equation. While part of the survey details the need for ready access to additional capacity, comments by some of the members indicate that simply adding capacity won't sustain the current demands. Here is one comment from a recent Computerworld article covering the survey: 

The Chicago Mercantile Exchange faces an even bigger problem, though: The performance of disk drives isn’t increasing fast enough to meet its processing requirements, Kulesa said. “Balancing growing capacity needs with increasing performance demands has been the real challenge we’re facing,” he added.

Rarely can capacity be evaluated without performance considerations, and vice versa. How many Oracle users are truly short of capacity, or perhaps just short of capacity that can suit the needs of today's demanding applications? If capacity is a constraint, one consideration could be to use SATA drives (typically cheaper and with higher capacities that Fibre Channel) and a centralized caching resource for performance. We covered this in an earlier post, Very SATAsfying.

A Tale of New Choices

The NAS market has continued to march forward with functionality and performance. 2006 proved to be a banner year particularly in terms of new performance achievements for NAS and the expansion of the number of market players.

But the drive to improve the performance of NAS configurations has left customers with less than optimal choices...until now.

Many customers enjoy their existing NAS installation, have comfort with the current operating system and management tools, and simply need to reduce latency and increase performance. They do not necessarily need to re-architect their entire configuration. Unfortunately, that has been the primary option to date.

Now new solutions based on centralized storage caching will provide customers with all the performance necessary, but more importantly will allow customers to retain existing storage solutions. This helps save money, reduce risk, and speed time to deployment.

In the following embedded presentation, A Tale of New Choices, we take a closer look at the details and offer a new way to tackle NAS performance issues with centralized storage caching.

Hybrid Drives

Consumer and enterprise markets are separate, but often follow similar trends. The x86 platform launched years ago as a personal computer is the ancestor of today's rack and stack data centers of hundreds or thousands of servers.

This week marked a major announcement from the Hybrid Storage Alliance. This group is deftly working around the inherent limitations of mechanical disks by incorporating flash memory into laptop drives to increase performance, reliability, and durability all while reducing power consumption. Hmmm. Sounds like a great way to solve the Server-Storage Performance Gap.

Except that flash memory isn't suited for enterprise environments. That flash drive on your iPod Nano would choke in a matter of minutes if it had to run a transactional database.

But RAM can be put to good use in the data center to compensate for poor storage performance.

Focusing on benefits, complementing disk drives with flash memory at the notebook level is analogous to complementing enterprise storage systems with centralized storage caching a the data center level. For example:

• Fast performance: while the Alliance promotes faster boot-up and resume times, the enterprise complement of caching means significantly greater I/O and application performance
• Lower power consumption: if you think saving some battery time on your laptop is a desire, how about reducing the enormous power drain on keeping hundreds of terabytes of disk spindles. Solutions that restore performance with the right balance of RAM and less over-provisioned disk means far lower power requirements.
• Higher reliability: As the last moving mechanical piece in the computer, drives are known to fail in notebooks, and enterprise configurations alike. Reducing the disk load by offsetting traffic with flash or RAM helps extend disk life. Users who manage hundreds of spindles in their data center are always looking for ways to simplify their environments. See our post: Drive Reduction.

The Hybrid Storage Alliance is right on time. And the emergence of memory-based, heterogeneous caching solutions for the data center is an equivalent breakthrough.

Making things TIERific

Not all storage is created equal. That maxim, and the truth it holds, led many vendors and customers to develop architectures to support multiple tiers of storage.

One of the simplest breakdowns to conceptualize, and a relatively easy structure to implement, is a tiered storage strategy based on disk type. For example, you could set up a performance driven tier with fast Fibre Channel drives, and then a capacity centric tier with SATA drives. Why put all of your data on expensive drives when much of it can be offloaded to less expensive, higher-capacity, SATA drives?

But while the tiering storage is easy to think about, it can be harder to manage than initially perceived. This is particularly true when performance considerations are in play, and when the number of tiers starts to require more attention than the problems they were initially designed to address. A recent quote (see graphic) from an article in Storage Magazine, Users want more from tiered storage, makes this abundtantly clear.

It seems that users want the benefits of tiering but not if it includes the burden of excessive management attention.

Another way to think about managing performance across a storage infrastructure is to break out of conventional thinking and focus on memory-based complements. As outlined in Very SATAsfying, it is easy to set up one simple, high-capacity, low-cost storage tier, and then complement that tier with perfomrance based caching solutions. This delivers all of the performance needed, and removes the need to manage multiple tiers, especially across controllers which can become a pain to administer.

Tiering is helpful, but having the minimum number of tiers supporting the most use cases is even more helpful. And getting to a solution that offers a single cost-effective, high-capacacity storage tier, and then a performance caching complement, makes an ideal solution.

Moving to More Memory

We saw an increasing awareness of the Server-Storage Performance Gap throughout 2006, and a need to address this issue with memory based alternatives. Gone are the days where disks alone can deliver the performance and efficiency today's demanding application environments.

A great piece titled Are RAM Blades in Your Future? accurately highlights this thinking and the trend underway. In particular, the author describes how using a large, centrally available pool of memory can help with
server virtualization, an area that has potential pitfalls if not addressed. See Virtual I/O Trap? from our previous post.

The author specifically points out how a shared network-based memory resource can help with:

• Higher Utilization
• Flexibility
• Lower Cost
• Component Simplification

Check out the full article for all the details. And take a look at an explanation of centralized storage caching to see how to deploy a similar architecture today.