Disk Storage

Disk Storage

The spinning disk is inevitably going to be the choke point for application performance in many server environments, as information is exchanged between the server bus and magnetized areas on a spinning platter. Typical storage deployments have changed radically over the past 30 years, starting with consolidated storage deployments (mainframe centric) toward distributed computing with captive storage (onset of open systems computing) and back to a consolidated storage deployment model (open systems with storage networking).

During the rapid proliferation of distributed servers with captive storage, each node had its own directly attached storage capacity, also known as direct-attached storage (DAS). With the onset of storage area networks (SANs), which enable consolidation of storage capacity onto a storage array device that can be logically provisioned and shared through a networking fabric, many organizations find themselves collapsing storage infrastructure once again to simplify deployment, control costs, and improve utilization and efficiency.

DAS can be as simple as a single disk within the server chassis or multiple disks within a dedicated unintelligent external chassis (just a bunch of disks, or JBOD), or as complex as a directly attached dedicated array using Redundant Array of Independent Disks (RAID) technology. With DAS, storage is completely captive to the server it is connected to and can be connected to that server using a number of storage interconnects, including Small Computer Systems Interface (SCSI), Serial Attached SCSI (SAS), serial advanced technology attachment (SATA), or Fibre Channel.

SANs provide logically the same function as DAS with the exception that the disk (physical or virtual) being accessed is attached to a network and potentially behind a controller that is managing disks that are accessible to other nodes on that network. With SAN-attached storage, disk capacity appears to the server as DAS but is managed as a provisioned amount of capacity on a shared array, accessible through a high-speed network such as Fibre Channel.

Network-attached storage (NAS) is similar to SAN in that storage capacity is made accessible via a network. NAS, however, does not provide a server with a physical or logical disk to perform direct operations against. With NAS, a file system is made accessible via the network, and not a physical or logical disk. This means that accessing information on the shared file system requires that it be accessed through a file system protocol such as CIFS or NFS.

Each of these types of storage interconnect have strengths and weaknesses. For instance, with DAS, the upfront cost is minimal, but longer-term management costs are extremely high, especially in dynamic environments where capacity may need to be repurposed. SAN, on the other hand, has a very high upfront cost but far lower long-term costs in terms of management and data protection. NAS generally has a lower initial cost than SAN but is commonly more expensive than DAS. As clients that access NAS for storage pool access utilize network file system protocols such as CIFS and NFS, NAS is generally less applicable in certain application environments. For instance, in database application environments where the server attempts to leverage its own file system or make direct calls against a physical volume, NAS may not be the best fit due to the abstraction of the file system protocol that must be used to access the capacity.

So how does this impact application performance? Simple. The slowest component in the path of an application is typically the spinning disk on the client and the server. If the storage subsystem is adequately sized and configured, application performance will still be dependent on the speed of the rotating disk, but the characteristics of how data is accessed may be changed to improve performance. For instance, some levels of RAID provide higher levels of data throughput than others, while some merely provide an added level of data redundancy without providing much of an improvement to overall throughput.

RAID implementations remain one of the most popular options for in-server and direct-attached storage because of their performance, price point, and simplicity to implement and support. In arrays attached to SANs, RAID is commonly used behind the disk controller to provide performance and high availability. Table 2-4 shows the characteristics of commonly used RAID levels.

In many enterprise environments, RAID levels are abstracted from servers because storage capacity is deployed in a SAN. It is important to note, however, that some applications prefer to use spindles that are configured for a certain RAID level. For instance, an application that needs a volume with extremely high availability characteristics would most likely prefer to use a RAID-1 protected volume. An application that is performing a large number of reads from disk would prefer RAID-5, because multiple spindles could be used concurrently for read operations. An application that is constantly writing data may prefer RAID-0, because of its ability to stripe data across spindles without performance penalty.

With some subsystems, RAID levels can even be mixed to provide the best of both worlds. For instance, RAID 10 provides mirroring across equal-sized stripe sets. In this way, it provides 1:1 redundancy of the entire stripe set, and stripes read and write data across the spindles within the stripe set.

When choosing a storage interconnect, it is important to examine the performance characteristics. Many servers deployed today are still using legacy SCSI technology that limits maximum disk performance to 20 MBps or less. SAS and Fibre Channel are the more commonly used storage interconnects available today, as described in the next sections.