Fault Tolerance Protection and Raid Technology for Networks: a Primer

Home , Data redundancy, Data striping, Disk mirroring, Fault tolerance, Parity drive

51-30-25

DATA COMMUNICATIONS MANAGEMENT FAULT TOLERANCE PROTECTION AND RAID TECHNOLOGY FOR NETWORKS: A PRIMER

Jeff Leventhal

INTRODUCTION According to a recent Computer Reseller News/Gallup poll, most networks are down for at least 2 hours per week. The situation has not got- ten any better for most companies in the past 3 years. If an organization has 1000 users per network, this equals one man-year per week of lost productivity. Even if a network is a fraction of that size, this number is imposing. For nearly a decade, many companies responded by deploy- ing expensive fault-tolerant servers and peripherals. Until the early 1990s, the fault-tolerant label was generally afﬁxed to expensive and proprietary hardware systems for mainframes and mini- computers where the losses associated with a system’s downtime were costly. The advent of client/server computing created a market for similar products created for local area networks (LANs) because the cost of network downtime can similarly be economically devastating. Network downtime can be caused by anything from a bad network card or a failed communication gateway to a tape drive failure or loss of a tape used for backing up critical data. The chances that a LAN may fail increase as more software applications, hardware components, and users are added to the network. This article describes products PAYOFF IDEA that offer fault tolerance at the sys- What can an organization do to increase server tem hardware level and those that uptime and reduce, or even eliminate, network downtime? In many cases, a RAID system — a use fault-tolerant methods to protect collection of disks in which data is copied onto the integrity of data stored on net- multiple drives — is added to a network to speed work servers. The discussion con- access to mission-critical data and protect it in cludes with a set of guidelines to the event of a hard disk crash. This article dis- help communications managers se- cusses RAID technology and the use of fault-tolerance protection to preserve the availability and lect the right type of fault-tolerant integrity of data stored on network servers.

solution for their network. This article also discusses RAID (redundant array of independent [formerly “inexpensive”] disks) technology, which is used to coordinate multiple disk drives to protect against loss of data availability if one of the drives fails.

DEFINING FAULT TOLERANCE PC Week columnist Peter Coffee noted the proliferation of fault tolerance in vendor advertising and compiled a list of seven factors that define fault tolerance. Coffee’s list included safety, reliability, confidentiality, integrity, availability, trustworthiness, and correctness. Two of the factors — integrity and availability — can be defined as follows:

• Availability is expressed as the percentage of uptime and is related to reliability (which Coffee deﬁned to be mean times between failures) because inﬁnite time between failure would mean 100% availability. But when the inevitable occurs, and a failure does happen, how long does it take to get service back to normal? • Integrity refers to keeping data intact (as opposed to keeping data se- cret). Fault tolerance may mean rigorous logging of transactions, or the capacity to reverse any action so that data can always be returned to a known good state.

This article uses Coffee’s descriptions of availability and integrity to distinguish between products that offer fault tolerance at the system hardware level and those that use fault-tolerant methods to protect the data stored on the network servers.

Availability The proliferation of hardware products with fault-tolerant features may be attributable to the ease with which a vendor can package two or more copies of a hardware component in a system. Network servers are an example of this phenomenon. Supercharged personal computers equipped with multiple power supplies, processors, and input/output (I/O) buses provide greater dependability in the event that one power supply, pro- cessor, or I/O controller fails. In this case, it is relatively easy to synchro- nize multiple copies of each component so that one mechanism takes over if its twin fails.

Cubix’s ERS/FT II. For example, Cubix’s ERS/FT II communications server has redundant, load-bearing, hot-swappable power supplies; multiple cooling fans; and failure alerts that notify the administrator audibly and through management software. The product’s Intelligent Environ- mental Sensor tracks ﬂuctuations in voltage and temperature and trans-

mits an alert if conditions exceed a safe operating range. A hung or failed system will not adversely affect any of the other processors in the system.

Vinca Corp.’s StandbyServer. Vinca Corp. has taken this supercharged PC/network server one step further by offering machines that duplicate any server on the network; if one crashes, an organization sim- ply moves all its users to its twin. Vinca’s StandbyServer exempliﬁes this process, known as mirroring. However, mirroring has a signiﬁcant drawback — if a software bug causes the primary server to crash, the same bug is likely to cause the secondary (mirrored) server also to crash. (Mir- roring is an iteration of RAID technology, which is explained in greater detail later in this article.)

Network Integrity, Inc.’s LANtegrity. An innovative twist on the mirrored server, without its bug-sensitivity drawback, is Network Integrity’s LANtegrity product in which hard disks are not directly mirrored. Instead, there is a many-to-one relationship, similar to a RAID system, which has the advantage of lower hardware cost. LANtegrity handles backup by maintaining current and previous versions of all files in its Intelligent Data Vault. The vault manages the most active files in disk storage and offloads the rest to the tape autoloader. Copies of files that were changed are made when LANtegrity polls the server every few minutes and any file can be retrieved as needed. If the primary server fails, the system can be smoothly running again in about 15 seconds without rebooting. Be- cause all the software is not replicated, any bugs that caused the first server to crash should not affect the second server.

NetFRAME Servers. The fault tolerance built into NetFRAME’s servers is attributable to its distributed, parallel software architecture. This fault tolerance allows the adding and changing of peripherals to be done without shutting down the server, allows for dynamic isolation and connection of I/O problems (which are prime downtime culprits), dis- tributes the processing load between the I/O server and the central processing unit (CPU), and prevents driver failures from bringing down the CPU.

Compaq’s SMART. Many of Compaq’s PCs feature its SMART (Self- Monitoring Analysis and Reporting Technology) client technology, although it is limited to client hard drives. If a SMART client believes that a crash may occur on a hard disk drive, it begins backing up the hard drive to the NetWare ﬁle server backup device. The downside is that the software cannot predict disk failures that give off no warning signals or failures caused by the computer itself.

DIAL RAID FOR INTEGRITY In each of the previous examples, the fault tolerance built into the systems is generally designed to preserve the availability of the hardware system. RAID is a technology that is probably the most popular means of ensuring the integrity of corporate data. RAID (redundant arrays of independent disks) is a way of coordinat- ing multiple disk drives to protect against loss of data availability if one of the drives fails. RAID software:

• Presents the array’s storage capacity to the host computer as one or more virtual disks with the desired balance of cost, data availability, and I/O performance. • Masks the array’s internal complexity from the host computer by transparently mapping its available storage capacity onto its member disks and converting I/O requests directed to virtual disks into oper- ations on member disks. • Recovers data from disk and path failures and provides continuous I/O service to the host computer.

RAID technology is based on work that originated at the University of California at Berkeley in the late 1980s. Researchers analyzed various performance, throughput, and data protection aspects of the different ar- rangements of disk drives and different redundancy algorithms. The fol- lowing table describes the various RAID levels recognized by the RAID Advisory Board (RAB), which sets standards for the industry.

RAID Level Description Beneﬁts Disadvantages

RAID 0 Disk stripping: Storage is maximized Has virtually no fault data is written across across all drives, tolerance multiple disk drives features good performance and low price RAID 1 Disk mirroring: data is Data redundancy is Slower write performance, copied from one drive to increased 100%; has fast but twice the disk drive the next read performance capacity, more expensive RAID 2 Spreads redundant data Has no physical benefits Has high overhead with across multiple disks; no significant reliability includes bit and parity data checking RAID 3 Data stripping at a bit level, Has increased fault Is limited to one write at requires a dedicated parity tolerance and fast a time drive performance RAID 4 Disk stripping of data Has increased fault Slower write performance, blocks, requires a tolerance and fast read not used very much dedicated parity drive performance RAID 5 Disk stripping of both data Features increased fault Write performance is slow and parity information tolerance, efficient performance, is very common

The redundancy in RAID is achieved by dedicating parts of an array’s storage capacity to check data. Check data can be used to regenerate in- dividual blocks of data from a failed disk as they are requested by the applications, or to reconstruct the entire contents of a failed disk to restore data protection after a failure. The most common forms of check data are a mirror (identical copy) of user data and shared parity, which involves appending mathematical code to data bits for later comparison, matching, and correction. Differ- ent combinations of mapping and check data comprise distinct RAID levels.

Striping Of the six well-deﬁned RAID levels, three are commonly used. Level 1 uses mirroring for data protection and may incorporate striping. Striping refers to the location of consecutive sequences of data blocks on succes- sive array members. Striping balances I/O load, thereby increasing performance. Levels 3 and 5 both use parity for data protection and almost always incorporate striping. RAID levels 3 and 5 use different algorithms for up- dating both user data and check data in response to application write requests. In a RAID level 3 array, the disks are physically or logically synchro- nized transmission, and each contributes to satisfying every I/O request made to the array (i.e., parallel access). In a RAID level 5 array, the disks are allowed to operate independently (i.e., independent access) so that, in principle, the array may satisfy multiple application I/O requests con- currently. Some RAID levels are theoretically faster than others, but in many sit- uations the existing hardware technology does not always enable these performance enhancements to be realized. Other factors that are signiﬁ- cant in overall system performance include the combinations of the disk drive, the host adapter, the tuning of the operating system, and how these components function together.

Parity In RAID level 3, parity information that is saved to one designated tape drive can be used to regenerate data from a failed drive or tape media (see Exhibit 1). RAID 5 offers improved storage efﬁciency over RAID 1 because parity information is stored rather than a complete redundant copy of all data. The parity information is essentially a number determined by adding up the value of all the bits in the data word. Parity requires some amount of overhead, ranging from 50% on RAID 1 to somewhat less than 20% on RAID 5.

EXHIBIT 1—Example of RAID 3 — Parity is Streamed to One Drive Device

Host System

Parrallel Data Queue

Data Data Data Data Parity Data Data Data Data Parity Data Data Data Data Parity Data Data Data Data Parity Data Data Data Data Parity Data Data Data Data Parity

The result is that three or more identical drives can be combined into a RAID 5 array, with the effective storage capacity of only one drive sac- rificed to store the parity information. Therefore, RAID 5 arrays provide greater storage efficiency than RAID 1 arrays. RAID 5 is an implementation in which parity information is striped across all the configured drives. This method can increase the array throughput by separating the parity information across all drives. This is the preferred method when using transaction or data base processing (see Exhibit 2). Data blocks and parity blocks are striped to the drives or tapes in a stair-step, or barber-pole fashion, allowing for full restoration even if a disk or tape is lost or damaged. In such an event, the data and parity blocks on the remaining drives or tapes contain enough information for the software to extrapolate the “lost” data.

Mirroring Modern operating systems that are built for the enterprise, such as Win- dows NT 4.0, provide both RAID 1 and RAID 5 fault-tolerance protection. RAID level 1 (disk mirroring) simultaneously streams data to two hard drives (or tape devices) throughout the entire job, not just in the event of the hardware failure. Both drives are considered to be one drive by the Windows NT software (NT’s fault tolerance driver is called FTDISK.SYS). Disk mirroring creates a duplication of partition data onto another physical disk. Any partition, including the boot or system partitions, can be mirrored. This strategy protects a single disk against failure. Disk mirroring is generally appropriate for small LANs, as its initial cost is limited to only two disk drives. As the need for more network stor-

EXHIBIT 2 —Example of RAID 5 — NOTE: This method is referred to as the barber pole because of the way the data and parity blocks spiral around.

Host System

Parrallel Data Queue

Parity Data Data Data Data Data Parity Data Data Data Data Data Parity Data Data Data Data Data Parity Data Data Data Data Data Parity Parity Data Data Data Data age capacity grows, mirroring may become more expensive per megabyte than other forms of fault tolerance because only 50% of the disk space is being used. However, with the advent of low-cost integrated drive electronics (IDE) (1 GB for under $300, for example), disk-mirroring controllers and large mirrored RAID drives could become viable al- ternatives, in workstations and small work groups, to the faster, more powerful small computer system interface (SCSI)-based arrays — provid- ed users do not need the full performance of the SCSI systems. For extra security, an entire RAID system can also be mirrored, but of course that doubles the cost. How communications managers determine whether mirroring or striping is best for their enterprise depends on their conﬁdence in the technology and how much they are willing to spend. Mirror drive sets provide a modest performance increase for reading because the fault-tolerance driver can read from both members of the mirror set at the same time. However, there is a slight performance decrease when writing to a mirror set because the fault-tolerance driver must write to both members simultaneously. When one drive in a mirror set fails, performance returns to normal because the fault-tolerance driver is working with only one partition. If a mirror set’s disk controller fails, then neither drive of the mirror set will be accessible. It is wise to install a second controller in the computer so that each disk in a mirror set has its own controller. This arrangement is called a disk duplex. Duplexing is a hardware solution to fault tolerance and should improve performance. It is helpful to add a duplicate power supply as well.

Tape Backup Devices Cheyenne Software (now owned by Computer Associates) was the ﬁrst to offer RAID 5 for tape devices. This allows the routine backup window to be met (albeit with a 10 to 15% decrease in performance when both drives are treated as one drive) and affords the user the security of ar- chiving the second tape offsite. Currently, other backup software products wait for the ﬁrst tape device to fail, in which case the entire backup job must be performed again from the start, requiring as much as double the time that it normally takes to perform the backup. Because by nature tape devices employ a sequential access method, RAID 5 is an ideal solution for a tape array. The ability to deliver several parallel data streams to an array of tape devices allows for a highly scal- able transfer performance. Although tape devices possess very intense error detection and correction algorithms, the user is still left vulnerable to mechanical or tape cartridge failure that could render a critical restore of data unobtainable. A RAID 5-based array of tape devices provides an economical way to protect the backup/restore session data. RAID level 5 (data striping) is only supported with three or more disk drives in NT, and three tape drives with Cheyenne’s tape software. If one of the drives is destroyed, a complete restore of the data set can still be made from the remaining tapes. If less than three tape drives are used, the data striping will occur, but without fault tolerance. The location of the parity information is moved from drive to drive in a barber-pole fashion. This is extremely important for tape RAID systems, where tape drives compress information.

CONCLUSION When selecting a network’s RAID level, the choice between the RAID 1 and RAID 5 depends on the level of protection desired and an organization’s budget. The major differences between disk mirroring and striping with parity are performance and cost. Generally, disk mirroring offers better I/O performance. It also has the advantage of being able to mirror the boot or system partition. Disk striping with parity offers better read performance than mirroring; however, the need to calculate parity information requires more system memory and can slow write performance. The cost per megabyte is lower with striping because disk utilization is much greater. For example, if there are four disks in a stripe set with parity, the disk space overhead is 25%, compared to 50% disk space overhead for disk mirroring. An effective means of achieving fault tolerance is to combine RAID 1 disk mirroring with RAID 5 data striping on the same computer. Consider

mirroring the system and boot partitions and protecting the rest of the drive with stripe sets with parity.

Jeff Leventhal is the founder and president of Remote Lojix, a computer services company that provides computer and network diagnostics and repair services in 40 cities nationwide. Mr. Leventhal’s company features 40 manufacturers’ authorizations and a 4-hour response time. He can be reached at (800) 565-4912.