Introduction Chapter seven is a distinct shifting of direction from the first six chapters. We have worked our way up through the lower layers of the OSI Model building a tool kit of connectivity technologies for constructing a network. Now the focus will move to the systems that populate the network. By analogy, we have been studying how to build roads and highways, now we will start studying how to build the cars and trucks that travel our roads. Purpose and History of Servers History Relationships and Types Sizing Rules Hardware Components CPU / System Board Memory Storage RAID Essential But Negleted Components Power Protection Backup and Recovery Conclusion Self Check
			[ I ] PURPOSE AND HISTORY OF SERVERS The primary purpose of a server is to provide a centralized place to access some kind of shared resource. A shared resource can be many things, ranging from information in a database to computing power to solve large mathematical problems, to pools of CDROM drives or modems. Modern network servers are somewhat like earliear mainframe and minicomputer systems in that users from across the network come to the server to access what it has to offer. However, the analogy can be taken only so far. The clients on the network are intelligent systems that can perform computing functions on their own, unlike the terminals associated with a mainframe or minicomputer. The relationship is somewhat less tilted: the server holds something the client wants, but the client is not entirely helpless without the server. What servers provide to their clients on the network can be broken down into three broad categories: Peripherals, Services, and Information. The table below shows what falls into each category. Type of Server Function Peripheral Shared CDROMs, modems Services File & Print, faxing, network directory, authentication Information databases, web servers Others divide servers into two broad categories -- application and infrastructure servers. This division puts peripheral and services servers in the infrastructure group, and populates the application group with email servers, databases, etc. Either method of dividing servers into types is necessarily an arbitrary division, useful for discussions and planning. Also, the specifics of how you would tune a server for optimal performance vary depending on the server's role.
			History The first "servers" in the early 1980s were simply large disk drives (usually 20-30MB) with connections for several computers. Each computer got a "slice" of the disk drive, with no sharing between them. These "disk servers" were a good idea -- and illustrate a principle in use throughout networking: sharing an expensive piece of hardware among many users to lower the cost per user -- but were very limited in functionality. The next step came in the middle '80s with the development of Novell NetWare and SUN's NFS (Network File System). Both gave multiple users shared access to large, centrally located, disk storage. NetWare was aimed at PC users, and NFS was aimed at UNIX workstations. Both represent the birth of the file server: a system whose job it is to provide access to files for uses on the network. A "large" disk at that time was 200 to 300MB. NetWare represents the first true NOS, or Network Operating System. It provided file services, printer sharing, security and access control. The "how" of each of these functions is discussed in Chapter Eight. Modern Server Types The relationship between a client system and a server can be described as being one of two types: client/server or peer-to-peer. In a peer-to-peer, or "peering" relationship, each system has some resource it shares with other systems on the network, and some resource(s) it looks to other systems to provide. For example, in a small office or workgroup someone might share access to his Zip drive, while others in the group share access to their hard drives and printers. Each person's PC is acting both as a server -- providing some kind of resource -- and as a client -- taking advantage of some kind of resource. When the resource providing and using has this "share and share alike" nature to it no one system is acting as "the server": the systems are operating as a collection of somewhat equal peers.
	"UNIX" means any of the common UNIX-like operating systems: AIX, HP-UX, SunOS, Solaris, Linux, UnixWare, SCO UNIX, etc....		The following operating systems have peer-to-peer capabilities: Apple MAC OS; Microsoft Windows v3.11 (Windows for Workgroups), '95, '98, and NT Workstation/Server; IBM OS/2; UNIX. A client/server relationship exists when one system (or a collection of systems in a larger network) are employed solely for the purpose of providing access to some resource, and where systems who do not have resources to share come to that system. A server running an Oracle database with fifty users connecting to it from their PCs is a good example of client/server in action. Servers in the client/server model tend to be either high-end PC class systems or minis/mainframes. The following operating systems can be used as servers for client/server systems: Apple MAC OS; Microsoft Windows NT Server; Novell NetWare; IBM OS/400, OS/390, OS/2; Banyan Vines; UNIX. Other operating systems certainly belong on this list: the list is partial and meant to enumerate the common, main-stream server operating systems. Sizing Rules This section is most appropraitely titled "Mr. Shipman's Rules of Division", or "When you have to pony up and buy a real NOS!" For small offices and workgroups peer-to-peer networking offers an easy and cost-effective way to share resources like files, laser printers, removable media disk drives (a la Zip and Jazz). But there comes a point where the number of computers sharing resources becomes large enough that maintaining all of the shares turns into an administrator's nightmare, and the load on individual systems due to peer traffic becomes intolerable. The table below shows my suggested guidelines for when to take the steps away from a peering environment toward a more traditional client/server environment. Number of Users Model 2 - 5 Simple peer-to-peer is OK. 5 - 15 Peer-to-peer is still OK, but consider setting up one PC as a "peer server" to handle all file storage and heavy printing. This PC should have no other jobs. EG: don't have a user sitting at this PC! 15+ Invest in a true client/server NOS and dedicated server hardware.
			[ II ] HARDWARE COMPONENTS The answer to the question "what makes a fast server?" is a moving target. The number of devices on a microprocessor chip doubles roughly every 18 months. The current rate of technological advance will probably continue into the first decade of the 2000s. Sometime in the next ten to fifteen years the cost of new chip fabrication plants -- needed to keep up the doubling of device count every 18 months -- will exceed the amount of capital available to the microprocessor industry. This will lead to a slowdown in the pace of processor chip advances, but it won't stop. The chart below shows the performance of Intel's processor chips (measured in Millions of Instructions Per Second, or MIPS).
	Read about Moore's Law, formulated by Gordon Moore, founder of Intel.		In an effort to provide information with lasting value, this section will focus on elements of server systems that, relative to each other, do not change rapidly. I will deal primarily with Intel-based systems. The general characteristics of a server-class system, are a computer with a fast processor, a large amount of memory, and a large amount of disk storage.
			CPU / System Board The Central Processor Unit or CPU lies at the heart of all computer systems. It is instinctive, and quite correct, to assume that servers should have the fastest processors available to insure speedy operation. A slow server can affect the performance of all its clients across the network. However, raw processor speed isn't the only thing to be aware of when specing out a new server or evaluating an existing system. The system board itself plays a large role in server performance. When reading a system's specifications it's easy to think that because it has a 300 MHz CPU that everything going on inside is operating at that high clock rate. This is very much not the case. Fabricating a single chip that can operate at high clock rates (hundreds of MHz) is a difficult process. As the pace of operations on the chip increases, the diificulty of keeping everything running in sync increases exponentially. A hefty percentage of the "landscape" on a modern processor chip is devoted to clock circuitry. The challenge is making sure that the ticking of the clock arrives everywhere at the same time. If keeping the beat correctly on a processor chip the size of a postage stamp is hard, think of the problems of keeping time at the same rate over something the size of an entire system board. Modern system boards are clocked at a fraction of the processor's speed. This practice started with Intel's 486DX2 chips (where a 50 or 66MHz chip ran on top of a 25 or 33MHz system board respectively) and has continued with Intel's Pentium family (as well as with Alpha and PowerPC chips). A big advancement came in 1998 with the availability of 100MHz system boards. Previously, everything from a Pentium 133MHz to a Pentium II 266MHz ran on top of a 66MHz system board. The difference in performance between a 300MHz Pentium II system with a 100MHz system board and a PII 266MHz with a 66MHz system board is substantially larger than the difference between the same PII 266MHz and a PII 233MHz (also with a 66MHz board). So what does this mean to you and me? The clock rate of a server's system board dictates how fast data will move from memory to the processor, from peripheral devices to memory, and vice versa. Ever increasing processor speeds won't mean great strides in system performance unless the speed of the underlying system board continues to advance as well. The outlook is that system board speeds will increase over time, but at a much slower pace than processor speeds due to the greater difficulties involved.
	You may have read about another bus inside your PC: USB or Universal Serial Bus. USB is intended to be a high-speed channel for attaching multiple external peripherals to your PC, such as modems and scanners. There are even USB hubs for connecting lots of USB devices to one PC. The technology is remeniscent of a polling architecture where the PC is in control of all the devices, and there can be only one controller (PC) on the bus.		Another component of the system board that affects overall performance is the peripheral bus (or referred to simply as the bus) type. The peripheral bus provides expansion slots on the system board for installing things like network interface cards and disk controllers. Three bus types are currently in use: Industry Standard Architecture (ISA), Extended Industry Standard Architecture (EISA), and Peripheral Component Interconnect (PCI). ISA and EISA are fading in popularity while PCI's star is still rising. Type Width Speed Notes ISA 16 bits 8.33 MHz Still used for modem, serial ports EISA 32 bits 8.33 MHz Most new servers have 1 or 2 EISA slots. ISA cards will work in EISA slots PCI 32, 64 bits 33, 66 MHz Most systems are still 32 bit, 33MHz PCI; 64bit PCI and 64bit/66MHz PCI is coming The table above details the current bus types, width, and speed. Width indicates the number of bits of data the bus can transmit at once. Speed indicates the clok rate used to time bus transfers. Wider busses, and busses with higher clock rates can move more data per second than narrower busses or slower busses. PCI is the current standard for server peripheral busses. EISA (pronounced "ee-sa") was the predominant server bus before the development of PCI. EISA bus slots have a feature which PCI slots do not: backward compatibility. You can use ISA adapters in EISA slots (but not the other way around), while PCI isn't backward compatible with either ISA or EISA cards. Most high-end servers will have a mix of PCI and EISA slots, with the trend being toward more PCI slots and fewer EISA slots. Newer entry-level or departmental servers may have only PCI slots, or PCI and one ISA slot for installing a modem card. In the early 1990s the VESA or VL (a.k.a. Local Bus) made a brief appearance in 486-class systems. The VL bus offered up to two slots per system at 32bits and 33MHz. VL bus cards were most often video cards or disk controllers. Once it was available on the market PCI rapidly displaced VL. PCI isn't limited in the number of slots per system, and had very broad industry support from the outset. Another once prominent bus type deserves some mention. In the late 1980s IBM introduced the Microchannel Architecture. Microchannel (or MCA) was a better bus than ISA, offering 16 and 32 bits slots and software configuration of adapters (both EISA and PCI have this feature). Unfortunately, IBM also made MCA expensive, and not backward compatible with existing ISA cards. Microchannel has pretty much died now that IBM has embraced PCI, however there are certainly still thousands of Microchannel-based IBM desktops and servers out in the field. PCI makes the same break with ISA compatibility that MCA did, however PCI has become a great success where MCA's lack of backward compatibility was a much derided liability. This is a testament to what solid industry backing can do for a new technology.
			Memory Main memory on a server has two primary roles. First is the obvious one; to store the programs and data that the operating system is actively working with. Second is to buffer interactions between fast components (namely the CPU) and slow components (usually the disk drives). There are a couple of names for memory used for this second purpose; cache memory, or buffer memory. The terms are largely interchangeable. Which you use depends most on which OS you tend to use most. NetWare people talk about "cache" while some UNIX types refer to "buffers". The role of cache in a system is critical. As the graphic below shows there is a distinct hierarchy of components in terms of speed. Bear in mind during this section that speed costs money. So, engineering tradeoffs are made to increase system performance while holding system costs down. The CPU is the fastest element of any computer system, particularly a modern system where the CPU's clock is running three to five times faster than the system board. The CPU is surrounded by several layers of cache which work to hide the "slowness" of the next layer out. The purpose of cache is two fold: To permit the system to write data out to a slower medium when it is convenient (instead of at the moment the data is created or changed). To "prefetch" or "read ahead" so that the next piece of data or instruction the processor needs has been brought in from slower media (eg: disk or main memory) before the CPU needs it. Layer 1 (L1) cache is closest to the CPU, and is the fastest memory in the system. In Pentium II and Pentium III systems L1 cache is built into the cartridge that holds the processor chip. L1 cache operates at the processor's clock rate, and tends to be fairly small (a few kilobytes). Layer 2 (L2) cache forms the next outer ring. It is on the system board and operates at the system board's speed, however the access time for L2 cache is shorter than that for main memory (nanoseconds vs. tens of nanoseconds). L2's job is to buffer data being written out of L1 cache, and to hold data or instructions that have been fetched from main memory but not yet moved into L1 cache. L2 cache tends to range from several hundred kilobytes to a few megabytes. Main memory, a.k.a. RAM (Random Access Memory) is the system's primary place to temporarily store the data and instructions it is working on. A server's main memory also serves to buffer writes to disk, and as a place to hold data that has been read from disk prior to its being needed by the CPU. If a server has too little memory for caching, the server will have to wait for disk reads and writes to complete before it can go on with its work. This slows the server down, and slows down all of the users who are attached to the server (by making them wait for the server to respond). Getting cache algorithms right is a critical part of operating system design. Making sure you have enough free memory for caching is a critical part of good system administration. Most operating system vendors will offer formulas to compute how much memory to install in a server. The formulas can take into account things such as amount of disk storage, number of concurrent users, and the type and size of applications running on the server. If you must choose between a faster CPU or more memory, choose more memory. Only the heaviest mathematical modelling applications are more CPU-intensive than they are memory-intensive. File servers and database servers live and die by how much memory they have available.
			Storage A good argument can be made that a server's disk drives are its most critical components. Disk drives are the server's means of permanently storing programs and data. They are also its slowest and most failure-prone subsystem. Both of these traits are due to the fact that, unlike any other part of the server, disk drives are electromechanical devices with moving parts. Disk drive latency (the time it takes to move the drive's read/write heads to the correct position and then for the disk to rotate until the desired data passes under the heads) has been reduced dramatically by decreasing the size of the disk platters (so the heads have less distance to move), and increasing the rotational speed of the drive. Current server disk drives spin at 7200 or 10,000 RPM. These advances also improve a drive's transfer rate -- the rate at which it can read or write data. Reliability has also improved dramatically in the past ten years. In 1989 a good quality disk drive had a MTBF (Mean Time Between Failure) of 20,000 to 30,000 hours. Current disks have ratings of 300,000 to 500,000 hours or more. But, with the exceptions of fans, disk drives are still the only moving parts inside a server, and thus are still more failure-prone than memory or CPUs. Most disks today are "intelligent" devices: the control circuitry for the drive is built into the unit. Older disk drives were "dumb" devices which required an external controller card to provide all the instructions a drive needed to operate. A simple way to tell a "smart" drive from a "dumb" one is to count the number of cables connecting the drive to the controller card. Older drives had two cables (one for data, one for control), where newer devices have just one cable. Here, as with peripheral busses, there are several types of system-to-disk connections. Each type has several sub flavors. The two primary disk interface technologies in use today are SCSI (Small Computer System Interface), and IDE (Integrated Device Electronics). The chart below shows the interface types and their flavors (each family is color-coded). Three obsolete interface types you might encounter in older equipment are: MFM, RLL, and ESDI. Type Devices per channel Speed Width Throughput IDE 2 5 MTS 8 bits 5 MB/sec EIDE/AT2 2 10 MTS 8 bits 10 MB/sec UDMA/ATA33 2 33 MTS 8 bits 33 MB/sec ATA66 2 66 MTS 8 bits 66 MB/sec SCSI 7 5 MTS 8 bits 5 MB/sec Fast SCSI2 (a.k.a. narrow SCSI) 7 10 MTS 8 bits 10 MB/sec Fast & Wide SCSI 15 10 MTS 16 bits 20 MB/sec Ultra SCSI (a.k.a. Ultra Narrow SCSI) 7 20 MTS 8 bits 20 MB/sec Ultra Wide SCSI (Sometimes {wrongly} called SCSI3) 15 20 MTS 16 bits 40 MB/sec Ultra2 SCSI LVD 15 40 MTS 16 bits 80 MB/sec Ultra-133 SCSI 15 66.5 MTS 16 bits 133 MB/sec
	A note about SCSI addresses. "Narrow" SCSI uses a three bit address, giving addresses from 0 to 7 (the controller is usually 7). "Wide" SCSI adds a fourth address bit to give addresses from 0 to 15. The controller is still usually 7. If you mix wide and narrow devices on the same bus be sure that no wide device has an address that is exactly eight higher than a narrow device. The result will be the same as if you had two devices with the same address (a Bad Thing). A simple way to avoid this when mixing narrow and wide devices is to only use addresses below seven.		SCSI is the preferable drive interface type for servers larger than very small departmental or workgroup servers. With current SCSI systems a single controller can have a chain of up to fifteen devices, and with the Low Voltage Differential (LVD) technology in Ultra2 SCSI, transfer eighty megabyts per second. Even in its most enhanced version (ATA66) IDE systems can only handle two devices per channel. Most system boards will have two IDE controllers for a total of four devices, but this is still very limited when compared to SCSI. The SCSI bus also supports a wider variety of peripherals than IDE. IDE is limited to disk drives, CDROMs, and certain other deivces like Zip drives. SCSI can support disk drives, tape drives, CDROMs, scanners, printer interfaces, and other 'exotic' peripherals. SCSI also supports a feature which IDE does not: device disconnect. SCSI disconnect permits a SCSI controller to issue a command to a drive, disconnect from it and issue a command to another drive, then reconnect to the first drive to check the result of the command. This frees up the controller and the channel to process multiple disk requests. IDE-derived controllers have to wait for the disk drive to complete the command before they can issue another one. Additionally, most SCSI disk drives support a feature that is not present on IDE drives called Tagged Command Queueing. Command queueing permits a controller to hand a SCSI drive a list of commands to execute. The drive's onboard controller will order the commands and execute them without further intervention from the SCSI controller. This further frees the SCSI bus and controller to deal with other drives and commands. As long as you do not attempt to mix Single Ended and Differential SCSI devices, you can have more than one type on a single SCSI bus. For example, your server might have a SCSI CDROM and SCSI tape drive (both Fast-SCSI2 {narrow} devices) on the same channel as a pair of Ultra-Wide SCSI disk drives. The SCSI controller will negotiate with each device on an individual basis to determine the maximum width and speed the device can support. (The negotiations are done using 8 bit transfers at plain SCSI speeds.) Some manufacturers advise not mixing device types in order to avoid having a fast device (like an Ultra Narrow disk drive) wait for a slow device (like a Fast-SCSI2 CDROM) to finish transferring data over the SCSI bus. This makes sense, so you might want to consider segregating older devices on a separate controller.
			[ III ] RAID RAID, in the computing sense, stands for Redundant Array of Inexpensive Disks. The original research work which defined six RAID techniques -- known as RAID levels 0 through 5 -- was done at the University of California Berkeley campus. The six RAID levels provide different techniques for building fault tolerant disk systems. A fault tolerant disk system is one that can survive the failure of one or more individual disk drives and still remain in operation. Using RAID is a popular way to increase the reliability (and thus availability of a server. For any mission critical server, some form of redundant disk system is mandatory. Different manufactuers may describe systems with RAID levels beyond 0-5. Only RAID 0 through RAID 5 have agreed upon meanings. Be very carful when examining a disk system billed as "RAID 6" or RAID 10". The most commonly used RAID levels are 0, 1, and 5. The chart below describes each of the common RAID levels and how it works. Schematic RAID Level and Description RAID0 - Striping Striping does not actually provide any protection against data loss due to disk failure. RAID0 is a performance enhancement technique: blocks of data are written evenly across all disks in the set (or "striped"). Read and write times for a striped set vs. a single disk drive are close to 1/N : 1, where N is the number of disk drives in the set. Because there is no data redundancy a single disk failure will take the entire striped set out of action. The only option is to replace the failed disk and restore the entire array from backup. RAID1 - Mirroring Mirroring is the simplest form of disk redundancy. Disks are set up in pairs. All of the data that is written to one disk is written to the other. If one disk in the mirrored set fails, the other can continue the operation of the set with no major degradation of performance. When the failed drive is replaced, the system will copy the data from the good drive to the replacement. Read performance vs. a single drive is roughly double. Write performance is often quoted as 1/2 that of a single disk. With duplexing (discussed below) and SCSI disconnect this write performance hit has almost been negated. RAID3 - Parity RAID3, also known as Data Guarding, uses a mathematical parity equation to protect data. In each three disk set, one disk is devoted to holding parity. When data is written to the set 1/2 of the data is written to disk 0, and 1/2 to disk 1. A parity value is computed from the two halves of the data. Using the parity value and one of the halves you can determine what the value of the other half of the data is. So, if either disk 0 or 1 fails, the system can use the remaining half of the data on the surviving disk, plus the parity information, to compute the value of the data stored on the failed drive. If the parity drive fails the system is not affected: it operates as if it were a two disk RAID0. Parity is less expensive in terms of hardware than mirroring, but system performance when a disk has failed is significantly degraded due to the overhead of computing the values of the missing data. Once the failed disk is replaced the system will rebuild the information on the drive. RAID5 - Distributed Parity RAID5, or Distributed Data Guarding, operates like RAID3 except that parity information is spread evenly over all of the drives in the set. RAID5 was developed to overcome one of RAID3's shortcomings. Because each disk write must also generate a write to parity, the single parity disk in a RAID3 set can become a bottleneck. Distributing parity information across all drives breaks that bottleneck. Duplexing is a technique for reducing the number of single points of failure within the disk subsystem to as few as possible. With simple RAID1 all of the mirrored disks are on one disk controller channel. There are any number of points of failure which could stop the array of mirrored disks from operating: a single power supply in the cabinet holding the disks, the single cable connecting the disks to the server, and the single controller card in the server. Duplexing seeks to remove these single points of failure by using two of everything. With duplexing, the disks of a mirrored pair are housed in different cabinets, with separate cables connecting to two separate disk controllers. With this setup a single failure will only stop one half of each mirrored set, and the system will continue to run. If you choose RAID1, and your operating system supports duplexing, you should use it. The graphic above shows a server with RAID1 and Duplexing. The server has two disk controller cards, each connected to a separate disk cabinet. Each cabinet has half of three mirrored pairs (indicated by color). Orthagonal RAID is a term used to apply to RAID3 or RAID5 arrays where the principles of duplexing have been applied. "Orthagonal" is a mathematical term meaning "separated". In an orthagonal RAID5 each disk is on a separate controller. This limits the impact of a single controller failure to taking out only one disk in the array. A number of manufacturers offer RAID controller cards which will implement one or more RAID levels in hardware. RAID controllers tend to be expensive items, and can represent a major single point of failure. But for operating systems that lack native support for RAID, a RAID controller can be a good buy. In large disk arrays, implementing RAID5 with a RAID controller can be substantially more cost effective than Mirroring or Duplexing. This cost savings comes with reduced reliability, as explained above. Large systems requiring maximum availability often use a mix of RAID0 and RAID1 concepts, either mirroring two RAID0 sets across different controllers, or creating a RAID0 set out of several mirrored pairs of disks.
	A paper on RAID is online in PDF format.		Which RAID level is right for your application is determined by your priorities and the type of application. Each level has different price, performance, and availability characteristics.
			[ IV ] ESSENTIAL BUT NEGLECTED COMPONENTS In designing small and midsized servers for small companies and workgroups, many people operate in a penny wise and pound foolish mode, skimping on essential components that they see as unnecessary, or luxury add-ons. Two of the most frequent victims of this cost-saving attitude are proper power protection and data backup. Industry research has shown repeatedly that the upfront purchase cost of a system, whether it is a PC or a server, is only a small portion of the lifetime cost of the system. Skimping on essentials like power protection and data backup are sure ways to maximize long-term cost while giving the (false) impression of saving money up front. Power Protection Proper power protection plays two roles in preserving your investment in time and equipment, and in protecting the health and well being of your business. Power protection, as it is used in this text means: providing your network systems with protection against surges, spikes, sags, brownouts, and outages in the electrical supply and providing automated shutdown capability for servers in the event of a prolonged blackout. The simplest form of protection against power surges, spikes, and lightning is a good quality surge suppressor, like the American Power Conversion SurgeArrest pictured below. Surge suppressors, or "surge strips" do not provide any backup power in the event of an outage. At a minimum, however, they are a good place to start. Every piece of network equipment should have at least a surge suppressor between it and the wall outlet. When shopping for a surge suppressor beware of similar-looking "power strips" which only provide extra outlets but no actual power protection: just because a power strip has a circuit breaker on it, dos not mean it's a surge suppressor!
	Links to several UPS vendors:		The preferred method for providing full power protection to network systems and servers is the UPS, or Uninterruptible Power Supply. A UPS is essentially a sureg suppressor coupled with a battery and the electronics needed to convert the battery's DC output to the AC output needed by your equipment. There are several types of UPS: online, standby, and line-interactive. Each type has a different way of providing output power to the load (your server or network hardware), and a different way of shifting from utility power to battery power in the event of an outage. Holy wars have been fought over which is better. A simple recommendation from my experience is this: stick with an online or line-interactive model. In theory a standby UPS can cause a momentary blip in its output power when it switches to battery, but in practice, they all get the job done!
	Because of the nature of AC power you can't just multiply Volts x Amps to get Watts like you can with DC. A handy conversion is: 1.6 VA = 1 Watt.		UPS units are available in both floor standing units (above left), and rack mounted units (above right). The UPS manufacturers provide sizing tools to determine how large a UPS you need for the size of your equipment and your desired run-time on battery. UPS power output is measured in Volt-Amps, or VA. Generally, a PC, router, or stack of hubs will require a unit in the 400 to 600VA range to provide about twenty minutes of battery power. A server with several gigabytes of disk storage will require a 1000VA unit or larger. Providing backup power to a server is only half the story, however. A UPS provides power to keep the server running when the lights go out, but if you're not there to bring the server down, it will crash when the UPS battery runs out, just as if you didn't have a UPS in the first place. (A Bad Thing, as a server crash with databases and files open can result in lost data or corrupt information.) The other half of the story then, is automatic shutdown software. All of the major UPS manufacturers provide software that enables the server OS to monitor the UPS and shut the server down automatically when the UPS is on battery and power is running out. The monitoring connection usually takes the form of a serial cable running from the server to the UPS. UPS monitoring software is generally easy to install and configure. Providing a UPS with automatic shutdown software is a must do for protecting your servers. Depending on your application, you might have a requirement for continuous power. UPS units are battery-based; just like the battery in your flashlight or Walkman they have a finite amount of stored energy. If a blackout is long enough your UPS will run out of stored energy and will shut down. Certain applications such as banking, health care, civil preparedness, and national defense require continuously available power. Engine generator units like the ones from Onan and Caterpillar pictured below can provide that power. Standby generator units are usually powered by diesel engines, or engines made to run on some kind of gaseous fuel like propane or natural gas. The diagram below shows a typical arrangement of generator, UPS, and servers. A critical component in this setup is the automatic transfer switch which will start the generator automatically when the commercial power fails. A UPS is still needed to provide power during the time between the start of the blackout and when the generator is up and running. (The UPS also protects the servers from variations in the power produced by the generator.) Click image for an animated illustration.
	See what HP has to say about backing up your data.		Backup and Recovery The other victim of up-front cost cutting, and unfortuately human laziness, is data backup. The first rule of keeping your data safe is to have it in at least two places. Though it is the most neglected task in small businesses and workgroups, there is absolutely no acceptible excuse for not taking the time to properly back up your data. There are many good quality backup hardware and software products on the market, covering every major commercial server operating system. The most common backup mechanism is some form of tape drive. Removable media drives such as ZiP, Jazz, or LS120 drives can be used, but they don't have the capacity to backup anything more than a very small server. Writeable CDROM is also used for data backup, but the cost per megabyte is more than that for most types of tape, and there is the limited capacity of each CD (650MB). In the tape arena there are 4mm and 8mm tape drives based respectively on DAT (Digital Audio Tape) and compact video casette tape technology. There are DLT (Digital Linear Tape) drives, and drives based on older 1/4 and 1/2 technologies. (Travan tape drives are descendants of 1/2 tape drives). HP 4mm DAT drives are pictured above. In the small to mid-sized server market Travan drives cost less up front, but 4mm DAT will give more bang for the buck over the long haul (more expensive drive, much cheaper tapes). The large server market is starting to become dominated by DLT. Tape drive manufacturers offer configuration charts to allow you to match the right tape drive with the amount of data you need to back up. The most direct method is to attach a tape drive (or drives) directly to the server and use a server based backup software package. Server based packages offer better performance and reliability than workstation based packages. The last word on this subject is simply do it! How to guard your business against disaster due to data loss is discussed in detail in Chaper 9. Maintaining good tape backups is just one tool to insure that a fire, flood, or disk drive failure won't put you out of business.
			[ V ] CONCLUSION Chapter seven covered the basics of reliable server server systems. We started at the innermost layer with the server's CPU, memory and system board, and moved outward to address disk storage, disk redundancy, power protection and data backup. A new server purchase is just the tip of the cost-of-ownership iceberg. Disk redundancy, power protection, and data backup are three tools to hold down the long term cost of owning a server, and to protect your business from financial losses driven by mechanical failures, power outages, or other disasters.
			[ VI ] SELF CHECK If a file server is has too little free memory, how will it perform? Which of the three common peripheral bus types can move the most data per second? Which can move the least? Why might PCI have succeeded in the PC market while Microchannel failed? Can I use a Fast-SCSI2 scanner on the same bus as a Ultra2 SCSI LVD disk drive? How many devices can you attach to a narrow SCSI bus? How many to a wide SCSI bus? How many devices on a bus with a mix of wide and narrow devices? Will a surge suppressor protect your equipment during a blackout? How? If it won't provided total protection, what will? Name an application not listed in the Power Protection section that might require continuous power. List three types of backup media. Why would you want to store your data in two separate locations?
			[ CH 6 ] [ TOC ] [ CH 8 ] 1999, 2000 Shipman | Created 3-17-99 | Updated 2-5-00