182 Networking: A Beginner’s Guide N Ultra SCSI Also called SCSI-3, this specification increases the SCSI bus speed even higher—to 20 MHz. Using a narrow, 8-bit bus, Ultra SCSI can handle 20 MBps. It can also run with a 16-bit bus, increasing the speed further to 40 MBps. N Ultra2 SCSI Yet another enhancement of the SCSI standard, Ultra2 SCSI doubles (yet again) the performance of Ultra SCSI. Ultra2 SCSI subsystems can scale up to 80 MBps using a 16-bit bus. N Ultra160 SCSI By now you should know the story: Ultra160 SCSI again doubles the performance available from Ultra2 SCSI. Ultra160 SCSI (previously called Ultra3 SCSI) is named for its throughput of 160 MBps. N Ultra320 SCSI Ultra320 SCSI can move data at a rate of 320 MBps. N Ultra640 SCSI Another doubling of the SCSI interface speed, Ultra640 SCSI was promulgated as a new standard in early 2003. NOTE A storage connection technology, called Fibre Channel, can use either fiber-optic or copper cable, is a much more flexible connection scheme than SCSI, and promises throughput many times faster than even that of Ultra640 SCSI. Based loosely on a network paradigm, Fibre Channel is initially expensive to implement, but large data centers will benefit greatly from its advances over SCSI. As you can see from the preceding list, a dizzying array of SCSI choices is available on the market today. Because of all the different standards, it’s a good idea to make sure you purchase matched components when building a SCSI disk subsystem or when purchasing one as part of a server. Make sure the controller card you plan to use is compatible with the drives you will use, that the card uses the appropriate cables, and that it is compatible with both the server computer and the NOS you will use. The good news is that once you get a SCSI disk subsystem up and running, it will run reliably and with excellent performance. Disk Topologies: It’s a RAID! The acronym RAID stands for redundant array of independent disks. RAID is a technique of using many disks to do the work of one disk, and it offers many advantages compared to using fewer, larger disks. The basic idea behind RAID is to spread a server’s data across many disks, seamlessly. For example, a single file might have portions of itself spread across four or five disks. The RAID system manages all those parts so you never know they’re actually spread across all the disks. You open the file, the RAID system accesses all the appropriate disks and “reassembles” the file, and provides the entire file to you. The immediate benefit you get is that the multiple disks perform much more quickly than a single disk. This is because all the disks can independently work on finding their own data and sending it to the controller to be assembled. A single disk drive would be limited by a single disk head and would take much longer to gather the same amount of data. Amazingly, the performance of a RAID system increases as you add more disks, because of the benefit of having all those disk heads independently working toward retrieving the needed data. 183 Chapter 13: Network Servers: Everything You Wanted to Know but Were Afraid to Ask If you think about a simple RAID array with data spread across many disks, you’ll probably notice that, while it improves performance, it also increases the chance of a disk failure. Using five disks to do the work of one means that five times more chances exist for a disk failure. Because the data is spread among all the disks, if one fails, you might as well throw away all the data on all the remaining disks because it’s useless if a big chunk is missing. Fortunately, different RAID schemes address this problem. There are many different ways to use multiple disks together in some sort of RAID scheme and, accordingly, a number of RAID levels are defined, each of which describes a different technique, as follows: N RAID 0 This scheme is a configuration whereby data is spread (striped) across multiple disks, although with no redundancy. Losing one drive in a RAID 0 array results in the loss of data on all the disks. RAID 0 is appropriate only for improving performance and should be used only with nonessential data. RAID 0 arrays can stripe data across two or more disks, as shown in Figure 13-1. N RAID 1 This type of array doesn’t stripe data across multiple disks. Instead, it defines a standard whereby data is mirrored between disks. Two disks are used instead of one, and the data is kept synchronized between the two disks. If one of the disks fails, the remaining disk continues working just fine, until the failed drive can be replaced. RAID 1 is often simply referred to as mirroring. An enhancement to RAID 1 is called duplexing; the data is still duplicated between two disks, but each disk has its own disk controller, adding another level of redundancy (because you can lose either a disk or a controller and still keep operating). Duplexing can also improve performance somewhat, compared to straight mirroring. Some RAID 1 implementations are also intelligent enough to read data from either disk in such a way that whichever disk has its drive head closest to the data performs the read request, while the other one sits idle. However, all writes must occur simultaneously for both disks. Figure 13-2 shows a typical RAID 1 array layout. Figure 13-1. A RAID 0 array stripes data across multiple disks File Disk 1 Disk 3Disk 2 184 Networking: A Beginner’s Guide TIP You can combine RAID levels 0 and 1 to achieve the performance benefit of RAID 0 with the high level of redundancy of RAID 1. Imagine a series of RAID 1 arrays with two disks each. Combine each of these RAID 1 arrays so that data is striped across them, and you have what is called a RAID 10 array (with 10 referring to a combination of RAID 1 and RAID 0). This is sometimes also called RAID 0 + 1 or RAID 1 + 0. N RAID 2 You probably won’t see RAID 2 implemented in the real world. RAID 2 is a technical specification that stripes data across multiple disks and then uses a Hamming Code ECC that is written to a set of ECC disks. The ratio of ECC disks to data disks is quite high with RAID 2: There are three ECC disks for every four data disks. RAID 2 isn’t used because of its inefficiencies. N RAID 3 This is where RAID starts to get interesting. RAID 3 implementations used to be fairly common, although these days you see RAID 5 used much more often than RAID 3. RAID 3 stripes data across multiple data disks and then uses an exclusive OR (XOR) bit-wise operation against all the stored data on each data disk to come up with ECC data, which is written to a single ECC drive. So, for example, you can have four data drives and one ECC drive to back them up. Figure 13-3 shows a RAID 3 array. The XOR data has an interesting mathematical property. If you remove one of the data drives, you can take the remaining data, plus the data on the ECC drive, and reconstruct what is missing from the failed drive. RAID disk controllers do this automatically if a drive fails, although the drives operate at a slower rate than normal because of the overhead of having to reconstruct the data on the fly. A more useful technique is to replace the failed drive and then use the ECC data to rebuild the lost data. Figure 13-2. A RAID 1 array mirrors data between two disks File Disk 1 Disk 2 185 Chapter 13: Network Servers: Everything You Wanted to Know but Were Afraid to Ask NOTE If more than one drive is lost from a RAID 3 or a RAID 5 array, all the array’s data will be lost. Still, these arrays provide good protection at relatively low incremental cost. N RAID 4 This is another of the RAID standards that isn’t used very much in the real world. RAID 4 is similar to RAID 3, except data is striped between the different data drives in much larger blocks than with RAID 3. RAID 4 still uses a single ECC disk for all the data drives. N RAID 5 RAID 5, depicted in Figure 13-4, is the current standard for RAID systems. (RAID 1 also remains a current standard, but it has different applications.) Recall how RAID 3 worked, with data striped to a set of data disks, and the ECC code written to a single ECC disk. RAID 5 improves on this scheme by interleaving the data and ECC information across all the disks. The big advantage of this approach over RAID 3 is that it doesn’t rely on a single ECC drive for all write operations, which becomes a bottleneck on RAID 3 systems. Because all the drives share the ECC work, performance with RAID 5 is slightly better than with RAID 3. There is a small drawback to this, though, that most commentators miss. In RAID 3, if you lost a data drive, the system slowed down (usually dramatically) as the data was reconstructed on the fly. If you lost the ECC drive, however, the system would still run just as fast as if no drive were lost. With RAID 5, if you lose a drive, you’re always losing part of your ECC drive (because its job is spread among all the disks), so you get a slowdown no matter what. N RAID 6 RAID 6 works the same as RAID 5, but stores parity data on two interleaving drives rather than the one of RAID 5. Also called dual data guarding, RAID 6 keeps your data safe while you are recovering from a single drive failure. (In RAID 5, if a drive failed while you were recovering from another drive's failure, you would lose the array’s data.) Figure 13-3. A RAID-3 array stripes data across multiple disks, with an ECC disk to protect the data 1-A 1-B 1-C 1-D ECC Data striped on these drives Last disk contains ECC data for drives 1–4 Disk 1 Disk 2 Disk 3 Disk 4 Disk 5 186 Networking: A Beginner’s Guide Which level of RAID should you use on your network server? Most network administrators favor RAID 5 because it requires only 20 to 25 percent of the total disk capacity for the redundancy function, yet it performs well and offers a measure of safety. However, RAID 3 and RAID 5 arrays do occasionally fail to recover data properly (although they very rarely lose data). For this reason, you usually should opt for either RAID 1 or a RAID 10 array for network servers that store vital data. In general, the different RAID configurations offer different levels of reliability. Ranked from best to worst purely in terms of the system’s likelihood of losing data would be RAID 1, RAID 10, RAID 6, RAID 5, and RAID 3. There are always trade-offs, though. A system with 20 disks using just RAID 1 would be unwieldy to manage, because you would have 10 logical drives to manage and use efficiently. However, if you configured those same 20 disks as two RAID 5 arrays, you would be able to manage more efficiently the two logical disks that would result. You must make your own decision based on the importance of the data, the required levels of performance, the capabilities of the server, and the budget available to you. One thing you should never do, though, is trust that any RAID level replaces regular, tested, reliable tape backups of network data! Server State Monitoring An important feature of most servers is the capability to monitor its own internal components and to notify you if any problems develop or appear to be developing. Higher-end servers can typically monitor the following: N Proper fan operation N System voltage N Memory errors, even if corrected by ECC memory N Disk errors, even if corrected automatically Figure 13-4. A RAID 5 array stripes data across multiple disks, and alternately uses all disks for ECC data Disk 1 Disk 2 Disk 3 Disk 4 Disk 5 Data and ECC distributed to all drives 1-A 2-A 3-A 1-B 2-B 3-B 1-C 2-C ECC ECC 1-D 3-C ECC 2-D 3-D . matter what. N RAID 6 RAID 6 works the same as RAID 5, but stores parity data on two interleaving drives rather than the one of RAID 5. Also called dual data guarding, RAID 6 keeps your data safe. Combine each of these RAID 1 arrays so that data is striped across them, and you have what is called a RAID 10 array (with 10 referring to a combination of RAID 1 and RAID 0). This is sometimes also. often than RAID 3. RAID 3 stripes data across multiple data disks and then uses an exclusive OR (XOR) bit-wise operation against all the stored data on each data disk to come up with ECC data, which