Parallel Systems

Parallel Systems

Most systems to date are single-processor systems; that is, they have onlyone main CPU. However, there is a trend toward multiprocessor systems. Such Systems have more than one processor in close communication, sharing the computer bus, the clock, and sometimes memory and peripheral devices. These systems are referred to as tightly coupled systems.

There are several reasons for building such systems. One advantage is increased throughput. By increasing the number of processors, we hope to get more work done in a shorter period of time. The speed-up ratio with n processors is not n, however, but rather is less than n. When multiple processors cooperate on a task, a certain amount of overhead is incurred in keeping all the parts working correctly. This overhead, plus contention for shared resources, lowers the expected gain from additional processors. Similarly, a group of n programmers working closely together does not result in n times the amount of work being accomplished.

Multiprocessors can also save money compared to multiple single systems because the processors can share peripherals, cabinets, and power supplies. If several programs are to operate on the same set of data, it is cheaper to store those data on one disk and to have all the processors share them, rather than to have many computers with local disks and many copies of the data.

Another reason for multiprocessor systems is that they increase reliability. If functions can be distributed properly among several proces35rs7then' the failure of one processor will not halt the system, but rather will only slow it down. If we have 10 processors and one fails, then each of the remaining nine processors must pick up a share of the work of the failed processor. Thus, the entire system runs only 10 percent slower, rather than failing altogether. This ability to continue providing service proportional to the level of surviving hardware is called graceful degradation. Systems that are designed for graceful degradation are also called fault-tolerant.

Continued operation in the presence of failures requires a mechanism to allow the failure to be detected, diagnosed, and corrected (if possible). The Tandem system uses both hardware and software duplication to ensure contin­ued operation despite faults. The system consists of two identical processors, each with its own local memory. The processors are connected by a bus. One processor is the primary, and the other is the backup. Two copies are kept of each process; one on the primary machine and the other on the backup. At fixed checkpoints in the execution of the system, the state information of each job (including a copy of the memory image) is copied from the primary machine to the backup. If a failure is detected, the backup copy is activated, and is restarted from the most recent checkpoint. This solution is obviously an expensive one, since there is considerable hardware duplication.

The most common multiple-processor systems now use the symmetric-multiprocessing model, in which each processor runs, an identical copy of. the operating system, and these copies communicate with one another as needed. Some systems use asymmetric multiprocessing, in which each proces­sor is assigned a specific task. A master processor controls the system; the other processors either look to the master for instruction or have predefined tasks. This scheme defines a master-slave relationship. The master processor schedules and allocates work to the slave processors.)

An example of the symmetric multiprocessing system is Encore's version of UNIX for the Multimax computer. This computer can be configured to employ dozens of processors, all running a copy of UNIX. The benefit of this model is that many processes can run at once (N processes if there are N CPUs) without causing a deterioration of performance. However, we must carefully control I/O to ensure that data reach the appropriate processor. Also, since the CPUs are separate, one may be sitting idle while another is overloaded, resulting in inefficiencies. To avoid these inefficiencies, the processors can share certain data structures. A multiprocessor system of this form will allow jobs and resources to be shared dynamically among the various processors, and can lower the variance among the systems.

Asymmetric multiprocessing is more common in extremely large systems, where one of the most time-consuming activities is simply processing I/O. In older batch systems, small processors, located at some distance from the main CPU, were used to run card readers and line printers and to transfer these jobs to and from the main computer. These locations are called remote-job-entry (RJE) sites. In a time-sharing system, a main I/O activity is processing the I/O of characters between the terminals and the computer. If the main CPU must be interrupted for every character for every terminal, it may spend all its time simply processing characters. So that this situation is avoided, most systems have a separate front-end processor that handles the entire terminal I/O. For example, a large IBM system might use an IBM Series/l minicomputer as a front-end. The front-end acts as a buffer between the terminals and the main CPU, allowing the main CPU to handle lines and blocks of characters, instead of individual characters. Such systems suffer from decreased reliability through increased specialization.

It is important to recognize that the difference between symmetric and asymmetric multiprocessing may be the result of either hardware or software. Special hardware may exist to differentiate the multiple processors, or the soft­ware may be written to allow only one master and multiple slaves. For instance, Sun's operating system SunOS Version 4 provides asymmetric multiprocessing, whereas Version 5 (Solaris 2) is symmetric.

As microprocessors become less expensive and more powerful, additional operating-system functions are off-loaded to slave processors, or back-ends. For example, it is fairly easy to add a microprocessor with its own memory to manage a disk system. The microprocessor could receive a sequence of requests from the main CPU and implement its own disk queue and scheduling algorithm. This arrangement relieves the main CPU of the overhead of disk scheduling. PCs contain a microprocessor in the keyboard to convert the key strokes into codes to be sent to the CPU. In fact, this use of microprocessors has become so common that it is no longer considered multiprocessing.