A second potential problem is due to the communication network. It can lose messages, which requires special software to be able to recover, and it can become overloaded. When the network saturates, it must either be replaced or a second one must be added. In both cases, some portion of one or more buildings may have to be rewired at great expense, or network interface boards may have to be replaced (e.g., by fiber optics). Once the system comes to depend on the network, its loss or saturation can negate most of the advantages the distributed system was built to achieve.. In Fig. 1-7 we see a bus-based multicomputer. It looks topologically similar to the bus-based multiprocessor, but since there will be much less traffic over it, it need not be a high-speed backplane bus. In fact, it can be a much lower speed LAN (typically, 10-100 Mbps, compared to 300 Mbps and up for a backplane bus). Thus Fig. 1-7 is more often a collection of workstations on a LAN than a collection of CPU cards inserted into a fast bus (although the latter configuration is definitely a possible design).. Because none of the CPUs have local memory and all programs are stored in the global shared memory, it does not matter on which CPU a process runs. If a long-running process is scheduled many times before it completes, on the average, it will spend about the same amount of time running on each CPU. The only factor that has any effect at all on CPU choice is the slight gain in performance when a process starts up on a CPU that is currently caching part of its address space. In other words, if all CPUs are idle, waiting for I/O, and one process becomes ready, it is slightly preferable to allocate it to the CPU it was last using, assuming that no other process has used that CPU since (Vaswani and Zahorjan, 1991).. Ousterhout also described several variations to this basic method to improve performance. One of these breaks the matrix into rows and concatenates the rows to form one long row. Withk processors, anyk consecutive slots belong to different processors. To allocate a new process group to slots, one lays a windowk slots wide over the long row such that the leftmost slot is empty but the slot just outside the left edge of the window is full. If sufficient empty slots are present in the window, the processes are assigned to the empty slots; otherwise, the window is slid to the right and the algorithm repeated. Scheduling is done by starting the window at the left edge and moving rightward by about one window’s worth per time slice, taking care not to split groups over windows. Ousterhout’s paper discusses these and other methods in more detail and gives some performance results.. On the other hand, if the processors exhibit Byzantine failures, continuing to run when sick and sending out erroneous or random replies, a minimum of2k+1 processors are needed to achievek fault tolerance. In the worst case, thek failing processors could accidentally (or even intentionally) generate the same reply. However, the remainingk+1 will also produce the same answer, so the client or voter can just believe the majority.. The consequences affect both hardware and software. Let us first look at the implications for the hardware. Designing a machine in which many processors use the same memory simultaneously is surprisingly difficult. Bus-based multiprocessors, as described in Sec. 1.3.1, cannot be used with more than a few dozen processors because the bus tends to become a bottleneck. Switched multiprocessors, as described in Sec. 1.3.2, can be made to scale to large systems, but they are relatively expensive, slow, complex, and difficult to maintain.. The first two points are self explanatory. The third may require some clarification. In Dash and most other modern UMA multiprocessors, remote access is slower than local access as well. What makes this property bearable is the presence of caching. When a remote word is touched, a block of memory around it is fetched to the requesting processor’s cache, so that subsequent references go at full speed. Although there is a slight delay to handle the cache fault, running out of remote memory can be only fractionally more expensive than running out of local memory. The consequence of this observation is that it does not matter so much which pages live in which memory: code and data are automatically moved by the hardware to wherever they are needed (although a bad choice of the home cluster for each page in Dash adds extra overhead).. As a more realistic example, one could imagine a system to provide sports fans with up-to-the-minute (but perhaps not up-to-the-nanosecond) scores for sporting events worldwide. Answering a query as if it had been made 2 nanoseconds earlier or later might well be acceptable in this case, especially if it gave much better performance by allowing multiple copies of the data to be stored. In this case strict consistency is not promised, delivered, or needed.. Fig. 6-23. A valid event sequence for release consistency.. In the next three cases, some other process is the owner at the timeP does the write. In all three cases,P must ask the current owner to invalidate any existing copies, pass ownership toP, and send a copy of the page unlessP already has a copy. Only then may the write take place. In all three cases,P ends up with the only copy of the page, which is inW state.. Originally, Munin supported some other categories as well, but experience showed them to be of only marginal value, so they were dropped. Each machine maintains a directory listing each variable, telling, among other things, which category it belongs to. For each category, a different protocol is used.. When the release is done, Munin runs a word-by-word comparison of each dirty write-shared page with its twin, and sends the differences (along with all the migratory pages) to all processes needing them. It then resets the page protection to read only.. Finally, the process descriptor also contains a descriptor for each thread in the process. The content of a thread descriptor is architecture dependent, but as a bare minimum, it contains the thread’s program counter and stack pointer. It may also contain additional information necessary to run the thread, including other registers, the thread’s state, and various flags. Brand new processes contain only one thread in their process descriptors, but stunned processes may have created additional threads before being stunned.. 9.2. PROCESS MANAGEMENT IN CHORUS. [Картинка: any2fbimgloader250].