Đang chuẩn bị nút TẢI XUỐNG, xin hãy chờ
Tải xuống
THE FRACTAL STRUCTURE OF DATA REFERENCE- P18:For purposes of understanding its performance, a computer system is traditionally viewed as a processor coupled to one or more disk storage devices, and driven by externally generated requests (typically called transactions). Over the past several decades, very powerful techniques have become available to the performance analyst attempting to understand, at a high level, the operational behavior of such systems. | Free Space Collection in a Log 73 must be moved i.e. read from one location and written back to another . For storage utilizations higher than 75 percent the number of moves per write increases rapidly and becomes unbounded as the utilization approaches 100 percent. The most important implication of 6.1 is that the utilization of storage should not be pushed much above the range of 80 to 85 percent full less any storage that must be set aside as a free space buffer. To put this in perspective it should be noted that traditional disk subsystems must also be managed so as to provide substantial amounts of free storage. Otherwise it would not be practical to allocate new files and increase the size of old ones on an as-needed basis. The amount of free space needed to ensure moderate freespace collection loads tends to be no more than that set aside in the case oftraditional disk storage management 32 . The final two sections ofthe chapter show in a nutshell that 6.1 continues to stand up as a reasonable rule of thumb even after accounting for a much more realistic model of the free space collection process than that initially presented to justify the equation. This is because to improve the realism ofthe model we we must take into account two effects 1. the impact of transient patterns of data reference within the workload and 2. the impact of algorithm improvements geared toward the presence of such patterns. Figure 6.1. Overview of free space collection results. 74 THE FRACTAL STRUCTURE OF DATA REFERENCE One section is devoted to each ofthese effects. As we shall show effects 1 and 2 work in opposite directions insofar as their impact on the key metric M is concerned. A reasonable objective for the algorithm improvements of 2 is to ensure a level of free space collection efficiency at least as good as that stated by 6.1 . Figure 6.1 illustrates impacts 1 and 2 and provides in effect a road map for the chapter. The heavy solid curve labeled linear model presents the
