THE FRACTAL STRUCTURE OF DATA REFERENCE- P22 pps

94 THE FRACTAL STRUCTURE OF DATA REFERENCE Figure 7.8. Persistent cylinder images as a function ofwindow size. Figure 7.9. Persistent file storage as a function ofwindow size. Transient and Persistent Data Access 95 Figure 7.10. Requests to persistent track images as a function of window size. Figure 7.11. Requests to persistent cylinder images as a function of window size. 96 THE FRACTAL STRUCTURE OF DATA REFERENCE Figure 7.12. Requests to persistent files as a function ofwindow size. and file storage that met the persistence criterion stated by (7.3). It should be noted that the percentage of persistent data depends even more strongly on the level of granularity, than does the percentage of active data. For track images, percentages in the range of 10 - 20 percent of active data were found to be persistent in a window of 24 hours; for files, the corresponding percentages were in the range of 50 to 75 percent. The phenomenon of persistence appears to be particularly important at the file level of granularity. Figures 7.10 through 7.12 present the amount of I/O associated with the persistent data just discussed. Again, we see that persistence is increasingly important at higher levels of granularity. At both of the installations presented in the figure, 90 percent or more of the I/O over a period of 24 hours was associated with persistent files. The results of Figures 7.10 through 7.12 provide a strong confirmation that I/O tuning is worth - while, despite the large fluctuations of load typical of measurements taken at different times or on different days. A substantial fraction of all files do exhibit persistent activity, and those that do tend to be the ones that dominate the overall I/O load. 3. We now focus strictly on file activity, as observed using the OS/390 System Measurement Facility ( SMF). The use of SMF, rather than I/O tracing, allows much longer time windows to be analyzed. The results of this section are based mainly on the file open/close event traces contained in the SMF record types PERIODS UP TO ONE MONTH Transient and Persistent Data Access 97 14, 15, 62, and 64, plus miscellaneous other records relating to file creates, renames, and deletes. Due to the use of this source of data, I/O activity is accounted for based upon the software - supplied EXecute Channel Program ( EXCP) counts, as placed into the SMF records just mentioned. This section presents the results of a study in which SMF data over a period of one month was obtained at two OS/390 installations. These were: C. A moderate - sized installation with a mix of on - line CICS, IMS, and DB2 database activity, plus TSO. D. A large installation with on - line DB2 database, batch, and TSO activity. Both installations had adopted active policies for Hierarchical Storage Man - agement (HSM). At both installations, general - purpose (primary) disk storage contained, for the most part, only files referenced within the relatively recent past. The policies for the management of the remaining files, administered via System Managed Storage (SMS), called for SMS to migrate unused data, first to a compressed disk storage archive, then to tape after a further period of non - use. SMS would also recall such unused data, back to primary storage, on an as - needed basis. To a remarkable degree, the files opened at both installations tended to be ones that had previously been open in the very recent past. Nevertheless, long gaps between open requests to a given file also occurred with a substantial probability. Figure 7.13 presents the resulting distribution of file interarrival times at each installation. Figure 7.13. Distribution offile interarrival times, based upon open requests. 98 THE FRACTAL STRUCTURE OF DATA REFERENCE It should not be surprising that the curves for both installations exhibit heavy - tailed behavior. Both curves appear to conform reasonably well to a mathematical model of the form (1.4), in that both resemble a straight line when plotted in a log - log format. Very few requests (about .5 percent at installation C, for example) ask for data that have not been used for five days or longer. Thus, Figure 7.13 suggests that an HSM policy calling for migration of unused data after one week or more would have a good chance of being acceptable from the standpoint of application performance. Figure 7.14 presents the average amount of storage associated with files that were active during various windows of time, ranging from about 15 hours up to 31 days. In cases where a file was created or scratched during a given study window, Figure 7.14 includes only the file’s storage while allocated. Figure 7.14 is adjusted, however, to ignore the impact of storage manage - ment. For example, if a file was migrated to tape during a given study window, then this action has no effect on the storage demand accounted for by the figure. To understand the implications of Figure 7.14, it is useful to think through what the figure would look like in several specific examples: 1. A collection of static, continuously active files. In this case, the figure would be a straight, horizontal line. 2. A series of transient files which are created at random times, referenced at the time that they are created, not referenced afterward, and never scratched. The longer such files are allowed to accumulate (the more generous we are Figure 7.14. Average active file storage over periods up to one month. . noted that the percentage of persistent data depends even more strongly on the level of granularity, than does the percentage of active data. For track images, percentages in the range of 10 - 20. levels of granularity. At both of the installations presented in the figure, 90 percent or more of the I/O over a period of 24 hours was associated with persistent files. The results of Figures. 10 - 20 percent of active data were found to be persistent in a window of 24 hours; for files, the corresponding percentages were in the range of 50 to 75 percent. The phenomenon of persistence