The following message is a courtesy copy of an article that has been posted to bit.listserv.ibm-main,alt.folklore.computers as well.
re: http://www.garlic.com/~lynn/2009p.html#12 Secret Service plans IT reboot there use to be a joke about TSO users not realizing how deplorable performance was because they couldn't see the difference by operating with & w/o MVS (actually in large part, CKD & multi-track search). CKD & multi-track search introduced with original 360 was scarce resource use trade-off of the period ... by the mid-70s, the relative amounts of resources had nearly inverted (which resources were the scarcest), starting to make multi-track search the exact wrong thing to do, there was a large national retail operation with a consolidated datacenter (large number of systems in loosely-coupled configuration) ... which started to run into severe throughput problem during peak periods. This went on for awhile, lots of experts being brought in over period of time, until they eventualy got around to calling me in. I was brought into a class room with large number of long class tables ... covered with high stacks of paper performance details from all the systems. while i started to leaf through all the pages (for shared disk activity, I had to aggregate drive activity from different systems/reports in my head ... while they started through overall summary of the symptoms). After about 20-25 minutes ... I started to notice a somewhat anomolous circumstance ... about the only correlation between "good" thruput and "nearly no thruput" was a specific pack had aggregated i/o counts between 6 and 7 during high-load/low-throughput (which would seem to hardly be a thruput limitation). After a little more investigation ... it turned out, the pack contained the shared application library for the whole complex ... more investigation was that the PDS had a three cylinder PDS directory. Back of the envelope calculations was that avg. depth of search was cylinder and half (PDS member lookup) ... that would be two multi-track search I/Os that took elapsed time of nearly 1/2 second. Assumption then was the two PDS directory lookup I/Os would be followed by a single I/O for a PDS member load. That accounts for aggregate of six I/Os per second saturating the drive ... basically limiting the whole national loosely-coupled infrastructure to performing an aggregate of two application (PDS) program library loads per second. Each full-cylinder multi-track search represented enormous busy elapsed time for the processor channel (locking out any other activity on the same channel). The full-cylinder multi-track searches also locked up the (shared) controller, string and drive ... locking out all systems from accessing anything else associated with those resources. The eventual result was reconfiguring everything to try and come as close as possible to eliminating the long multi-track searches (drastically reduced PDS directory size) ... and replicating the shared application library on non-shared drives for each system. PDS directory (& vtoc) multi-track searches alleviated needing the real storage to contain the directory information (at enormous cost in I/O resources). By the mid-70s, real storage was becoming plentiful enough that it was practical to keep high-useage (vtoc &) PDS directory information cached in system storage (allowing fast lookup of instorage index) ... so program loads could happen at "normal" disk activity thruput speeds (say 30-50/second) ... instead of at 2/second (limited by the enormous PDS directory multi-track search penalty). This resource trade-off also showed up with RDBMS ... the original relational/sql was done on vm system in bldg. 28 ... system/r ... misc. past posts: http://www.garlic.com/~lynn/subtopic.html#systemr In the 70s, there was somewhat rivalry between the IMS group in STL and system/r in bldg. 28 on the main plant site. IMS group claimed better trade-offs because record pointers were exposed as part of the data ... and it was possible to go directly to a specific piece of data. This was contrasted with RDBMS implementation that had an implicit index ... which could take 4-5 disk i/os to eventually find the location of the desired data record. This implicit index also tended to double the physical disk space required (vis-a-vis same data in IMS). The system/r group countered that the exposed record pointers created a significant administrative and maintanance overhead ... especially for adding data ... nearly eliminated by the implicit indexes). The resource trade-offs argument changed with combination of enormous disk size increases and drastic fall in cost/mbyte (muting the issue regarding doubling disk space for the indexes). At the same time there was significant increase in available system real storage ... making it practical to cache a large portion of the (implicit) RDBMS indexes (drastically reducing separate physical disk i/os to find data record). -- 40+yrs virtualization experience (since Jan68), online at home since Mar1970 ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@bama.ua.edu with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html