Files in this item
|(no description provided)|
|Title:||A Study of the Interactions Between Operating System Memory Management and Database Buffer Management Strategies|
|Doctoral Committee Chair(s):||Belford, Geneva G.|
|Department / Program:||Computer Science|
|Degree Granting Institution:||University of Illinois at Urbana-Champaign|
|Abstract:||Database management systems (DBMS) are usually built "on top of" operating systems (OS), using the memory management facilities that those systems provide. But in addition, for efficiency of data retrieval and manipulation, the DBMS generally provides its own data buffering. The DBMS designer will often choose to put the DBMS-managed buffer pool in user space to save on overhead.
The implication of these facts is that in a virtual memory system the DBMS buffer pool is likely to be paged like any other part of user space. The buffer replacement strategy implemented by the DBMS will in that case be affected by the page replacement strategy implemented by the OS. Unfortunately, there is insufficient guidance in the literature as to how buffering and paging strategies might affect each other; or indeed as to whether there exists an effect large enough for DBMS and OS designers to worry about. Accordingly, a thorough simulation study was undertaken to answer these questions.
Our study showed that performance of a DBMS is affected by both paging and buffering algorithms; the performance of a specific combination results from a kind of "addition" of effects of the two algorithms. Outside of providing a warning against LFU, our simulations did not yield any definitive choice for buffering/paging combination nor did it turn up any combinations that worked poorly together. Because most DBMSs run on top of a given operating system, it is also important to study efficient buffer management under the paging algorithm of the given operating system. It is recommended that, instead of prefetching pages at the time of a buffer hit, pages should only be tag-along fetched at the time of a demand fetch.
For more investigation, we recommend that experiments be done with working DBMSs to verify/implement the results. There are neither experimental run length distributions nor empirical data as to whether or not there is locality in relational-type reference strings and it is strongly suggested that such an investigation be made. The results shown in this thesis will be helpful in designing/implementing an "intelligent" buffer manager.
Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 1988.
|Date Available in IDEALS:||2014-12-15|