| WebIV:View NOTE:159898.1 |   |
 |
 |
 |
Article-ID: <Note:159898.1> Alias: VIEW:X$KCBKPFS Circulation: PUBLISHED (INTERNAL) ***Oracle Confidential - Internal Use Only*** Folder: server.Internals.General Topic: ** X$ Table Definitions Title: VIEW: X$KCBKPFS - PreFetch Statistics - (9.0) Document-Type: REFERENCE Impact: LOW Skill-Level: NOVICE Server-Version: 09.00 Updated-Date: 05-OCT-2001 05:27:02 References: Shared-Refs: Authors: MEPRESTO.US Attachments: NONE Content-Type: TEXT/PLAIN Products: 5/RDBMS (9.0); View: X$KCBKPFS [K]ernal [C]ache [B]uffer chec[K]point management [P]re[F]etch [S]tats Column Type Description -------- ---- -------- ADDR RAW(4|8) address of this row/entry in the array or SGA INDX NUMBER index number of this row in the fixed table array INST_ID NUMBER oracle instance number BUFFER_POOL_ID NUMBER Buffer Pool TIMESTAMP NUMBER Timestamp PREFETCH_OPS NUMBER number of prefetch operations PREFETCH_BLOCKS NUMBER number of blocks prefetched WASTED_BLOCKS NUMBER number of prefetched blocks wasted CLIENTS_PREFETCH NUMBER number of clients actually prefetching buffers PREFETCH_LIMIT NUMBER Limit to be used by for each prefetch operation Notes: This is maintained by the CKPT process which fills in an entry for each timeout period. Size of prefech history is limited to 50 per buffer pool. This section deals with controlling prefetching through the cache. This is used to determine whether prefetch is wasteful i.e are prefetched blocks being aged out of the cache before they can be used? This can happen, for example, if there is a lot of recycling activity in the cache and each client of the cache prefetched a large number of blocks into the cache. To prevent this the following algorithm is used to limit the amount of prefetching done in each buffer pool: A history buffer is maintained per buffer pool. This buffer maintains a history of prefetching performance. This is populated through a timeout action. The history buffer contains the following information in each entry (struct kcbkpfs): - Timestamp at which entry was created. - Number of prefetch operations since the last timeout action. - Number of blocks prefetched in this period. - Number of prefetched blocks that were wasted - this refers to blocks that were prefetched but had to be aged out before they could be pinned. - Prefetch limit (this is the value computed by this function) - Number of prefetching clients (as a snapshot at the time this function is executed) - Number of buffers being read at this time. The algorithm uses the history to determine the prefetch limit. It the last few history entries (defined by KCBK_HIST_WINDOW) and then computes the cumulative number of prefetch operations, prefetched blocks and wasted prefetch blocks over this window. It then applies the following rules to adjust the limit: - If there are no prefetched blocks, then set the limit to Q/C where Q is the prefetch quota (in number of buffers) and C is the number of clients performing prefetching - If there are prefetched blocks and some of them were wasted. In this case, the limit is reduced by the fraction (P - W)/P where P is the number of blocks prefetched over the history window and W is the number of prefetched blocks that were wasted. If W happens to be greater than P, this implies that the buffers that were prefetched before the history window were wasted in this time interval. In this case, we reduce the limit to half its value. - If there are no wasted prefetch blocks, then there are 3 cases: (a) The number of clients has gone down - in this case we double the prefetch limit. If this causes any wasted prefetches it will be reduced by that fraction in the next timeout. Note that the increase is limited by the ratio (Q/C') where Q is the prefetch quota and C' is the new number of clients. (b) The number of clients has increased - retain the limit as long as the it is less than (Q/C') (same ratio as above). If not, set it to Q/C' (c) The number of clients remains the same - The limit is doubled if the number of prefetched buffers has decreased by atleast 25% (again subject to the Q/C limit), otherwise the limit is retained. The rationale is that if the number of prefetched buffers has gone down, the clients have reduced their prefetching, so increasing the limit may not lead to wasted prefetches. Note that if the increase in the limit leads to wasted prefetching, then at the next timeout, the limit will be reduced by the fraction of the prefetched blocks that were wasted.
