Re: [Q] Cry for help: how to make exp faster?

index=y index=n - the time saving will be marginal, since the only thing exported is the definition of the index, not the index itself.

The difference in time is more likely to be the switch from direct=y to direct=n - which saves a lot of CPU because it bypasses various format conversion routines (even so 25%/33% seems a big difference).

Since you are exporting to tape through a pipe, have you tried putting 'compress' between the export and the tape ? I doubt if it will make things go faster since your tape probably does hardware compression anyway - but it's worth a try.

The other option is to run multiple exports: do a rows=n full export first to get the structure, then run one export per relevant user with rows=y. If you have multiple tapes you can then run concurrent tape copies to save time.

BTW - compress=y is a bad idea - it can cause ridiculous space management problems at import time if the object definitions are not already in place.

I doubt, also, that any buffer size beyond a couple of MB helps very much.

One of the standard ways to improve the speed of export, which has been documented at times by Oracle, but is also frowned upon, is to create a single task version of exp, as this eliminates the 2K pipe between the exp process and its shadow. Look in $ORACLE_HOME/rdbms/lib/ins_rdbms.mk for 'expst'.

--

Jonathan Lewis
Yet another Oracle-related web site: http://www.jlcomp.demon.co.uk

Igor P. Merku wrote in message <37B269D3.EBF95A7_at_acegas.ts.it>...
>Hello,
>we are running an Oracle 7.3.4.4.0 on raw device on RS/6000 AIX
>V4.3.2.-02 (node has 4 CPU at 332 MHz), disk subsystem is 7133-020 SSA
>technology (disks are 9.1 GB).
>I have written a procedure that makes backup of Oracle in "raw mode",
>means saving the data files / raw logical volumes, which works all right
>and, of course, runs in always the same time unless data files would be
>added.
>
>The application/database is not completely in production yet as we are
>converting data from "old" mainframe system to new Unix system in
>cycles. We have still some months to go...
>In the meantime, logical export seems to be a good way (seemed) to save
>data for the development team. In the first place I saved data via pipe
>immediately onto tape drive (IBM 3590) which was all right for the first
>converted cycle as it took up about 45 min. Now, with four cycles and,
>to improve performance, exporting to a file on file system (jfs) for 2,3
>GB of resulting export file it takes up about 6 hours (tape took about
>7,5 hours).
>Monday I used direct=Y, compress=Y and index=N options (6 hours).
>Yesterday I used direct=N, compress=Y and index=Y options (8 hours). As
>I see from resulting export file index=N or Y seems not to make really
>difference as speaking of export file length, but as I saw makes a
>difference in time.
>Buffer for export is set to about 75 MB, I could afford some more, I
>guess. Maybe a renice could help some?
Received on Thu Aug 12 1999 - 09:16:17 CDT