| Oracle FAQ | Your Portal to the Oracle Knowledge Grid | |
 |
 | |||
| Oracle FAQ | Your Portal to the Oracle Knowledge Grid | |
|
| |||
Home -> Community -> Mailing Lists -> Oracle-L -> Re: oraperf comment
Yechiel,
You had mentioned only one possible scenario (i.e. "user A accesses table while user B simultaneously accesses index") where there are several other possible, equally-likely scenarios (i.e. "user A accesses table while user B simultaneously accesses table", "user A accesses index while user B simultaneously accesses index", etc). Separating tables and indexes to separate devices does nothing for those other, equally-likely scenarios, does it? That's the reason for the question "why?" in the beginning of my last reply...
At issue here is not the concept of parallelism in I/O. At issue (at least for me) is the "conventional wisdom" that states/implies that there is some performance benefit of separating tables and indexes to separate devices. My assertion is that this is irrelevant for two reasons: a) within a single process the accessing of table blocks and index blocks are purely sequential and b) tables and indexes have different I/O characteristics which make it less likely that they will conflict with each other. In fact, in most situations datafiles/tablespaces containing indexes generate far fewer physical I/Os than datafiles/tablespaces containing tables. From an I/O perspective, the key is not to focus on whether the datafile/tablespace contains tables or indexes but rather to focus on the volume and type of physical I/O they generate.
By focusing on the I/O statistics rather than whether they are tables or indexes, one can make better determinations on how to distribute I/O across non-RAID devices.
Hope this helps...
-Tim
I do not understand the WHY in the beginning.
I said that it is better to split according to the I/O load, but without more data, split between indexes and tables as a typical sql select will use both.
Yechiel Adar
Mehish
...resending, as the original send encountered some kind of "locking problem" at fatcity...
Why? What are the chances of precisely that scenario happening, as opposed to Oracle doing concurrent I/O to tables for both users A and B? Or to indexes for both users A and B simultaneously?
Splitting tables and indexes into separate tablespaces makes sense, but mainly for recovery purposes. This has little to do with the placement of the datafiles of those tablespaces on devices (non-RAID or RAID).
Generally, indexes tend to cache extremely well in Oracle (because they are more compact and because of the nature of the I/O), so they usually don't get as much physical I/O as tables. Check V$FILESTAT on a busy application to prove it for yourself...
After seeing this performance data, why would you place a datafile/tablespace which only gets a small amount of I/O on one device while placing a much busier datafile/tablespace onto another device, just because one contains indexes and the other tables?
Please think in terms of I/O counts, not poorly-conceived but oft-repeated "conventional wisdom". Keep indexes and tables segregated to different tablespaces, but for decisions on placement of datafiles upon devices, use empirical performance data only.
Hello Tim
I beg to differ. Without raid it is better to put indexes and tables on different disks and controllers.
This way Oracle can do I/O to a table for user A while doing I/O to the index for user B.
It is better if you can find the high I/O areas of the database and split them across disks, but as a rule of thumb splitting indexes and tables make sense (again - when you work without raid).
Yechiel Adar
Mehish
----- Original Message -----
From: Tim Gorman
To: Multiple recipients of list ORACLE-L
Sent: Tuesday, October 22, 2002 12:39 AM
Subject: Re: oraperf comment
Ray,
I don't know exactly what was intended with the comment, but I agree with your interpretation.
---
As far as any other reasons for the comment...
<RANT>
In terms of myths that have persisted with Oracle over the years, the idea that some performance benefit exists from I/O parallelism due to separating tables and indexes to different devices has been especially persistent. I've even heard it described as "conventional wisdom". As a matter of fact, there is no possibility for "parallelism" benefits on indexed I/O operations. Never has been; might never be (though "never" is a long time)...
</RANT>
The reason is that navigating a B*Tree index structure is inherently sequential. Think about it -- first you have to access the "root" block. Looking inside the contents of the "root" directs you to the next "branch" or "leaf" block in the index B*Tree structure. You cannot seek for the next block in parallel; you've got to look inside one block in order to know what block to access next. Then, once you've accessed down to the final "leaf" block, reading its contents tells you which row in the table to access. If you are doing a "range scan" operation, then you have to go back to the index "leaf" block in order to find the next table row to access.
The name of the wait-event for this type of I/O (a.k.a. "db file sequential read", a.k.a. single-block random-access read) also suggests this "sequentialiality" (is that a word?). Jeff Holt wrote a great paper on the reasons for the apparent mis-naming of the wait-events "db file sequential read" and "db file scattered read" -- I'm sure that it is downloadable from http://www.hotsos.com. Even when "asynchronous I/O" is available and configured, indexed I/O operations are still essentially synchronous (and non-parallel)...
There is a possibility of some form of "parallelization" in "range-scan" operations, but there is no evidence that this is happening. For example, while performing an indexed range-scan, if we wanted to read a batch of index entries from the index "leaf blocks" and submit a list of I/O requests for data blocks on the corresponding table, we could do so. However, when I've performed "truss" operations on an Oracle server process performing such a range-scan operation (at least through Oracle8i), I've not seen this happening. Purely generic "read()" operations, one at a time, sequentially...
---
The only real advantages of separating tables from indexes into different tablespaces are:
a.. different recoverability requirements
a.. indexes can be rebuilt instead of restored
b.. data (tables and clusters) must be restored -- cannot be "rebuilt" from anything
b.. different types of I/O requests
a.. indexes are predominantly accessed using single-block, random read I/O (i.e. UNIQUE scans, RANGE scans, FULL scans)
a.. relatively seldom are accessed with multi-block sequentially-accessed read I/O (i.e. FAST FULL scans)
b.. while tables are often accessed with a mix of the two types of I/O, depending on the application
a.. OLTP usually has heavier single-block, random read I/O due to heavy use of indexes
b.. DW usually has heavier multi-block, sequentially-accessed read I/O due to heavy use of FULL table scans
c.. may be advantages from this in Oracle9i where different blocksizes are possible for different tablespaces
These last points are related to performance, but not in the sense that the mythical "conventional wisdom" dictates...
Hope this helps...
-Tim
----- Original Message -----
From: "Ray Stell" <stellr_at_cns.vt.edu>
To: "Multiple recipients of list ORACLE-L" <ORACLE-L_at_fatcity.com>
Sent: Monday, October 21, 2002 2:43 PM
Subject: oraperf comment
>
> An recent oraperf report included the comment: "Never split index
> and data files to different sets of disks." It goes on to state that
> striping is better. If the system in question does not have
> raid support, wouldn't it be better to split the index and data across
> spindles? That would make the word "Never" inappropriate here? Maybe
> this is their way of saying don't use old technology. Is there some
> other reason I am missing?
> ===============================================================
> Ray Stell stellr_at_vt.edu (540) 231-4109 KE4TJC 28^D
> --
> Please see the official ORACLE-L FAQ: http://www.orafaq.com
> --
> Author: Ray Stell
> INET: stellr_at_cns.vt.edu
>
> Fat City Network Services -- 858-538-5051 http://www.fatcity.com
> San Diego, California -- Mailing list and web hosting services
> ---------------------------------------------------------------------
> To REMOVE yourself from this mailing list, send an E-Mail message
> to: ListGuru_at_fatcity.com (note EXACT spelling of 'ListGuru') and in
> the message BODY, include a line containing: UNSUB ORACLE-L
> (or the name of mailing list you want to be removed from). You may
> also send the HELP command for other information (like subscribing).
-- Please see the official ORACLE-L FAQ: http://www.orafaq.com -- Author: Tim Gorman INET: Tim_at_SageLogix.com Fat City Network Services -- 858-538-5051 http://www.fatcity.com San Diego, California -- Mailing list and web hosting services --------------------------------------------------------------------- To REMOVE yourself from this mailing list, send an E-Mail message to: ListGuru_at_fatcity.com (note EXACT spelling of 'ListGuru') and in the message BODY, include a line containing: UNSUB ORACLE-L (or the name of mailing list you want to be removed from). You may also send the HELP command for other information (like subscribing).Received on Tue Oct 22 2002 - 15:00:01 CDT
 |
 |