|
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
----- Original Message -----
Sent: Tuesday, October 22, 2002 10:09
AM
Subject: Re: oraperf comment
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
----- Original Message -----
Sent: Tuesday, October 22, 2002 5:14
PM
Subject: Fw: oraperf comment
...resending, as the original send encountered
some kind of "locking problem" at fatcity...
----- Original Message -----
Sent: Tuesday, October 22, 2002 6:35 AM
Subject: Re: oraperf comment
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.
----- Original Message -----
Sent: Tuesday, October 22, 2002 3:43
AM
Subject: Re: oraperf comment
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 -----
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:
- different recoverability
requirements
- indexes can be rebuilt instead of
restored
- data (tables and
clusters) must be restored -- cannot be "rebuilt" from
anything
- different types of I/O requests
- indexes are predominantly accessed using
single-block, random read I/O (i.e. UNIQUE scans, RANGE scans, FULL
scans)
- relatively seldom are accessed with
multi-block sequentially-accessed read I/O (i.e. FAST FULL
scans)
- while tables are often accessed with a
mix of the two types of I/O, depending on the application
- OLTP usually has heavier single-block,
random read I/O due to heavy use of indexes
- DW usually has heavier multi-block,
sequentially-accessed read I/O due to heavy use of FULL table
scans
- 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 -----
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 [EMAIL PROTECTED] (540) 231-4109
KE4TJC 28^D
> --
> Please see the
official ORACLE-L FAQ: http://www.orafaq.com
> --
> Author: Ray Stell
> INET: [EMAIL PROTECTED]
>
> 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: [EMAIL PROTECTED] (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).
|
