Re: NT Raw partitions seem to provide no benefit

It appears that for the specific configuration and test procedure (outlined in the previous post) that there is no difference in the performance of "raw" and "cooked" partitions. But then again, I wouldn't expect these specific tests to produce any measurable difference, since a "raw" partition doesn't make a disk spin any faster, doesn't reduce the head access time, and doesn't increase the available bandwidth of the hard drive or the i/o channel.

> I don't see why anyone would bother implementing this
> scheme on production servers.

It's possible that some production servers exhibit i/o characteristics significantly different than those of your test system and procedure. Deep in the bowels of the Oracle test labs, the engineers and technicians are busy tweaking systems to wring every bit of horsepower from the database engine, readying systems for the next race through the transaction benchmark tests. (How closely any of these transaction benchmarks compare to any "real world" workloads is a topic we can leave for others to debate in another discussion.)

Suffice it to say that, for a high volume workload, the i/o subsytem can become a significant bottleneck in the overall performance of the database system. By using "raw" partitions, the database bypasses an unnecessary source of additional overhead.

On the other hand, it is likely that a large number of "real world" production servers do _not_ experience the type of transaction loads such that the overhead imposed by the file system significantly degrades performance.

"Pat Boivin" <lori.pat_at_ns.sympatico.ca> wrote in message news:3925DBBD.71533F25_at_ns.sympatico.ca...
> Hi,
>
> I just created a couple of partitions on an NT server (NT 4
 SP5; Pentium
> III
> 350MHz, IDE 6.4G drive). The database is in noarchivelog
 mode, RDBMS
> 7.3.4.5.2. for NT. This is a test machine.
> Both partitions are identical in size, and on the same
 physical disk.
> Both
> partitions are 15M in size, one is a raw partition, one is an
 ntfs
> partition.
> Now.
> I put the same table into each of the two partitions, 188,266
 rows.
> When I do a full table scan of the table, the result is the
 same every
> time:
> 20 seconds on the ntfs partition, 21 seconds on the raw
 partition. I
> created an index and did a select * from ... which would make
 Oracle
> ignore
> the index anyway, this time I got 20 seconds for each
 partition. Then I
> did
> select indexed_column from ... with no where clause. Again it
 should
> just
> do a full table scan, I got 20 seconds for each partition.
 Then I did a
>
> select indexed_column from ... with a where clause containing
> temperature/2
> = trunc (temperature/2,0). Times were 2 seconds for each
 tablespace.
> Then I did a select temperature from ... where temperature in
 (22, 23.1,
>
> ..., 27.2) and got 2 seconds for each tablespace.
> The tables in the tablespaces are called raw_table and
 ntfs_table, and I
>
> double-checked my sql script for the test, it indeed queries
 two tables
> and
> not the same one twice.
> I even increased db_file_multblock_read_count from 8 to 16,
 and it did
> nothing for the full table scans: I still get 20 seconds in
 each
> tablespace.
> Am I missing something here?
> I know that a table with 188K rows is not large, but I would
 have
> expected
> to see at least a 10% difference between the raw parititon and
 the ntfs
> partition.
> Did I forget to set a registry setting somewhere? Is NT
 forcing all I/O
>
> calls to go through the kernel and the file system routines
 even if
> Oracle
> is writing to a raw partition? Is it because I am using IDE
 drives that
> I
> cannot see a benefit?
>
> If I see no difference on this test machine, I don't see why
 anyone
> would
> bother implementing this scheme on production servers.
 Hopefully I
> forgot
> to do something and the raw partitions do make a difference.
>
> Ironically the Oracle technical paper explaining how to set up
 raw
> partitions on NT starts by saying that significant performance
 benefits
> can be achieved by placing redo logs and datafiles on raw
 partitions...
>
> Have you placed your datafiles on NT raw partitions? Did you
 see any
> benefits? If so, what did I forget to do??
>
> TIA
>
> Pat.
>
>
Received on Sat May 20 2000 - 00:00:00 CDT