topic FC10 Operation in Disk Enclosures

FC10 Operation

George Nikoloudis_1 — Wed, 21 Nov 2001 09:09:48 GMT

Dear All
I would like to ask about the FC10 operation.
Since the link is Fibre channel and the disks are SCSI daisy chained then there is no way where I can read in parallel from each disk.Is this true?

I believe that such thing can perfromed only when using LUNs where you can have up two disks work as one through different controller. Such system could be FC60 or XP series.

How the kernel parameters affect the Fibre Channel operation?

Re: FC10 Operation

Duncan Edmonstone — Wed, 21 Nov 2001 09:50:10 GMT

Not really... If this were the case no-one would ever have used striped disks with all the SCSI disk enclosures that have been around for so long... Perhaps I'm not understanding your question, are you weighing up whether to stripe across the disks in the FC10 or not?

As for kernel parms... I can't think of any I'd use specifically for fibre channel - but I'm happy to be proved wrong.

HTH

Duncan

Re: FC10 Operation

George Nikoloudis_1 — Wed, 21 Nov 2001 09:58:17 GMT

I am not using striping. But how the FC10 works in case that the striping is not performed and in case that the striping is perfromed.
Can I stripe along 10 disks?
How can I define the striping size for ORACLE?

Re: FC10 Operation

Duncan Edmonstone — Wed, 21 Nov 2001 10:42:24 GMT

You sure can, if you want to - striping on a FC10 is no different from on any other disk cabinet. As for how you lay it all out it depends on:
Your config - how many FC HBAs you have, how many FC10s you have and whether your FC10s have 2 LCCs
Your availability requirement... (Do you need to mirror your data?)
Your Database - what the expected profile of your database is (OLTP or DSS? many small random writes/large sequential writes etc.)- the attached info is from Oracle's 'Designing & Tuning for Performance' Manual.

______________________________________________
Striping Disks with Operating System Software
As an alternative to striping disks manually, use operating system utilities or third-party tools, such as logical volume managers, or use hardware-based striping.

With utilities or hardware-based striping mechanisms, the main factors to consider are stripe size, number of disks to stripe across (which defines the stripe width), and the level of concurrency (or level of I/O activity). These factors are affected by the Oracle block size and the database access methods.

Table 20-14 Minimum Stripe Size
Disk Access Minimum Stripe Size
Random reads and writes
The minimum stripe size is twice the Oracle block size.

Sequential reads
The minimum stripe size is twice the value of DB_FILE_MULTIBLOCK_READ_COUNT.

Table 20-15 Typical Stripe Size
Concurrency I/O Size Typical Stripe Size
Low
Small
k * DB_BLOCK_SIZE

Low
Large
k * DB_BLOCK_SIZE

High
Small
k * DB_BLOCK_SIZE

High
Large
k * DB_BLOCK_SIZE * DB_FILE_MULTI_BLOCK_READ_COUNT

Where k = 2,3,4...

In striping, uniform access to the data is assumed. If the stripe size is too large, then a hot spot may appear on one disk or on a small number of disks. Avoid this by reducing the stripe size, thus spreading the data over more disks.

Consider an example in which 100 rows of fixed size are evenly distributed over 5 disks, with each disk containing 20 sequential rows. If your application only requires access to rows 35 through 55, then only 2 disks must perform the I/O. At a high rate of concurrency, the system may not be able to achieve the desired level of performance.

Correct this problem by spreading rows 35 through 55 across more disks. In the current example, if there were two rows per block, then we could place rows 35 and 36 on the same disk, and rows 37 and 38 on a different disk. Taking this approach, we could spread the data over all the disks and I/O throughput would improve.

HTH

Duncan

Re: FC10 Operation

Duncan Edmonstone — Wed, 21 Nov 2001 10:46:37 GMT

Oops - the table didn't paste quite right - hopefully it will look OK now...

Concurrency I/O Size Typical Stripe Size
------------------------------------------
Low Small k * DB_BLOCK_SIZE

Low Large k * DB_BLOCK_SIZE

High Small k * DB_BLOCK_SIZE

High Large k * DB_BLOCK_SIZE * DB_FILE_MULTI_BLOCK_READ_COUNT

Re: FC10 Operation

Duncan Edmonstone — Wed, 21 Nov 2001 10:49:01 GMT