db_block_size and OS block size Correlation

[cdutta]

Hi everybody,

I have a database used for datawarehousing. I set my
db_block_size to 16K. I have heard conversations that
a OS block size of 512 k will greatly improve reading performance for DSS on Widows NT. It is that so? What is the correlation between OS block size and db_block_size? Did anyone tried to optimize the I/O by changing OS block size from a default 64K to 512k? The database is Oracle 9i running on Windows 2000.

Thanks,
CD

[tamilselvan]

Do you use Volume Manager for your disks?

Volume Manager overrides the OS block size. For example, Veritas volume manager's default I/O block size is 8K.

[cdutta]

I am running Windows 2000 not Unix. I am not familiar
with Unix volume manager.

[jrpm]

It has been my experience that if you use the 16K block size, you can/should adjust the data read amount to match the OS read buffer size (amount of data the io subsystem returns in 1 read.) If this is 512K, then a 32block multi-block-read-count would be appropriate.

I am not sure what you mean by os block. Is it the read buffer or the cluster size (1000-512 byte sectors.)?
A 512K disk block feels a little large. I don't know that I have seen a read amount over 32K at the io level.

[cdutta]

We ran 2 tests so far and the results are somewhat puzzling.
#1) Low level format of our disk data array to 1024k blocks.
Our Raid controller was set to 1024k stripe, (Raid0)
Oracle DB_BLOCK_SIZE=16384 , DB_BLOCK_BUFFERS=500M

#2) Low level format of our disk data array to 64k blocks.
Same Oracle config.

When building a identical tablespace on Item#2(driveset) it ran 15x faster than #1. I thought the larger the formatted blocksize on the disk itself, the quicker the throughput... and better oracle performance.. not true.. but why??