Optimum (theoretical) blocksize

[Casper]

Hi,

I just read up on some documentation regarding performance and blocksize. I always used to use 4K blocks on Linux. Now I saw in a presentation from Exchange that if you use typeII storage area's then it is better to use 8K blocks on Linux. Can anyone comfirm this for me?

Furthermore am I interested in the optimum blockszie for bi and ai. Is this 4K or 8K on Linux? We tend to use 8k blocks for bi and ai. But reading up it seems to be suggested that 4K blocks are better to use. Is this correct or is 8K good? I am not in a real position to do extensive testing with this, therefore my question

TIA,

Casper.

 

[TomBascom]

It depends.

In most tests that I've done 8k is somewhat faster and packs data tighter than 4k. But the difference isn't usually huge and unless you need every imaginable bit of performance both 4k and 8k are good choices. 1k, OTOH is just awful.

I'd have to see whatever you were looking at that said 4k might be better for ai & bi. That seems unlikely. I usually use 16.

Someday we may see larger db block sizes. It gets mentioned frequently but I haven't heard of a delivery date.

 

[Casper]

Sorry for the late response. But thanks for the info.
The bi blocksize suggestion I found in a KB article:
http://progress.atgnow.com/esprogress/Group.jsp?bgroup=progress&id=P19806

5. Set bi log block size to 8 kb.
Allows for more efficient writing of the bi log, which is always done using synchronous writes.

On Linux, use 4K.

On Windows, use 4K.

Casper.

 

[TomBascom]

I suspect that someone is being overly cautious with regards to the theoretical "torn page" issue.

It might be interesting to benchmark it but I wouldn't expect it to be a big deal.