SCSI RAID 0 Performance - 64bit vs. 32 bit

[Braxus]

Recently been doing benches on my AMI/LSI Logic MegaRAID controller and looks like there is a HUGE performance gap. Basically using two 36GB 10K IBM ultrastar drives...

Tiger MPX (32bit/33MHz)

If you look at that bench (ATTO) there is a HUGE performance difference between writing (slow as hell) and reading (resonable). Since the card I have is only has provisions for 64bit/33MHz operation and the board only has 64/66 slots I can't take use of the 64 bit transfers the card offers. Even so, shouldn't write speeds at least be a lot more closer to the read marks?

Then I decided to go with the Tiger MP which had support for 64bit/33MHz slots.

Tiger MP (64bit/33MHz)

Big difference... However, on cached I/O performance on the MP seems to get around 150-200MB/s LESS compared to the Tiger MPX. I would assume that cached I/O performance would be better as there is more bandwidth available. Any thoughts?

Also, other than buying a 64bit/66MHz RAID controller, any other ways to bump up performance when using 32bit/33MHz? Would really like to get those write scores up.

 

[dkozloski]

Changing the cache line size to 64K in my AMI SCSI RAID controller made all the difference in the world in the writes.

 

[Braxus]

That's what I'm using....

- 64K stripe size
- Read Ahead
- Write Back
- Cached I/O

That configuration seems to work well on the MP board but not on the MPX board. Tried fiddling with all the settings as as well (save the 64K blocks) and nothing seems to improve scores on the MPX board.

 

[dkozloski]

I didn't say stripe size, I said cache line size. It is a rather obscure parameter that can be set in the controller at startup. I have an AMI caching controller with 128Megs of cache memory and a battery backup module for the memory.