Slow sequential write performance on Perc 3/DC

I have recently installed a new Sybase server and have found that the transaction processing throughput is very slow.  This has been traced to slow sequential write performance on the log array (controlled by Perc 3/DC).

My system configuration:

- Windows 2000 server running on a dual xeon 3Ghz, 2GB ram
- Sybase log array - 2 x 18GB 15K rpm Seagate Cheetah drives configured Raid-0 on Dell Perc 3/DC
- Sybase data array - 8 x 36GB 10K Atlas IV drives configured Raid-0 on Dell Perc 3/DC

I am using Raid-0 because we are looking for maximum performance and the server isn't intended for production.  Stripe size on raid arrays is 64K.

Using IOMeter I have benchmarked sequential writes on 2KB blocks at 8MB/s.  Enabling write-back cache drops sequential writes to around 6MB/s.  This is much slower than an equivalent test on a single 7200RPM IDE disk (Around 25MB/s).  Using a larger block size of 64K in IOMeter results in 90 MB/s, but this drops to 8MB/s when the write-back cache is enabled.  Sybase uses 2K blocks, so this is the figure that I need to increase.

I have tried various configuration combinations with little effect.  This includes:
- Various stripe sizes (2KB - 128KB)
- All permutations of the array cache settings (write-back/write-through, read-ahead/no read-ahead, direct IO/Cache IO)
- Swapping the controller card to different PCI slots including 32Bit/33Mhz, 64Bit/66Mhz, 64Bit,133Mhz
- Utilising Windows 2000 server striping over individual logical drives for each disk
- Different disks - the data array (Atlas IV) and the log array (Seagate Cheetah) have the same write performance.
- Various firmware and drivers including conversion to LSI MegaRaid 1600 firmware and drivers.
- I have confirmed the SCSI interfaces and drives are all operating at Ultra160.

My expectation is that the write-back cache should improve write performance - particulrary on small blocks.  This hasn't proven to be the case.  I would greatly appreciate any pointers on improving the sequential write performance on small blocks.  Also, if this is the expected performance of the Dell Perc 3/DC card then I would be grateful for that feedback so that I can start looking for a new controller card!

Thanks

Message Edited by M0cha on 03-30-2004 03:42 AM