The Age Old Unclearly Answered Question: Write-Through or Write-Back

We're going to a new colo facility where they have all their clients using Sun StorEdge T3s for database backend. These things have a 1Gbps FC (100MB/s) interface, with 1GB of battery-backed write-cache. It also has hardware RAID.

Our database I/O is mostly read with a bit of small random writes (and is about 50GB in size total).

The cost of this darn T3 is crazy, and I told my CIO that is was overkill for what we need. I suggested instead that we forego the hardware RAID and use Veritas VM with RAID 10 on an UltraSCSI3 (160MB/s) StorEdge 3310, which has dual internal SCSI buses with 6 disks per bus in a unit. I plan to use Veritas to mirror across the busses and stripe each bus of drives. On top of that, we're getting Veritas vxfs filesystem with Oracle Edition Quick I/O to make the files appear as raw disk. The OS will be Solaris 8. In the end, we end up saving thousands of dollars getting THREE StorEdge 3310s for various purposes instead of purchasing ONE StorEdge T3.

At first, the colo tried to fight back by saying, "HW RAID is *always* faster than SW RAID." I killed that argument quick with a variety of benchmarks and real-life testimonials from several colleagues. Incidentally, I've heard horror stories about the T3 from several people, too.

Now, they're trying to say that we need the write cache for performance. I recall reading that, with my type of workload, Oracle would be better off with more investment in host RAM and optimal disk striping than disk write cache. I can't remember where I read this online, though.

Does anyone have any URLs, preferably from Oracle, that explains this?  It would be very helpful to have for a meeting scheduled tomorrow morning with the colo facility. I already have docs about the HW RAID vs. SW RAID myth.

Thanks in advance. Received on Wed Sep 17 2003 - 17:59:02 CDT