Transcontinental iSCSI for databases
I talked about the advantages of the separated ZIL and the L2ARC even on rotating rust in my article “iSCSI for I/O intensive tasks” in March 2008. I didn´t found the time to write down the findings of my tests as i started LKSF not much later. But Jignesh Shah did some tests with postgres and wrote down his results. He even took the concept a step farther:
Unfortunately the storage that's available is in Colorado while my PostgreSQL server is located in Massachusetts.
He tested an database with local sZIL and L2ARC and a remote iSCSI pool and got excellent results:
Not bad. Cutting the latency of writes of something that would have taken in excess of 8-10 minutes is atleast recorded within 4 seconds on nonvolatile cache/log combination and allowing ZFS to sync it up to the actual storage.
The whole concept of an ZFS augmented iSCSI shows promise even for latency-dependend applications.