> > Which is why ZFS isn't a replacement for proper array controllers > (defining proper as those with sufficient battery to leave you with a > seemingly intact filesystem), but a very nice augmentation for them. ;) > > Nothing prevents a clever chap from building a ZFS-based array > controller > which includes nonvolatile write cache. However, the economics suggest > that the hybrid storage pool model can provide a highly dependable > service > at a lower price-point than the traditional array designs.
I don't have finished results that are suitable for sharing yet, but I'm doing a bunch of benchmarks right now that suggest: -1- WriteBack enabled is much faster for writing than WriteThrough. (duh.) -2- Ditching the WriteBack, and using a ZIL instead, is even faster than that. Oddly, the best performance seems to be using ZIL, with all the disks WriteThrough. You actually get slightly lower performance if you enable the ZIL together with WriteBack. My theory to explain the results I'm seeing is: Since the ZIL performs best for zillions of tiny write operations and the spindle disks perform best for large sequential writes, I suspect the ZIL accumulates tiny writes until they add up to a large sequential write, and then they're flushed to spindle disks. In this configuration, the HBA writeback cannot add any benefit, because the datastreams are already optimized for the device they're writing to. Yet, by enabling the WriteBack, you introduce a small delay before writes begin to hit the spindle. By switching to WriteThrough, you actually get better performance. As counter-intuitive as that may seem. :-) So, if you've got access to a pair of decent ZIL devices, you're actually faster and more reliable to run all your raid and caching and buffering via ZFS instead of using a fancy HBA. _______________________________________________ zfs-discuss mailing list zfs-discuss@opensolaris.org http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
