3 ms·
Very good points, but I think his conclusion is wrong. I think the fix will come from the filesystem. A filesystem that automatically mirrors (or distributes e
by ars 17y ago
Very good points, but I think his conclusion is wrong.
I think the fix will come from the filesystem. A filesystem that automatically mirrors (or distributes ecc code for) files, and checks checksums is the future.
- docmach 17y agoThat is exactly what ZFS does.
- wmf 17y agoMoving the same functionality from one layer to another won't help much. For example, rebuilding a full disk in ZFS takes just as long as in traditional RAID because the same work is being done. ZFS does help if your filesystem is not full since it doesn't bother rebuilding empty space, but this doesn't provide the order of magnitude improvement that we are looking for.
- rbanffy 17y agoExcept that a filesystem can tune the redundancy of the data to the amount of free space in the storage pool. It could be considered full when a minimum of redundancy was reached (like every block being stored at least on two different disks). As soon as a disk fails and is taken out of the pool, the file system could (conceivably, at least) start replicating all data on the remaining disks maintaining the redundancy minimum and shrinking the window of vulnerability to multi-drive failures. By the time a second disk fails, all data on it could be replicated elsewhere and all you would observe would be a shrinking file system (and urgent messages from the server).
- jasonwatkinspdx 17y agoThat's not entirely true. If we're at the block level, it's more difficult for us to optimize recovery. If we're at a higher level where we have access to some sort of summary information, like the checksummed tree of blocks in zfs, then we can quickly isolate mismatched data and copy only that from the available to the recovering drive. If we want the same capability at a layer below the file system, say the raid controller, than we end up needing a much more complicated implementation that is in effect a layering of two filesystems. Given general trends towards clustered commodity hardware I'd expect the software to keep getting smarter and the block level devices to keep getting dumber. It seems like there's a role that's missing in the storage hierarchy. Something like an extent manager that doesn't maintain all the actual metadata, leaving that to the file system above, but that does have a concept of a tree of indirectly referencing extents. Something similar to how allocation and garbage collection can be provided by libraries in c++.