Actually writes faster then reads are typical fora Copy on Write FC (or Write Anywhere). I usually describe it like this.
CoW in ZFS works like when you come home after a long day and you ust want to go to bed. You take of one pice of clothing after another and drop it on the floor just where you are - this is very fast (and it actually is copy on write with block allocation policy of "closest"). Then the next day when you have to get to work (in this example assuming that you wear the same underwear again - remember not supported ! :) - you have to pick up all the cloths one after another and you have to move across all the room to get dressed. This takes time and it is the same for reads. So in CoW it is usual that writes are fast then reads (especially for RaidZ/RaidZ2, where each vdev can be viewes as one disk). For 100% synchronous writes (wcd=true), you should see the same write and read performance. So for your setup I assume: 4 x 2 disk mirror with Nearline SATA: Write (sync, wcd=true) = 4 x 80 IOPS = 320 IOPS x 8 KB Recordsize = 2,6 MB/Sec if you see more thats ZFS optimizations already. If you see less - make sure you have proper partition alignment (otherwise 1 write can become 2). Read = 8 x 100 IOPS (some more IOPS because of head optimization and elevator) = 800 IOPS x 8k = 6,4 MB /sec from disk. Same problem with partiton alignment. For 128k block size ? Write: 320 x 128k = 42 MB/sec Read: 102 MB/sec ZFS needs caching (L2ARC,ZIL etc.), otherwise it is slow - just as any other disk system for random I/O. For sequencial I/O ZFS is not optimimal because of CoW. Also with iSCSI you have more fragmentation becase of the small block updates. So how to tune ? 1) Use ZIL (this will make your writes more sequencial, so also optimize the reads) 2) Use L2ARC 3) Make sure partition aligment is ok 4) try to disable read-ahead on the client (otherwise you case eben more random I/O) 5) use larger block size (128k) to ave some kind of implicit read-ahead (except for DB workloads) Regards, Robert Heinzmann -- This message posted from opensolaris.org _______________________________________________ zfs-discuss mailing list zfs-discuss@opensolaris.org http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
