Hello all, We run two Lustre on ZFS filesystems running Lustre 2.4.0 and ZFS 0.62 (the defaults from the ZFS on Linux repo). One filesystem has roughly 150GB of metadata and the other has 6GB. I've been using ZFS send/receive functionality to make full/incremental backups of the metadata on a daily/hourly basis. However, this has become more and more difficult to do as the size of our metadata has increased. On our filesystem with 150GB of metadata, we are using ZFS RAID 10 with 4x15K 300GB drives. Our backup metadata server has the same configuration. When I'm using ZFS send/receive to do a full backup, I can only reach somewhere between 3-5MB/s in throughput over 1GB Ethernet. I've used a tool called mbuffer to determine that the send side definitely is causing the slow performance. For our larger metadata pool, a full backup can take an entire day and decreases performance of the file system for other uses. CPU wait is quite high on the system during normal operations and increases substantially during backups. Our other, smaller, ZFS metadata partition is a poor test subject as the loads are far lighter and the partition uses ZFS RAID10 on 4xSSDs. There I see performance at 10-20MB/s and much friendlier CPU wait times. Even there, that's a lot lower performance than those drives are capable of doing, but the smaller metadata size makes it very reasonable for now. What are other people doing for backing up ZFS metadata partitions? What kind of performance do have you normally see for these backups? If anyone has found any gotchas regarding performance of these types of ZFS operations, please let me know! Thanks, Andrew -- Andrew Wagner Research Systems Administrator Technical Computing UW-Space Science and Engineering AOSS Room 439 608-261-1360