On Jun 16, 2010, at 6:46 AM, Dennis Clarke wrote: > > I have been lurking in this thread for a while for various reasons and > only now does a thought cross my mind worth posting : Are you saying that > a reasonably fast computer with 8GB of memory is entirely non-responsive > due to a ZFS related function? >
The problem is that ZFS ends up trying to do too much work in syncing context. Because of the way the ZFS transaction model works, it's important that the txg sync phase remain constant (ZFS currently shoots for 3 seconds, though this was recently changed to 1). If this phase takes too long, then all other ZFS work is blocked. If this happens to your root pool, this would have the appearances of a hard hang. These problems have generally been due to one of two root causes: 1. Destroying snapshots, where the deadlist must be processed entirely within one txg. 2. Freeing blocks in a deduped dataset, which requires updating the DDT in syncing context. Most of the pathological aspects of these problems have been fixed (or will soon be fixed) in the latest source: 6922161 zio_ddt_free is single threaded with performance impact 6938089 dedup-induced latency causes FC initiator logouts/FC port resets 6948890 snapshot deletion can induce pathologically long spa_sync() times 6948911 snapshot deletion can induce unsatisfiable allocations in txg sync 6949730 spurious arc_free() can significantly exacerbate 6948890 6957289 ARC metadata limit can have serious adverse effect on dedup performance 6958873 lack of accounting for DDT in dedup'd frees can oversubscribe txg 6960374 need auxiliary mechanism for adjustment of write throttle There are still some extreme cases that can result in long sync times when using dedup, but nothing pathological (i.e. 30 seconds, not 30 hours). Expect to see fixes for these remaining issues in the near future. - Eric -- Eric Schrock, Fishworks http://blogs.sun.com/eschrock _______________________________________________ zfs-discuss mailing list zfs-discuss@opensolaris.org http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
