On Wed, Mar 2 at 9:58, David Dyer-Bennet wrote:
On Tue, March 1, 2011 10:35, Garrett D'Amore wrote:a) do you need an SLOG at all? Some workloads (asynchronous ones) will never benefit from an SLOG.I've been fighting the urge to maybe do something about ZIL (which is what we're talking about here, right?). My load is CIFS, not NFS (so not synchronous, right?), but there are a couple of areas that are significant to me where I do decent-size (100MB to 1GB) sequential writes (to newly-created files). On the other hand, when those writes seem to me to be going slowly, the disk access lights aren't mostly on, suggesting that the disk may not be what's holding me up. I can test that by saving to local disk and comparing times, also maybe running zpool iostat. This is a home system, lightly used; the performance issue is me sitting waiting while big Photoshop files save. So of some interest to me personally, and not at ALL like what performance issues on NAS usually look like. It's on a UPS, so I'm not terribly worried about losses on power failure; and I'd just lose my work since the last save, generally, at worst. I might not believe the disk access lights on the box (Chenbro chassis, with two 4-drive hot-swap bays for the data disks; driven off the motherboard SATA plus a Supermicro 8-port SAS controller with SAS-to-SATA cables). In doing a drive upgrade just recently, I got rather confusing results with the lights, perhaps the controller or the drive model made a difference in when the activity lights came on. The VDEVs in the pool are mirror pairs. It's been expanded twice by adding VDEVs and once by replacing devices in one VDEV. So the load is probably fairly unevenly spread across them just now. My desktop connects to this server over gigabit ethernet (through one switch; the boxes sit next to each other on a shelf over my desk). I'll do more research before spending money. But as a question of general theory, should a decent separate intent log device help for a single-user sequential write sequence in the 100MB to 1GB size range?
ZIL, as I understand it, is only for small synchronous writes, the opposite of your workload. If you don't have a SLOG, the ZIL is embedded in your pool anyway. Above a certain size, the writes go straight to the pool's final storage location. I'd be curious if you're getting errors in your SMB stream, or maybe your server is set to hold onto too much data before flushing (default is 45 seconds, and there's been reports of the system not always force-flushing early when the buffers fill) I've heard reports of short-stroking the amount of time it accumulates write data resulting in improved performance in some workloads. --eric -- Eric D. Mudama edmud...@bounceswoosh.org _______________________________________________ zfs-discuss mailing list zfs-discuss@opensolaris.org http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
