SLOG tuning for sync writes?

[dswartz]

So I have a 4x2 raid10. NL SAS drives. Serving a datastore to vsphere via NFS over a 10gbe link. If I disable sync on the datastore, crystaldiskmark shows a write speed to the pool of about 2/3 the read speed. The writes were maxing out at about 400MB/sec (for 4 NL drives, that seems close to max). I have a ZEUS STEC SAS SSD which I erased and threw in as SLOG. The sequential writes went down to about 50MB/sec. To eliminate the SSD as an issue, I created an 8GB ramdisk using the brd driver, and replaced the STEC as the SLOG. Re-ran the test. Write speed maxes out at about 200MB/sec (about half that of async mode.) Since the SLOG is a ramdisk, it's hard to imagine there is a physical limitation, so I am suspecting some kind of ZIL tuning issue? I want to run this using a pacemaker cluster to give me HA storage, so obviously sync disabled or a ramdisk SLOG ain't gonna cut it :) Any hints/tips welcome...

[DeHackEd]

This is an issue tracker, not a support forum. Please use the mailing list.

[dswartz]

We have had more than one issue here vis-a-vis sync writes (NFS or otherwise.) My suspicion is that this may be in the same category. I have repeated this experiment on a local dataset with sync=always, and it seems like ZIL is feeding the disks on the pool small enough chunks of data that IOPS of the spindles is the limiting factor (which would explain why a ramdisk SLOG didn't help.) I wrote 16GB of data to the dataset in question, for an aggregate write throughput of 142MB/sec. On 4 mirrored vdevs. That's about 35MB/sec per mirrored spindle. If the IOPS is limiting the writes to the pool due to some ZIL tuning default (or whatever), wouldn't it be helpful to have a better default? Or at least have this documented somewhere?

[dswartz]

Digging through the issues, this looks like it might be #1012.