SLOG bottleneck on sync writes with smaller block sizes

[mav@]

As I've told, I am just going to investigate this NFS behavior. From what I saw, with iSCSI situation is better, but I haven't tested exactly multiple SLOGs recently.

NVMe supposed to have lower latency then SAS, which, when true, is good for SLOG.

 

[xyzzy]

For what it's worth, I checked out the iSCSI situation with multiple SLOGs and can confirm it's much better than NFS.

Using 3 Intel SLOGs and 3 Samsung data drives, testing with CrystalDiskMark 5.2.1 with 5 passes at 4 GiB, here are the results:

Code:

					NFS	 iSCSI
				  ------   ------
Seq Q32T1 Read	1234.0   1052.0
Seq Q32T1 Write	277.2	817.7
4K Q32T1 Read	  334.3	393.5
4K Q32T1 Write	  45.8	124.1
Seq Read		   968.3	841.7
Seq Write		  261.0	460.7
4K Read			 42.3	 32.9
4K Write			16.3	 16.4

Writes are a LOT better. Reads are a bit slower than NFS....not sure why.