4

I'm a bit embarrassed that I've never seen or used the -iops parameter as part of the esxcli deviceconfig options.

I stumbled upon an article and proceeded to benchmark before and after using various block sizes and read/write values and am seeing ungodly gains in performance of our soon-to-be-deployed v3700 SAN, regardless of disk configuration.

Monitoring throughput of the 8 iSCSI interfaces (it's an active/active design) with Solarwinds free SNMP bandwidth monitor on the stacked switches, I saw each interface jump up from 120ish Mbps (or 12% utilization) to 325ish Mbps (~33% utilization). Some of them were even pegged, which led me to think there was packet loss on that particular port (didn't get a chance to check the port statistics before the IOMeter job completed).

So what's the downside here, besides oversaturating a particular path (link)? What's a safe and happy setting people are using? Seems too good to be true.

gravyface
  • 13,957
  • 19
  • 68
  • 100

1 Answers1

3

Most of the iSCSI MPIO configurations I run with VMware end up using the SAN vendor's recommended path-selection policies.

For round-robin, most of the storage systems I've used recommend lowering the default number of I/O operating before switching paths from the default 1000 to 1.

Edit: I had a v3700 at a recent job. The IBM best practices (see page #12) reflected that the paths should be fixed on ESXi 5.1 and round-robin on 5.5. For tuning, we tuned down to 1 I/O operation.

ewwhite
  • 197,159
  • 92
  • 443
  • 809
  • Yeah, interesting that I'm not seeing that for the IBM stuff, at least not in the Redbook I'm reading. I've also seen two instances where VMware support reps are saying two different things: "never change it" or "follow SAN vendor recommendations". Just to make things more interesting is that I've seen others who are using a byte value based on average ethernet frame sizes (8800 for 9k Jumbo Frames) so that new paths are used when the frames been maxed out. Either way, it's hard to ignore the games, whether that translates into real world or not. – gravyface Jan 16 '14 at 18:33
  • 1
    @gravyface See the edit above. – ewwhite Jan 16 '14 at 18:44
  • Ah nice. Navigating IBM's site is a nightmare. :/ – gravyface Jan 16 '14 at 18:50
  • actually it looks like they're not really recommending Fixed or Round Robin, just that the former is the default for <= 5.1 and the latter is the default for >= 5.5. – gravyface Jan 16 '14 at 18:52