1 Reply Latest reply on Nov 24, 2014 1:28 AM by artist-06

    EMC VNX Round Robin and "I/O Operation Limit"

    vmproteau Expert

      We don't use PowerPath with are ESXi Hosts. Currently we are using "VMW_SATP_ALUA_CX” array type and a “Round Robin" PSP Policy.

       

      The default "I/O Operation Limit" is 1000. I have read tuning it to 1 can provide better throughput and response time depending on workload. It appears sequential workloads have the most benefit. Although random and OLTP type 2 workloads are also optimized with a setting of 1, it's a relatively small improvement from 1000. The papers I've read were specific to the VMAX so I'm not sure if the same would hold true for the VNX. It was also noted that tuning it to 1 will increase ESXi CPU and VMAX director utilization.

       

      1. Would I expect this data to be consistent with the VNX arrays?
      2. Using this data, I'd only moidify this setting if I could demonstrate my storage traffic was predominantly sequential. What is the best way to determine the storage characteristics of my environment?
      3. I suppose if I could identify enough predominantly sequential VMs to support a seperate sequential cluster that might make sense.
        • 1. Re: EMC VNX Round Robin and "I/O Operation Limit"
          artist-06 Novice

          Hi there,

           

          1. Would I expect this data to be consistent with the VNX arrays?

                    - VMAX and VNX arrays are like Orange vs Apple. Setting Round Robin IOPS to 1 on VNX is not the best practice. When I worked at EMC as a vSpecialist we recommended changing this setting to 50 or 100 depending on your workload. When I worked with a customer we tuned this setting on a mixed workload environment and 50 was the sweet spot.

          2.Using this data, I'd only moidify this setting if I could demonstrate my storage traffic was predominantly sequential. What is the best way to determine the storage characteristics of my environment?

                    - I would not agree to this, you can benefit from this setting if your IOs are Random too. Imagine your DB is sending 500 Random IOs if your setting is default at 1000 you would only use one path. If the setting was 50 then requests would be divided in to 10 paths.

          3. I suppose if I could identify enough predominantly sequential VMs to support a seperate sequential cluster that might make sense.

               - Not necessarily refer to my answer in question 2 above. There are tools to identify your IO characteristics. A good start is ESXTOP or Remote ESXTOP. You can ask your EMC presales and partner presales to run an assessment on your VNX and VMware hosts. They can send you instructions to enable performance logging on VNX and enabling ESXTOP on VMware hosts. You will upload the logs and receive a very nice powerpoint presentation of your environment.


          Regards,


          Burak Uysal