5 Replies Latest reply on Feb 23, 2015 3:05 PM by designerfx

    Issue with Volume Capacity Forecast for Physical Memory

    bluefunelemental

      Confused on what I am seeing on the the new Volume capacity forecast for physical memory for at least my net-snmp-Linux nodes:

      1. First record on my Top 10 Capacity Problems tells me that 'D----MGR2' physical memory volume utilization is critical and warning Now and at capacity in 6days
      2. Clicking into that net-snmp - Linux node details page the memory is  6% of 32GB used with a downward trend slope
      3. The Node Capacity Forecast resource says >1year for percent memory usage

       

      Thoughts on the 6 day to capacity vs > 1 year for memory?

       

      0.png4.png

      3.png2.png

      1.png

        • Re: Issue with Volume Capacity Forecast for Physical Memory
          aLTeReGo

          There are differences between how memory is calculated for the "Physical Memory" volume vs the "Percent Memory Used" that rob.hock may be able to explain in better detail, but the fact that this data is polled from two different sources is likely the reason why the values are different.

            • Re: Issue with Volume Capacity Forecast for Physical Memory
              bluefunelemental

              Looking forward to why that might be the case as it's confusing for my users.

              I don't want to keep it so I'm adding the resource filter for instancecaption <> 'Physical Memory'  for now.

               

              Another issue for rob.hock

              We had a server fail today due to being 100% fixed disk volume usage. Yet Capacity forecast said we had weeks left

              It would seem to me that your linear slope is being fooled by the large deviation into forecasting weeks till capacity when it's currently 99%

              Can you address why Capacity forecast would completely miss this? I cannot use it in this condition


              Screen Shot 2015-02-13 at 12.05.55 PM.png

              Screen Shot 2015-02-13 at 11.59.24 AM.png

                • Re: Issue with Volume Capacity Forecast for Physical Memory
                  rob.hock

                  Indeed sir- the sharp saw-tooth pattern is would be a classic case where linear regression would not be the most accurate. Maintaining utilization thresholds is still a good backstop for these conditions.

                  The below diagram illustrates how run-out could be calculated under similar conditions:

                   

                  linear_regression.png

                    • Re: Issue with Volume Capacity Forecast for Physical Memory
                      bluefunelemental

                      Rob,

                      Just spoke with a person who taught Statistics at Harvard (god that sounds pedigree) and the response was that your first chart with a constant slope coefficient would suffice being applied to our issue if you accounted for a smaller time-scale. While your forecast resource seems to indicate a 7day window - I would think a more granular time-scale would allow a simple linear regression to more accurately predict the at threshold/capacity date. I do not think changing the setting to 'calculate exhaustion using peak daily values' would make a difference in this "saw-tooth" case- if so let me know to try it.

                       

                      My recommendation is calculating detail, daily, and monthly coefficients for each object and allow us to use the zoom interface like chart resources have to switch between time-scales.

                       

                      The data notes on our graph below would be that around Nov 24th an application setting changed and no longer purged backup files. That resulted in a constant slope coefficient until the volume was purged then again 3.5 times. If you show 99% full in the disk space used resource and weeks or months in the forecast then our engineers will begin to suspect forecast is not "reliable" and ignore it. It's amazing how quickly things get ignored. Being able to see slopes based on this hour, today, this week, this month might help us to focus our efforts.

                       

                      Screen Shot 2015-02-13 at 11.59.24 AM.png

                      • Re: Issue with Volume Capacity Forecast for Physical Memory
                        designerfx

                        rob.hock what about ways to sort the capacity forecast by category? Such as: filtering out all memory and resorting specifically to actual volumes?  Or is there a way to put that into a report I have missed?