4 Replies Latest reply on May 11, 2014 7:31 PM by Renato Bonomini

    Load Balancers

    Jorge Xifra
      Share This:

      Has anyone done "Capacity Planning" on Load Balancers (F5 or another).

      In particular considering metrics like: CPU, Memory, hits/sec, number of users & network I/O.


      What has been the experience?

      How you loaded the metrics into BCO?

      Was BCO useful in doing the CP?




        • 1. Re: Load Balancers

          Yes, it can be done and it has been done.

          BTW, we are thinking about devoting the next best practice webinar to Network Capacity Management. Stay tuned.



          • 2. Re: Load Balancers
            Jorge Xifra

            Nice the Network Capacity webinar

            • 3. Re: Load Balancers

              We would also be very interested in a Network webinar


              • 4. Re: Load Balancers
                Renato Bonomini

                Hello Jorge,

                late is better than ever, sorry for the delay


                Make sure to follow the upcoming network webinar where Riccardo Casero and I will show with the BMC team how to manage network devices and links.


                Yes, Load Balancer are quite interesting and their analysis can show good use of correlation of parameters.

                According to my personal experience, the following metrics are useful

                • CPU Utilization: complex forwarding and balancing rules stress the CPU of these dedicated devices
                • Number of connections: this metric defines the actual throughput and can be used for discussions with Application Developers
                • Input or Output Bandwidth usage: as the number of connections, we can estimate the maximum bandwidth of information a device can manage. This number changes according to different types of worload


                While each packet or request or connection can be different, using the published 'maximum number of connections' or 'maximum bitrate' of a device can be misleading.

                The actual maximum sustainable rate of connections or bitrate can vary and the following technique helps understanding values under the current workload.


                The following analysis shows that CPU behaves like the workload, memory is flat: so we'll focus on CPU.

                PLOT_785_7_14 copy clean.png


                One possible correlation is with the output bitrate

                Network pre PLA PR_RUN_285_2 clean.png


                Or another type of correlation is with number of connections, for example one load balancer in the application was changed and it's CPU utilization dropped significantly:


                PLOT_971_3_6 clean.png



                But what about the actual workload?

                The extrapolation model allowed to estimate that

                • old capacity = 10k conn/s
                • new capacity > 30k conn/s

                PR_RUN_262_1 clean.png

                PR_RUN_263_1 post upgrade clean.png


                As usual, depending on the type of worload your bottleneck may shift.


                In any case, given the extremely high physical interface bandwidth of modern devices and ever increasing complexity of balancing/forwarding rules, one could expect the CPU to still be the limiting factor.