Rethinking performance testing

12. Load model and login storm


Some time ago I published about the load model in Article 2 - Workload in the series Applying modelling on, which also describes the transaction blend with an example in more detail. More recently I posted Rethinking performance testing – The load model. When many users are actively at work with an application, the transactions they execute form a certain blend. If this blend contains relatively more “heavy” transactions you need more capacity then when “light” transactions prevail. The blend has quite an impact in a load model.

You need a load model in any performance test no matter if it is with load and stress testing or model based. Creating a good load model is not trivial. A load model can take some effort to create, requiring investigation at various stakeholders and some engineering, which could be the task of Ops engineers or performance testers. By the way, keep in mind that load model and performance model are two entirely different things.

When preparing a load model one should not only cover the steady state, but also take the login or boot storm into account. With load testing this can be addressed in the ramp up. Question is if the ramp up accurately represents a load model. It takes some study on login logs to gain insight in what happens during the login storm. I am always keen on getting my hands on this type of data but I don’t succeed too many times. So if anyone has login logs to share I’ll be quite happy to analyse them and return the results. With performance modelling you may use different load models for Login storm and steady state. I’ll come back on this soon.

Especially with corporate applications most of the users arrive at the office and start logging in in a short period of time. This may cause a peak volume on the login transaction type changing the transaction blend and make it different from the stationary state. How this effects performance depends on the weight of the login compared to the other transaction types.

Many businesses use Virtual Desktop Infrastructure (VDI) for their corporate applications. Starting their virtual desktop instances can cause a boot storm that can become an extra area of concern. The login storm may charge the servers and LANs, in addition the Boot storm can cause extra load on the hard disks or SSDs.

Optimising the capacities for the stationary situation may not be sufficient. Extra capacities may be required. How much extra capacities are needed depends on a number of factors:

  • The number of users that arrive at the office in the peak period
  • The period of arrivals
  • The weight of the landing page
  • The weight of the Login transaction
  • The size of the Windows boot profiles
  • . . .

In order to sustain a constant service level one may need more than three times the CPU capacity for some of the servers.