Once a model is built, testing of the user concurrency and data load levels occurs, and then optimizing the system for the specific use case and conditions is carried out. Then, we have three main options in order to tune for optimum performance. These are the main optimization options:
1. Model design
Is the model designed correctly?
Have you reduced sparsity and unnecessary complexity?
Is the model too big?
Have you neatly designed the model to have input, engine, and output modules?
Have you cleaned up as you go?
Problems often exist when you have added to the model, tested something that did not work out, and then not removed what you tested that didn’t work. This piece is not fulfilling any requirements. Sometimes we refer to this as model debt. Remember, Anaplan is a living, breathing model and so any line items that exist in the model, whether used or not, are used by the engine. A surplus piece (model debt) is an inefficient use of model space.
2. Model calculations
Check that calculations are as efficient as possible. Are you using standard functions to be more efficient?
3. Platform code
Do we need to engage L3 and/or engineering to look at code optimization?
Performance issues including data volume and user concurrency
Performance and the experience the end user has are of critical importance when deploying applications to a wide audience. Therefore, several factors need to be considered when deploying, in order to optimize performance and determine whether a single instance or distributed instance strategy is best:
In order for end users to enjoy the best possible experience and have an average response less than two seconds to most popular requests, model size and concurrency must be managed appropriately. In many cases a base model is produced that contains all the dimensionality and calculation logic and then the model is subjected to a series of tests that determine what the end user experience and model performance will be.
The first test is a load test where data is loaded into the model to simulate what the production model volumes would actually be. During this test, basic functions are performed such as data input, allocations, filtering, pivoting, sorting, list formatted item drop down manipulation, etc. This is done both in an automated fashion and via human intervention. If you determine that some or many functions are slow and server memory and CPU are used to the maximum, this is likely a case for distribution. If however, the model is slow, but user concurrency is minimal, then this could form a case for a single model instance as the system is merely processing numbers and not being accessed by a user community. Otherwise, this model could also be split to provide a better user experience.
The size of a model measured in number of cells or in memory size is a good indicator for splitting a model. We are setting the expectation that a model size should not go beyond 15B cells or 120 GB of memory. Therefore, if an application requires 30B cells, it should be split in two models. Here’s an example of how a split model decision can be made:
First, estimate the size of the application: List the main dimensions that will be used in each application and define the expected number of cells for each the valid combinations of dimensions (these will be modules).
# of cells for the group
Customer: 80 (incl. hierarchy rollups)
Product: 1500(incl. rollups)
Time: 36 months, 12 Q, 3 Year, 3 YTD
Version: 2 (Actual, Budget)
Line items: 50 metrics
Time (should be same for each group)
Versions (should be same for each group)
Total Application 1
Then summarize how many models will be needed for each application.
Estimated size in cells
Estimated size in GB
# required models
The second test is user concurrency. If you have an application that requires a large user base to interact with it, a user concurrency test should be performed. As a general rule, user concurrency is approximately 10% of the total user community. Therefore, if you have a total user base of 1000, around 100 people will be on the live system performing tasks at any given time. It is usually unlikely that many more than that number would be accessing the system at the same time.
In some cases though, applications follow a set high concurrency pattern and this needs to be taken into account. For example, a weekly sales forecast may have 1000 users on the system, but very likely each Sunday (if forecasts are due Monday) the user concurrency will be quite high, maybe as high as 50–60%. Your processes and experience will determine exact concurrency in high traffic applications or periods. The best approach to get to the right number of users in a model is to test the concurrency with automated tests, and then with manual tests that include a large number of real users.
First, start with User Acceptance Testing (UAT). In short, UAT involves human users simultaneously performing scripted tests inside the platform. During these tests, system behavior will be monitored and reported by each of the human testers, which may be provided via a user survey that is launched post UAT.
Then, automated testing can be performed in the platform. Automated testing simulates user actions across the platform. To do this, coordinate with the Anaplan QA Team to schedule automated testing of load, performance, and concurrency.
It is also important to monitor the server while the automated testing is in place to monitor memory and CPU usage. The Anaplan QA team can obtain server monitoring metrics as part of the model performance testing process. In either case, application tuning needs to happen to optimize for all conditions needed.
Multi-model application optimization
The application tuning lifecycle includes a 2-step, iterative tuning process that reoccurs during the model building process. Step 1 is carrying out the complete build. Step 2 is tuning at the application level (i.e. optimizing the design and the calculations or business rules) by Anaplan’s L3 Support team and the solution architect. You may also make additional platform level or code optimizations with the assistance of Anaplan’s engineering department on rare occasions.
In many situations, enterprises need to split very large and complex models for various reasons including:
Performance issues, including data volume and user concurrency
Metadata time cycle differences
Regional / business process differences
Anaplan is a platform designed to enable businesses to build models in almost endless configurations, so there is no pre-set size recommendation for where a model can be distributed. It is not uncommon for a 15-billion-cell model performing complex calculations to remain a single model, used by only a single person or just a few people. However, in contrast to that, it is also not uncommon to have a distributed model as small as 1 billion cells, with complex calculations and multiple people in multiple locations using the model.
As a general guide, this table takes into consideration the factors that influence a single model or distributed model solution.
Large Data Volumes, (> 10GB)*
High User Concurrency*
Sample Model 1
Sample Model 2
Sample Model 3
Depends on actual volume
Sample Model 4
Sample Model 5
Depends on actual volume
Sample Model 6
Depends on user concurrency
Sample Model 7
Depends on actual volume
* As always, apply appropriate testing and tuning to optimize the model. Different combinations can have a dramatic effect on desired performance and experience.
Anaplan has robust security across its platform. In some cases, it’s possible to achieve region-specific experiences using selective access. If this is the case, then distributed models are not necessary. But in mixed environments where model builders and end users operate in the same model, and where various business processes exist, at times it makes sense to separate or distribute models rather than have them in a single instance. For example, you may have different countries that all need access to a workforce planning application. You also have model builders from each country modeling and maintaining their section. By distributing the models and restricting access, this problem is abated.
Note: Where there is a need to segregate administration (model builder) roles, the split models will need to be in different workspaces, as the admin role is by workspace, not by model.
Metadata time cycle differences
A single instance of a model serving the world across multiple time zones does not respect the different business cycles involved, and therefore updates to data and/or metadata of a model will affect the entire community, some of whom may be in the middle of their planning cycle. These changes may be small, but in many instances are large-scale and frequent changes, which require pauses in the application cycle for end users.
However, a configuration that does respect business cycles and time zones and distributes the model can be beneficial to the business as business regions that are in down-time (e.g., in the middle of their night, where usage is very low) can, independently, carry out updates to data and metadata without affecting other regions.
ALM application: Metadata time cycle differences
Alternatively, ALM prevents pauses in the application cycle altogether by providing a development environment for each model. You may edit development models at any time without disrupting live production models for end users. Then, once you have completed your edits on the development model, you may deploy them to live production models without any disruptions or down-time for end users. As a result, using ALM removes any risk for pauses in the application cycle for any user at any time.
Regional / business process differences
Similar to the workforce planning example above, regional differences may exist. It may not be practical to attempt to include all regional variances that exist across countries for workforce planning in a single instance. Much of the functionality would not be relevant to every region, and so confusion and frustration would occur, as well as complication of user interface. In this instance a distributed model would be the best solution.
Another consideration is that of differing business processes. That is to say, both processes are intrinsically the same, but different enough to warrant separate treatment and business processes that are completely different.
An example of this may be a process where a business updates a forecast. Perhaps they get to the same point in a revenue forecast, but how different parts or divisions of a business get to that point is different. One may do an initial bottom-up forecast, submit up to management for draft approval, and then do a final submit. Another may do a top-down approach where they set a target and that target needs to be validated. These are connected, yet separate, processes that may warrant separate instances of an application.
ALM application: Regional / business process differences
If regional and business processes are similar between satellite models, and the metadata between them can be synced from a single development (primary) model, then ALM can be used to develop, test, and produce the single development model that feeds the satellite models.
If the regional and/or business processes cannot conform to use the same metadata from a single development model, then multiple development models must be used. In this case, ALM would be used to update, test, and produce each development model, which would then feed into each respective satellite model.
There is an easy way to see to which dashboards a module has been published. This can be particularly helpful when you are making changes to a module and need to know which dashboards the changes could impact. It can also be useful to reduce sparsity by identifying modules that might not be needed within a model. In other words, if a module is not used for any dashboards you can check to see if it’s needed for anything else and if it’s not, eliminate it.
There are two different types of distributed models to consider as early as possible when a client chooses to implement Anaplan:
A split model is where one model, known as the primary model, is partitioned into multiple satellite models that contain the exact same structure or metadata (such as versions and dimensions) as the primary model. The split models will be 90% identical to the primary model and will have about a 10% difference. The split model method is most common when a client's workspace involves multiple regions.
For example, the primary model may contain three different product lines. Region 1 sells product lines A and B, while Region 2 sells only product C. In this case, a split model may provide consistency in structure across the models, but variation with the product lines since not all product lists are applicable to each region.
ALM application: Split models
For split models, ALM allows clients to maintain the primary model as well as all satellite models in their workspace using one development model. Clients may make changes to their development model, and then deploy updates to their live models without disrupting the application cycle.
Similar models are models that vary slightly in structure or metadata. The degree of difference is usually less than 5%. If it gets to be greater than this, or there’s a greater difference in user experiences, it may be impractical to use similar models. For example, you could use the similar models method if you have multiple regions that must view the same data, ideally from a master data hub.
ALM spplication: Similar models
For similar models, ALM requires clients to maintain one development model for each similar model in use. Comparable to split models, each development model may be edited, tested, and then deployed to the production model without disrupting the application cycle.
It is important to understand what Application Lifecycle Management, or ALM, enables clients to do within Anaplan.
In short, ALM enables clients to effectively manage the development, testing, deployment, and ongoing maintenance of applications in Anaplan. With ALM, it is possible to introduce changes without disrupting business operations by securely and efficiently managing and updating your applications with governance across different environments and quickly deploying changes to run more “what-if” scenarios in your planning cycles as you test and release development changes into production.
Learn more here: Understanding model synchronization in Anaplan ALM
Training on ALM is also available in the Education section 313 Application Lifecycle Management (ALM)
Have you ever wondered where, within a model, a list property is in use? The Referenced By property will tell you!
Within Model Settings select the desired list and click on the Properties tab.
From here just look for the column labeled Referenced By. It displays where the list is currently in use or being referenced.
This is especially useful if you want to edit or delete a property but you don’t know if it’s being used. Please note this same feature is available for list subsets.
Have you ever wondered where, within a model, a line item or line item subset is in use? The Referenced By property will tell you!
Open the model which contains the line item.
Toggle Blueprint mode on and look for the column labeled Referenced By. It displays where the line item is currently in use or being referenced.
Functional areas should be sorted by grouping dashboards and modules separately. Doing this allows for quick access to dashboards, as well as improved control over user access assignments of these areas.
Use the Reorder button to sort the functional areas. Select the rows that should be moved and then click the Reorder button to choose where to move them to.
The Contents panel provides end users with links to dashboards and modules that are accessible by their user role. Workspace administrators should remove all unnecessary dashboards and modules for each role to keep the navigation options succinct. Always keep the Contents panel in line with the business process and the user role.
You should create user roles for each business function. You should then apply Selective Access to all lists, which helps to control the access that each end user needs. Avoid creating different roles, with varying access rights, for the same type of end user. This will help avoid the need for additional model maintenance.
Sort roles in a sensible fashion using the Reorder button (e.g. most privileges, some privileges, least privileges).
Consider using a module to control user access. This will allow model builders to provide clear instructions on the roles and access rights in the model, along with the ability to change user access rights from a convenient dashboard. Additionally, you can create an import and run it as part of a process to import user access from this module. Note that only model builders will have access to import data into the user list.
More information on User Roles and Selective Access can be found in Learning Center under Advanced Topics.
When user roles are given access to lists (for edit), memory is pre-allocated for those users to increase model size. Give user role access only to the lists that they will possibly update through actions.