3 of 3 people found this helpful
Question (1) and (2) are very good questions that are, I believe, impossible to answer. The reason is that it all depends on:
1) How the analysis is configured
2) The data density of the inputs
3) The type of data reference of the inputs
4) Network performance
5) The performance of the PI Data Archive
6) The performance of the AF Server
7) The degree of dependency of the dependent analyses (though this is less important as of 2018 for global service performance).
8) The degree of dependency of the inputs
For example, we have customers running 1 million analyses and aren't encountering any performance issues. We have customers running several thousands calculations that are having performance issues. The easiest way to explain this is that if the analysis triggers at a faster rate than the time it takes to calculate the analysis, calculations will fall behind. If LoadSheddingEnabled is set to true, the service will skip evaluations to keep up. If it is disabled, or in the case of Event Frame analyses and upper rank dependent analyses, the service will fall behind.
In almost all analysis performance cases I've worked on, the root cause of a performance issue is attributed to a few badly configured analyses. As a super simple example, consider an analysis that does a TagAvg() of multiple PI Points over the time span of several years and is calculated every second. It's easy to see that this type of analysis would perform poorly. Of course, this configuration isn't very logical as it doesn't make sense to calculate a moving average of several years every second.
Question (3) is easier to answer, we currently don't support load balancing for the Analysis Service. We have the request on user voice here: Load balance PI Analytic Equation – Customer Feedback for OSIsoft & the PI System
We are working on a sizing spreadsheet to give estimates on the hardware requirements for Asset Analytics. I assumed this post was related to this one: AF analysis Trigger not working - PI Points,
If you were asking for a different or new system, could you specify how many expression, rollup and event frame analyses this system would scale to? I could provide some guidelines on the hardware required.
Thank you for detailed information related AF analysis. Currently we have 19 K calculations which are trigger based. We are expected another 20 K - 30K calculations in future. Current hardware contains with 32 GB RAM with 4 cores. As per suggestion in previous post : AF analysis Trigger not working - PI Points, we are planning to upgrade AF and check performance but before upgrade we would like to check all possible options like hardware sizing and configuration. As mentioned we will move some calculations (TagAvg,TagMax(*-7d) etc.) from trigger based to periodic and check if performance improves.
PI Data Archive and AF server are not loaded currently but will definitely check network latency between servers to understand if there is any bottleneck. We are planning to upgraded AF analysis 2017 R2 in production and would like to know if there are any significant changes in AF analysis 2018 compared to previous version.
:Lal Babu Shaik
I mentioned this in the other post, as of 2.9 you can manually set the number of evaluation threads to 4 times the number of cores. I haven't had a look at your system, but if you're facing a concurrency issue, I would recommend upgrading the software first so you can configure it to utilize the max potential of the hardware it is installed on.
Hope this helps,
Thank you for all recommendations.
Moved few calculations which are involving summary calculations (multiple PI Points over the time span) which calculated every second based on input trigger to periodic and left other inexpensive calculations as-is. We could notice a drop in latency and recalculation time.