Poor performance while many users

[Katlitipkins]

Hello everyone!

We are currently migrating from QlikView to Qlik Sense. We’ve prepared a super-powerful servers in cluster, but we’re facing application performance issues. When developing an application and testing it, everything runs quickly. Then, we test it with colleagues, there are about 10 of us, and everything starts to slow down terribly. Interestingly, the server handling the calculations is loaded to only about 40% at peak. Telemetry logs show that everything is calculated very quickly, but on the screen, it lags horribly. We’re baffled... We don’t even know where to start troubleshooting. Please help.

[korsikov]

По букве Ы предположил что родной язык русский.
Точно ли Ваши новые сервера мощнее чем те что были? Базовая тактовая частота выше чем на старых серверах? Это железные сервера или виртуализация? Когда начинает тормозить? какие параметры мониторите? ЦП ОЗУ? Добавьте еще дисковые очереди.
Приложения, они имеют идентичную модель данных? Можете начать с сравнения скорости работы идентичного скрипта загрузки в QV и QS? 
Если отбросить детали то рекомендации общие такие:

сравнить производительность старых и новых серверов без ПО

сравнить производительность на максимально эталонном приложении. 
определить в чем разница в производительности? в скорости работы скриптом или отклик интерфейса для конечных пользователей?

From the letter 'Ы,' I assumed your native language is Russian. Are you certain your new servers are more powerful than the previous ones? Is the base clock speed higher than on the old servers? Are these physical servers or virtualized? Do the applications use the same data model? Could you start by comparing the speed of an identical load script in QV and QS?

In short, my general recommendations are:

- Compare the performance of the old and new servers without any software.
- Compare performance on a etalon Qlik application.
- Identify the main performance differences—is it  Load script  speed or user interface response time?

[Katlitipkins]

Thank You for the quick response.

An identical application in Qlik Sense works about the same as in QlikView when it’s used by 1-3 users. The problems begin when more users start working with the application. We plan to compare the servers without software soon. We have a virtualized cluster, and the clock speed is approximately the same. The performance just drops drastically, not just somewhat worse. The application itself takes 79GB in memory, and in QlikView, we get a 5-second response time, whereas here, similar actions take minutes. RAM usage is at 10%, and CPU doesn’t even approach 100%. We’re inclined to think it’s some hardware settings, but it’s unclear which ones.

[Katlitipkins]

What is also strange is that the cache does not accumulate on our nodes. At some point, it starts to build up, but after a minute or two, it is cleared. Could it be a memory settings of VM? The cache purge settings in QMC are set to the default - 28,800 seconds

[korsikov]

are you sure it's qlik sense cache? Is the memory consumption of Engine.exe increasing or the overall memory consumption of the whole system? There were no reload at this point - working only with published sheets and objects?

Maybe it's the memory consumed for calculating hypercubes

[Katlitipkins]

My colleagues report that on our dedicated QV node for users, we observe a memory distribution where the blue line represents the suggested memory cache. This indicates the expected memory consumption, which is above the 70% RAM limit.

And on the QS node, it's completely different; memory consumption never reaches the 70% low limit, which is strange. This node is dedicated solely to user analysis, with no reloads ever occurring

[korsikov]

Don't mix apples and oranges.

What you highlighted in the screenshot is the operating system cache and has nothing to do with Qlik. The computation cache for Qlik is stored within the engine.exe process. It can't be isolated separately, just as you can't isolate memory consumed by individual user sessions.

The screenshots you provided show the total memory usage by all processes in the system. How can you be certain that this isn’t memory reserved by the disk subsystem driver for NVMe disks? When discussing memory usage for Qlik's calculations and application processing, you need to look at the memory consumed by the engine.exe process.

Please take a look at the Windows Memory Manager to understand these parameters. You may also find tools like Process Explorer and RAMMap useful for managing memory.

[Katlitipkins]

Thank you. You may definitely be right, and I will try using the tools you mentioned. We're actually completely lost in this situation and are trying to guess at any possible clues. 🙂