TM1 memory usage

[Istvan]

Hi All,

I've got a potentially very dumb question.
The TM1 model at my current client is 15Gb big if I check the total memory usage with the performance monitor in TM1.
On the other hand the tm1sd windows service uses almost 21Gb RAM. I thought they should be more or less the same.
Can anyone tell me what is the cause of this difference? Currently we are close to reach the total physical memory installed in our TM1 server, that's why it's quite crucial to reduce the RAM usage of TM1.
I am not sure whether it's relevant but we have:
- parallel interaction switched on
- increased the maximum paging file size to 30Gb to support spreading on big cubes
- also increased slightly the VMM attribute of some cubes to improve query time (helped big time)
- there is sufficient disk space on the TM1 server
- feeders look OK

Thanks,
Istvan

[Duncan P]

Just for interest Istvan, how many raw detail and how many calculated (including consolidated) cells do you have in the model?

[tomok]

If you are running out of memory all you can really do is cut down the size of your model or add more RAM to the box, if that's an option. You could reduce the VMM setting but I wouldn't recommend doing that. Your'e just experiencing one of the potential drawbacks of an in-memory tool. It happens to the best of us.

[Istvan]

Thanks for the help.
I've switched off the parallel interaction parameter out of curiosity and surprisingly the RAM usage went down by 7GB.
I don't quite understand it so I need to go back to the drawing board and read more forum topics about the parallel interaction best practices.
But now I deserve a bottle of quality red wine.
Thanks again

[Harvey]

You say the feeders look ok, but over-feeding is usually the bottleneck of any tm1 model with excessive RAM usage. You may get some improvements changing settings, but if I were you, I'd spend some time looking for suspect feeders.

[Istvan]

I've already gone through the feeders and made as efficient as possible.
My original problem was that I didn't understand why the tm1sd windows service showed 6GB more RAM usage than the TM1 performance monitor (}StatsByCube cube).
And found it has something to do with parallel interaction, as when it's switched off the memory usage goes down considerably.

[Andy Key]

}StatsByCube doesn't include garbage memory. Look at }StatsByServer and I suspect you will find your missing 6GB.

As far as ParallelInterction is concerned, you should expect a server with PI to use up to 30% more RAM than one without it.

[Martin Erlmoser]

2 other hints:
- re-ordering dimensions of the bigger cubes (if the order is not optimized)
- turn maximumcubeloadthreads to 1 (or delete it) and groupscreationlimit to 20 if its higher -- if you don't use these parameters, you don't have to change anything of course.


maximumcubeloadthreads can create tons of garbagememory and a overall high ram useage.