Size tmpfs File System for Hana Fast Restart

[maximino_suarez]

Hi experts,

We want to use Fast Restart in order to speed up the Hana restart. Our system has 11 TB of RAM, the backup size is 7 TB and now the memory used is 6,3 TB with peaks of 10 TB. As the system has 8 numa nodes we decide to limite the size of the tmpfs file systems creating 8  with 625 GB in each (5 TB in total)

When we request our operating system team to implement them, they answer that the system has only 2,6 TB of free RAM and this is less than the 5 TB for the tmpfs. In fact, the exit of the command 'free -h' is:

                             total   used   free  
Mem:                    11T     8.7T   2.6T 
-/+ buffers/cache:             1.3T 10T
Swap:                    0           0      0

Now we are concern if this 5 TB of tmpfs can be a problem with this amount of free RAM

What do you think about? 

Regards, Maxi

[HakanHaslaman]

The tmpfs filesystem, stored in RAM, directly affects available RAM. Allocating 5TB to tmpfs reserves that RAM, making it unavailable for other processes. Your system has 2.6TB of free RAM, so allocating 5TB to tmpfs would overcommit your RAM, potentially causing errors and instability. This is likely why your OS team is reluctant to implement the requested tmpfs size. Consider reducing your tmpfs size to fit within your available RAM, increasing your RAM for larger tmpfs, or using a different method for faster Hana restarts that requires less RAM. Ensure your tmpfs size aligns with your free RAM to avoid issues.

[maximino_suarez]

Thank Hakan for answering,

This is our main concern: overcome the available RAM. But in 2700084 - FAQ: SAP HANA Persistent Memory I saw the next:
"Data in persistent memory isn't redundant to data in DRAM. Column store main storages stored in persistent memory aren't stored in DRAM"

So, I understand that the tables loaded in tmpfs are not in RAM and after a Hana restart, the resident memory will be release and the free RAM will increase too.

Do you think this will be correct?

Regards.