I have some hosts running MongoDB. Below is a picture monitoring the host's free memory.
It seems that when the free memory keep dropping below some threshold, the system will trigger some memory reallocated mechanism. But it just releases too fast (the orange line release 19G memory in few seconds). And that clears all the MongoDB buffer, increases query latencies.
How to solve this problem or limit the amount of memory release everytime (e.g. make the figure like a saw), like setting some parameters in the /proc/sys/vm/?