How does landscape calculate memory usage?

I'm trying to debug an OOM situation in an Ubuntu 12.04 server, and looking at the Memory graphs in Landscape, I noticed that there wasn't any serious memory usage spike. Then I looked at the output of the free command and I wasn't quite sure how both memory usage results relate to each other.

Here's landscape's output on the server:

$ landscape-sysinfo 
  System load:  0.0               Processes:           93
  Usage of /:   5.6% of 19.48GB   Users logged in:     1
  Memory usage: 26%               IP address for eth0: -
  Swap usage:   2%

Then I ran the free command and I get:

$ free -m
             total       used       free     shared    buffers     cached
Mem:           486        381        105          0          4        165
-/+ buffers/cache:        212        274
Swap:          255          7        248

I can understand the 2% swap usage, but where does the 26% memory usage come from?

Solution 1:

In Landscape

landscape-sysinfo actually gets its data directly from /proc/meminfo:

dpb@helo:~$ cat /proc/meminfo |egrep 'MemTotal:|Active:'
MemTotal:       12286760 kB
Active:          3794832 kB
dpb@helo:~$

The calculation of "Memory Usage" in this case would be:

((MemTotal - Active) / MemTotal) * 100

You can see these calculations in:

/usr/share/pyshared/landscape/sysinfo/memory.py
/usr/share/pyshared/landscape/lib/sysstats.py

gets its data directly from /proc/meminfo:

In free

The free utility also gets its data directly from /proc/meminfo:

Mem

total: MemTotal
used: MemTotal - MemFree
free: MemFree
buffers: Buffers
cached: Cached

Buffers/cache

used: MemTotal - MemFree - Buffers - Cached
free: MemFree + Buffers + Cached

Swap

total: SwapTotal
used: SwapTotal - SwapFree
free: SwapFree

Total

total: MemTotal + SwapTotal
used: MemTotal - MemFree + SwapTotal - SwapFree
free: MemFree + SwapFree

Corrected cached -- lzap

Solution 2:

Those graphs don't reflect every single memory allocation/freeing event, but samples from /proc/meminfo (exactly as dpb described) at given intervals. A slightly speculative explanation for why it's not showing in the graph would be that it simply occurred between to points when memory usage was sampled.

I suspect that what has happened here is that some process acquired lots of memory in a hurry and OOM killer disposed of it before a sample could be made. That would be a fairly extreme circumstance, and also one that would mean the whole machine was running slowly as it was swapping heavily. This loading on the machine would reduce the likelihood of the system having available time to sample memory usage during that window and report it back to the Landscape server.