Docker utilizes Linux namespace to isolate resources (network, filesystem and pid etc.) for containers. However traditional system performnce tools (uptime, top etc.) are unaware of the namespace. When runnig inside containers, they report the same metrics as running on host.
This project provide a solution to correctly get system performance metrics within each container, without any need to modifying any exisitng monitoring tools.
Each docker container belongs to one separate cgroup, resources consumed(CPU, Memory, etc) are recorded in cgroup files. Instead of reading system-wide metric sources (e.g, /proc/meminfo), calculating metrics from cgroup records results in more accurate results for each container. This is how most container monitor tools work.
What make the diffrence of this project is that it take usage of LD_PRELOAD to enable existed tools to calculating matric from cgroup records. On the initialization of inject.so, libc functions like open, lseek are hijacked. hijacked lib c calls will calculate from cgroup records once application try to read related system-wide metric sources, otherwise it will pass through request to original version.
> make && make test
-
run a container with required volumes and get right metric
> docker run -ti --rm -v <inject.so>:/usr/lib/inject.so -v /sys/fs/cgroup:/sys/fs/cgroup:ro ubuntu bash > # in container > export CONTAINER_PROC_INJECT_TARGETS=free:top:uptime > export LD_PRELOAD=/usr/lib/inject.so > uptime
- (DONE)open, fopen, and lseek on /proc/uptime
- (DONE)open, fopen, and lseek on /proc/cpuinfo
- (DONE)open, fopen, and lseek on /proc/meminfo
- (DONE)open, fopen, and lseek on /proc/stat
- (DONE)open, fopen, and lseek on /proc/diskstats
- (DONE)open, fopen, and lseek on /system/devices/system/cpu/online
- (DONE)sysinfo
- (DONE)sysconf(_SC_NPROCESSORS_ONLN)
- (TODO)open, fopen, and lseek on /proc/loadavg
- lxcfs: https://linuxcontainers.org/lxcfs/introduction/ Some metric calculating algorithms come from lxcfs