Skip to contentSkip to conversation

:Containers(V2): Documentation

patch01/05 Rohit Seth 2 participants 4 messages Sep 20, 2006 Documentation

lore
Share & tools
lore

Rohit Seth

author
#0 lore raw

This patch contains the documentation for containers.

Documentation/containers.txt +65 -0
@@ -0,0 +1,65 @@
+Containers allow different workloads to be run on the same platform with
+limits defined on per container basis.  This basically allows a single
+platform to be (soft) partitioned among different workloads (each of
+which could be running many tasks).  The limits could be amount of
+memory, number of tasks among other features.  These two features are
+already implemented in the patch set that I posted.  But it is possible
+to add other controllers like CPU that allows only finite amount of time
+to the processes belonging to a container.
+
+For example, users can run batch jobs like backups using tar, which if run
+uncontained could use lot of page cache, inside a container.  This way the
+memory footprint of the backup job can be contained.
+
+Currently we are tracking user memory (both file based
+and anonymous).  The memory handler is currently deactivating pages
+belonging to a container that has gone over the limit. Even though this
+allows containers to go over board their limits but 1- once they are
+over the limit then they run in degraded manner and 2- if there is any
+memory pressure then the (extra) pages belonging to this container are
+the prime candidates for swapping (for example).  The statistics that
+are shown in each container directory are the current values of each
+resource consumption.
+
+Configfs support is needed in kernel as the container's user interface is 
+through configfs. So first enable CONFIG_CONFIGFS_FS and CONFIG_CONTAINERS 
+and recompile the kernel.
+
+1- Mount a configfs (for example):
+	mount -t configfs none /mnt/configfs
+   This will create a /mnt/configfs mount point.
+
+2- As the support of containers is built into kernel, so the mount point
+   /mnt/configfs will automatically contain a directory "containers"
+
+3- Create a container by name test_container
+	cd /mnt/configfs/containers
+	mkdir test_container
+
+All the current implemented attributes in the kernel will show up in the
+directory /configfs/containers/test_container
+
+4- Add a task to container
+	cd /mnt/configfs/cotnainers/test_container
+	echo <pid> > addtask
+
+Now the <pid> and its subsequently forked children will belong to container
+test_container.
+
+5- Remove a task from container
+	echo <pid> > rmtask
+
+6- Set a page limit for the container
+	echo some_number_of_pages > page_limit
+
+7- Read the id for the container
+	cat id
+
+8- Get the statistics for this container
+	cat num* (will print active pages, anon_pages, file_pages, num_files, 
+			and num_task)
+	cat *hits (will print page_limit_hits and task_limit_hits: the number
+		of times container has gone over page_limit and task_limit)
+9- Freeing a container
+	cd /mnt/configfs/containers/
+	rmdir test_container
#1 lore raw

Documentation/containers.txt line 22

Quoted 9 lines · Rohit Seth
Documentation/containers.txt +9 -0
+Currently we are tracking user memory (both file based
+and anonymous).  The memory handler is currently deactivating pages
+belonging to a container that has gone over the limit. Even though this
+allows containers to go over board their limits but 1- once they are
+over the limit then they run in degraded manner and 2- if there is any
+memory pressure then the (extra) pages belonging to this container are
+the prime candidates for swapping (for example).  The statistics that
+are shown in each container directory are the current values of each
+resource consumption.

Containers via cpusets allow a clean implementation of a restricted memory area. The system will swap and generate an OOM message if no memory can be recovered.

Documentation/containers.txt line 50

Quoted 9 lines

+4- Add a task to container

Documentation/containers.txt +8 -0
+	cd /mnt/configfs/cotnainers/test_container
+	echo <pid> > addtask
+
+Now the <pid> and its subsequently forked children will belong to container
+test_container.
+
+5- Remove a task from container
+	echo <pid> > rmtask

Could you make that compatible with the way cpusets do it?

Documentation/containers.txt line 65

Quoted 3 lines

+9- Freeing a container

Documentation/containers.txt +2 -0
+	cd /mnt/configfs/containers/
+	rmdir test_container

Adding and removal is the same way as cpusets.

author
#2 lore raw
Quoted 28 lines · Christoph Lameter

Documentation/containers.txt line 22

Quoted 9 lines · Rohit Seth
Documentation/containers.txt +9 -0
+Currently we are tracking user memory (both file based
+and anonymous).  The memory handler is currently deactivating pages
+belonging to a container that has gone over the limit. Even though this
+allows containers to go over board their limits but 1- once they are
+over the limit then they run in degraded manner and 2- if there is any
+memory pressure then the (extra) pages belonging to this container are
+the prime candidates for swapping (for example).  The statistics that
+are shown in each container directory are the current values of each
+resource consumption.

Containers via cpusets allow a clean implementation of a restricted memory area. The system will swap and generate an OOM message if no memory can be recovered.

Documentation/containers.txt line 50

Quoted 9 lines

+4- Add a task to container

Documentation/containers.txt +8 -0
+	cd /mnt/configfs/cotnainers/test_container
+	echo <pid> > addtask
+
+Now the <pid> and its subsequently forked children will belong to container
+test_container.
+
+5- Remove a task from container
+	echo <pid> > rmtask

Could you make that compatible with the way cpusets do it?

Quoted 5 lines

Documentation/containers.txt line 65

Quoted 3 lines

+9- Freeing a container

Documentation/containers.txt +2 -0
+	cd /mnt/configfs/containers/
+	rmdir test_container

Adding and removal is the same way as cpusets.

Most of the syntax of cpuset is in terms of nodes and CPU numbers. That is not the case here in containers.

-rohit

#3 lore raw
Quoted 2 lines · Rohit Seth

Most of the syntax of cpuset is in terms of nodes and CPU numbers. That is not the case here in containers.

Think about a container as a virtual node that can be configured to have a certain size.