Simple cluster manager (SCM) is a small C program to manage a really small cluster.
Supercomputer employs the job schedule system like slurm as the manager.
For a small cluster, there is no need for such system. On a small cluster, the user normally 1) submit a job using command like mpirun -n 40 -host node4 ~/code/prog_run 2) login to a compute node and then run it.
Under such circumstances, the user want to know which compute node is available. The plain approach is to login to a compute node and check if this node is running any computing job by top.
SCM is a small utility program reducing the efforts for the user to know which node is available.
SCM currently supports the following features
- Deploy the server on all the compute nodes and start the client on the master node
- Run the client on the master node to print the CPU loading of all compute nodes in the terminal
- Use pthread to send/receive sockets to/from all the compute nodes to reduce the communication time. 8 nodes cost almost no additional time. 1 second is spent on calculating the CPU usage by reading
/proc/stat2 times. The interval between 2 reading operations is 1 second. Maybe a better way of calculating the CPU loading could be done.
The output is as follows,
Simple Cluster Manager (SCM) shows CPU loading of compute nodes.
SCM is implemented by Jingchang Shi.
Go to https://github.com/desperadoshi/SimpleClusterManager/issues for issues.
It is possible that due to the slow network, you have to wait for max 2 seconds of timeout.
=== Explanations begin ===
CPU usage is in [0, 100%]
If CPU usage is about 85% on a node,
that node is probably running all its 40 CPU cores.
If CPU usage is less than 1% on a node,
that node is probably occupied by nobody.
=== Explanations end ===
node1: Not connected!
node7: CPU usage = 100.00%
node5: CPU usage = 100.00%
node6: CPU usage = 100.00%
node2: CPU usage = 0.05%
node3: CPU usage = 0.07%
node4: CPU usage = 0.10%
node8: CPU usage = 0.07%
TODO for SCM
- Parse a configuration file instead of hard coding the configuration
scm_shownodeuser implements the same feature.
For example, on our small cluster, there is a master and 7 compute nodes node[2-8].
Thus run scm_shownodeuser gives the following output.
| Node No. | CPU usage | Users |
| node2 | .5 | User1 |
| node3 | 2.1 | User1 User2 |
| node4 | 4.3 | User1 User4 User2 |
| node5 | 3884.7 | User2 |
| node6 | 3988.4 | User3 |
| node7 | 3892.8 | User2 |
| node8 | .2 | User1 |