DC/OS Spark limits
Mesosphere has scale-tested Spark on DC/OS by running a CPU-bound Monte Carlo application on the following hardware:
Cluster characteristics
- 2560 cores total
- 40 m4.16xlarge EC2 instances
Single executor per node
- 40 executors
- Each executor: 64 cores, 2GB memory
- CPU utilization was > 90%, with majority of time spent in task computation
Multiple executors per node
On a smaller, 1024-core, 16 node (m4.16xlarge) cluster, the following variations were tested:
| Executors | Time to launch all executors | Executors per node |
|---|---|---|
| 82 | 7 s. | 16 |
| 400 | 17 s. | 64 |
| 820 | 28 s. | 64 |
In all tests, the application completed successfully.
Spark Documentation