|
| 1 | +# ElasticDL: Kubernetes-Native Distributed Deep Learning with Elastic Scheduling |
| 2 | + |
| 3 | +## Description |
| 4 | + |
| 5 | +In addition to online services, users have been running distributed AI jobs on |
| 6 | +Kubernetes. Related projects add Kubernetes operators to start distributed deep |
| 7 | +learning programs calling TensorFlow or PyTorch. In these solutions, Kubernetes |
| 8 | +plays the role to launch pods and restart preempted ones. However, such retrials |
| 9 | +often fail due to the same reason that caused the perception -- the lack of |
| 10 | +resources. If couldn't maintain the constant number of worker pods, jobs fail. |
| 11 | +Maintaining the constant number, it becomes gang scheduling. Either case leads |
| 12 | +to low utilization of clusters. |
| 13 | + |
| 14 | +ElasticDL changes the paradox by using the statistical properties of distributed |
| 15 | +learning theory and making jobs tolerable to varying numbers of workers. Taking |
| 16 | +it as the basis, ElasticDL realizes elastic scheduling by introducing a master |
| 17 | +pod per job, to replace the Kubernetes operator per cluster. It makes full use |
| 18 | +of residual resources and improve the utilization significantly. |
| 19 | + |
| 20 | +## Benefits to the Ecosystem |
| 21 | + |
| 22 | +ElasticDL boosts the cluster utilization up to 90%, on on-premise clusters at Ant |
| 23 | +Financial and on Google Cloud, as it makes full use of residual resources to run |
| 24 | +deep learning jobs with elastic scheduling. Moreover, it enables the running of |
| 25 | +deep learning jobs in lower priority than online services on the same cluster. |
| 26 | +It senses and uses resource left by online services. |
| 27 | + |
| 28 | +The master is important to elastic scheduling, it takes three roles. |
| 29 | + |
| 30 | +1. The master dynamically partitions the training data so to decouple the number |
| 31 | +of partitions and that of the workers. |
| 32 | +2. The master also works with Kubernetes to watch the cluster utilization and a |
| 33 | +good chance to restart failed workers. Before the chance, the master always |
| 34 | +leverages living workers. |
| 35 | +3. The master starts and monitors parameter servers when training large models |
| 36 | +using the asynchronous SGD algorithm, and cooperate workers to implement a |
| 37 | +Kubernetes-native fault-tolerable AllReduce operation for the synchronous SGD |
| 38 | +counterpart. |
| 39 | + |
| 40 | +Deep learning researchers like ElasticDL as it reduces the pending time of each |
| 41 | +job as it makes full use of residual resources to run as many concurrent |
| 42 | +experiments as possible. Deep learning jobs depend on many parameters, |
| 43 | +including the optimizer, activator, and cost, hyperparameters. Users are eager |
| 44 | +to see the status of the first few iterations of the learning job so to ensure |
| 45 | +the configuration is mathematically correct. ElasticDL provides a cure. Using |
| 46 | +the residual resource, ElasticDL also shortens the total time to run a batch |
| 47 | +of training jobs. |
| 48 | + |
| 49 | +ElasticDL provides an easy-to-use interface. Users define models using |
| 50 | +TensorFlow 2.x API just like filling the map and reduce functions required by |
| 51 | +the MapReduce framework, without the need to consider anything about |
| 52 | +distributed programming. The interface allows users to test their models |
| 53 | +locally and run big data using ElasticDL without changing their source code. |
| 54 | + |
| 55 | +## Open Source Projects |
| 56 | + |
| 57 | +[ElasticDL](https://github.com/sql-machine-learning/elasticdl) |
| 58 | + |
| 59 | +## Speakers |
0 commit comments