High-Performance Scheduling

Application Scenario 1: Hybrid Deployment of Multiple Types of Jobs

Multiple types of domain frameworks are developed to support business in different industries. These frameworks, such as Spark, TensorFlow, and Flink, function irreplaceably in their service domains. They are not working alone, as services and businesses are becoming increasingly complex. However, resource scheduling becomes a headache as clusters in these frameworks grow larger and a single service may have fluctuating loads. Therefore, a unified scheduling system is in great demand.

Volcano abstracts a common basic layer for batch computing based on Kubernetes. It supplements Kubernetes in scheduling and provides flexible and universal job abstractions for computing frameworks. These abstractions (Volcano Jobs) are implemented through multi-task templates to describe multiple types of jobs (such as TensorFlow, Spark, MPI, and PyTorch). Different types of jobs can be run together, and Volcano uses its unified scheduling system to realize cluster resource sharing.

Application Scenario 2: Scheduling Optimization in Multi-Queue Scenarios

Resource isolation and sharing are often required when you use a Kubernetes cluster. However, Kubernetes does not support queues. It cannot share resources when multiple users or departments share a machine. Without queue-based resource sharing, HPC and big data jobs cannot run.

Volcano supports multiple resource sharing mechanisms with queues. You can set the weight for a queue. The cluster allocates resources to the queue by calculating the ratio of the weight of the queue to the total weight of all queues. You can also set the resource capability for a queue to determine the upper limit of resources that can be used by the queue.

For example, in the following figure, queue 1 is allocated 40% of the cluster resources, and 60% for queue 2. In this way, two queues can be mapped to different departments or projects to use resources in the same cluster. If a queue has idle resources, they can be allocated to jobs in another queue.

Application Scenario 3: Multiple Advanced Scheduling Policies

Containers are scheduled to nodes that satisfy their requirements on compute resources, such as CPU, memory, and GPU. Normally, there will be more than one qualified node. Each could have different volume of available resources left for new workloads. Volcano automatically analyzes the resource utilization of each scheduling plan and help you achieve the optimal deployment results in great ease.

The following figure shows how the Volcano scheduler schedules resources. First, the scheduler loads the pod and PodGroup information in the API server to the scheduler cache. In a scheduler session, Volcano goes through three phases: OpenSession, action calling, and CloseSession. In OpenSession, the scheduling policy you configured in the scheduler plugin is loaded. In action calling, the configured actions are called one by one and the loaded scheduling policy is used. In CloseSession, final operations are performed to complete scheduling.

Volcano scheduler provides plugins to support multiple scheduling actions (such as enqueue, allocate, preempt, and backfill) and scheduling policies (such as gang, priority, drf, proportion, and binpack). You can configure them as required. The APIs provided by the scheduler can also be used for custom development.

Application Scenario 4: High-Precision Resource Scheduling

Volcano provides high-precision resource scheduling policies for AI and big data jobs to improve compute efficiency. Take TensorFlow as an example. Configure affinity between ps and worker and anti-affinity between ps and ps, so that ps and worker to the same node. This improves the networking and data interaction performance between ps and worker, thereby improving the compute efficiency. However, when scheduling pods, the default Kubernetes scheduler only checks whether the affinity and anti-affinity configurations of these pods conflict with those of all running pods in the cluster, and does not consider subsequent pods that may also need scheduling.

The task-topology algorithm provided by Volcano calculates the task and node priorities based on the affinity and anti-affinity configurations between tasks in a job. The task affinity and anti-affinity policies in a job and the task-topology algorithm ensure that the tasks with affinity configurations are preferentially scheduled to the same node, and pods with anti-affinity configurations are scheduled to different nodes. The difference between the task-topology algorithm and the default Kubernetes scheduler is that the task-topology algorithm considers the pods to be scheduled as a whole. When pods are scheduled in batches, the affinity and anti-affinity settings between unscheduled pods are considered and applied to the scheduling processes of pods based on priorities.

Application Scenario 5: Hybrid Deployment of Online and Offline Jobs

Many services see surges in traffic. To ensure performance and stability, resources are often requested at the maximum needed. However, the surges may ebb very shortly and resources, if not released, are wasted in non-peak hours. Especially for online jobs that request a large quantity of resources to ensure SLA, resource utilization can be as low as it gets. Resource oversubscription is the process of making use of idle requested resources. Oversubscribed resources are suitable for deploying offline jobs, which focus on throughput but have low SLA requirements and can tolerate certain failures. Hybrid deployment of online and offline jobs in a cluster can better utilize cluster resources.

The default scheduler of Kubernetes schedules resources by pod, regardless of the type of services running in pods. This mechanism fails in hybrid deployment. Here is where Volcano comes in. Volcano provides intelligent algorithms to schedule jobs by type based on how many resources are required and available. Idle cluster resources can be further used with mechanisms such as resource preemption and time division multiplexing.

Benefits

Running containers on high-performance GPU-accelerated cloud servers significantly improves AI computing performance by three to five folds. GPUs can cost a lot and sharing a GPU among containers greatly reduces AI computing costs. In addition to performance and cost advantages, CCE also offers fully managed clusters that will hide all the complexity in deploying and managing your AI applications so you can focus on high-value development.

Advantages

By integrating Volcano, CCE has the following advantages in running high-performance computing, big data, and AI jobs:

• Hybrid deployment of HPC, big data, and AI jobs
• Optimized multi-queue scheduling: Queues can be effectively managed for scheduling jobs. Complex job scheduling capabilities such as queue priority and multi-level queues are supported.
• Advanced scheduling policies: gang scheduling, fair scheduling, resource preemption, and GPU topology
• Multi-task template: You can use a template to define multiple tasks in a single Volcano Job, beyond the limit of Kubernetes native resources. Volcano Jobs can describe multiple job types, such as TensorFlow, MPI, and PyTorch.
• Job extension plugins: The Volcano Controller allows you to configure plugins to customize environment preparation and cleanup in stages such as job submission and pod creation. For example, before submitting a common MPI job, you can configure the SSH plugin to provide the SSH information of pod resources.
• Hybrid deployment of online and offline jobs: Hybrid deployment is supported, and CPU and memory resources can be oversubscribed to better utilize cluster resources.

Related Services

GPU-accelerated Cloud Server (GACS), Elastic Load Balance (ELB), and Object Storage Service (OBS)

Figure 1 How AI computing works