The Run:AI Scheduler
At the heart of the Run:AI solution is the Run:AI scheduler. The scheduler is the gatekeeper of your organization's hardware resources. It makes decisions on resource allocations according to pre-created rules.
The purpose of this document is to describe the Run:AI scheduler and explain how resource management works.
Run:AI differentiates between two types of deep learning workloads:
- Interactive build workloads. With these types of workloads, the data scientist opens an interactive session, via bash, Jupyter notebook, remote PyCharm or similar and accesses GPU resources directly. Build workloads typically do not tax the GPU for a long duration. There are also typically real users behind an interactive workload that need an immediate scheduling response.
- Unattended (or "non-interactive") training workloads.Training is characterized by a deep learning run that has a start and a finish. With these types of workloads, the data scientist prepares a self-running workload and sends it for execution. Training workloads typically utilize large percentages of the GPU. During the execution, the researcher can examine the results. A Training session can take anything from a few minutes to a couple of weeks. It can be interrupted in the middle and later restored.
It follows that a good practice for the researcher is to save checkpoints and allow the code to restore from the last checkpoint.
Projects are quota entities that associate a project name with a deserved GPU quota as well as other preferences.
A researcher submitting a workload must associate a project with any workload request. The Run:AI scheduler will then compare the request against the current allocations and the project's deserved quota and determine whether the workload can be allocated with resources or whether it should remain in a pending state.
For further information on projects and how to configure them, see: Working with Projects
A Department is a second hierarchy of resource allocation above Project. A Department quota supersedes a Project quota in the sense that if the sum of Project quotas for Department A exceeds the Department quota -- the scheduler will use the Department quota rather than the Project quota.
For further information on departments and how to configure them, see: Working with Departments
Pods are units of work within a Job.
- Typically, each Job has a single Pod. However, in some scenarios (see Hyperparamter Optimization and Distribute Training below) there will be multiple Pods per Job.
- All Pods execute with the same arguments as added via
runai submit. E.g. The same image name, the same code script, the same number of Allocated GPUs, memory.
Basic Scheduling Concepts¶
Interactive vs. Unattended¶
The Researcher uses the --interactive flag to specify whether the workload is an unattended "train" workload or an interactive "build" workload.
- Interactive workloads will get precedence over unattended workloads.
- Unattended workloads can be preempted when the scheduler determines a more urgent need for resources. Interactive workloads are never preempted.
Guaranteed Quota and Over-Quota¶
Every new workload is associated with a Project. The project contains a deserved GPU quota. During scheduling:
- If the newly required resources, together with currently used resources, end up within the project's quota, then the workload is ready to be scheduled as part of the guaranteed quota.
- If the newly required resources together with currently used resources end up above the project's quota, the workload will only be scheduled if there are 'spare' GPU resources. There are nuances in this flow which are meant to ensure that a project does not end up with over-quota made fully of interactive workloads. For additional details see below
Allocation & Preemption¶
The Run:AI scheduler wakes up periodically to perform allocation tasks on pending workloads:
- The scheduler looks at each Project separately and selects the most 'deprived' Project.
For this deprived project it chooses a single workload to work on:
- Interactive workloads are tried first, but only up to the project's guaranteed quota. If such a workload exists, it is scheduled even if it means preempting a running unattended workload in this Project.
- Else, it looks for an unattended workload and schedules it on guaranteed quota or over-quota.
The scheduler then recalculates the next 'deprived' project and continues with the same flow until it finishes attempting to schedule all workloads
During the above process, there may be a pending workload whose project is below the deserved capacity. Still, it cannot be allocated due to the lack of GPU resources. The scheduler will then look for alternative allocations at the expense of another project which has gone over-quota while preserving fairness between projects.
The Run:AI scheduler determines fairness between multiple over-quota projects according to their GPU quota. Consider for example two projects, each spawning a significant amount of workloads (e.g. for Hyperparameter tuning) all of which wait in the queue to be executed. The Run:AI Scheduler allocates resources while preserving fairness between the different projects regardless of the time they entered the system. The fairness works according to the relative portion of GPU quota for each project. To further illustrate that, suppose that:
- Project A has been allocated with a quota of 3 GPUs.
- Project B has been allocated with a quota of 1 GPU.
Then, if both projects go over quota, project A will receive 75% (=3/(1+3)) of the idle GPUs and project B will receive 25% (=1/(1+3)) of the idle GPUs. This ratio will be recalculated every time a new job is submitted to the system or existing job ends.
This fairness equivalence will also be maintained amongst running jobs. The scheduler will preempt training sessions to maintain this equivalence
Bin-packing & Consolidation¶
Part of an efficient scheduler is the ability to eliminate defragmentation:
- The first step in avoiding defragmentation is bin packing: try and fill nodes (machines) up before allocating workloads to new machines.
- The next step is to consolidate jobs on demand. If a workload cannot be allocated due to defragmentation, the scheduler will try and move unattended workloads from node to node in order to get the required amount of GPUs to schedule the pending workload.
Run:AI Elasticity is explained here. In essence, it allows unattended workloads to shrink or expand based on the cluster's availability.
- Shrinking happens when the scheduler is unable to schedule an elastic unattended workload and no amount of consolidation helps. The scheduler then divides the requested GPUs by half again and again and tries to reschedule.
- Shrink jobs will expand when enough GPUs will be available.
- Expanding happens when the scheduler finds spare GPU resources, enough to double the amount of GPUs for an elastic workload.
Run:AI provides a Fractional GPU sharing system for containerized workloads on Kubernetes. The system supports workloads running CUDA programs and is especially suited for lightweight AI tasks such as inference and model building. The fractional GPU system transparently gives data science and AI engineering teams the ability to run multiple workloads simultaneously on a single GPU.
Run:AI’s fractional GPU system effectively creates virtualized logical GPUs, with their own memory and computing space that containers can use and access as if they were self-contained processors.
One important thing to note is that fraction scheduling divides up GPU memory. As such the GPU memory is divided up between jobs. If a Job asks for 0.5 GPU, and the GPU has 32GB or memory, then the job will see only 16GB. An attempt to allocate more than 16GB will result in an out-of-memorty exception.
GPU Fractions are scheduled as regular GPUs in the sense that:
- Allocation is made in fractions such that the total of the GPU allocation for a single GPU is smaller or equal to 1.
- Preemption is available for non-interactive workloads.
- Bin-packing & Consolidation work the same for fractions.
- Elasticity is not supported with fractions.
- Hyperparameter Optimization supports fractions.
Distributed Training, is the ability to split the training of a model among multiple processors. It is often a necessity when multi-GPU training no longer applies; typically when you require more GPUs than exist on a single node. Each such split is a pod (see definition above). Run:AI spawns an additional launcher processs which manages and coordinates the other worker pods.
Distribute Training utilizes a practice sometimes known as Gang Scheduling:
- The scheduler must ensure that multiple pods are started on what is typically multiple nodes, before the job can actually start.
- If one pod is preempted, the others are also immediately preempted.
Gang Scheduling essentially prevents scenarios where part of the pods are scheduled while other pods belonging to the same job are pending for resources to become available; scenarios that can cause deadlock situations and major inefficiencies in cluster utilization.
The Run:AI system provides:
- Inter-pod communication.
- Command-line interface to access logs and an interactive shell.
For more information on Distributed Training in Run:AI see here
Hyperparameter optimization (HPO) is the process of choosing a set of optimal hyperparameters for a learning algorithm. A hyperparameter is a parameter whose value is used to control the learning process, to define the model architecture or the data pre-processing process, etc. Example hyperparameters: learning rate, batch size, different optimizers, number of layers.
To search for good hyperparameters, Researchers typically start a series of small runs with different hyperparameter values, let them run for a while and then examine results to decide what works best.
With HPO, the researcher provides a single script which is used with multiple, varying, parameters. Each run is a pod (see definition above). Unlike Gang Scheduling, with HPO, pods are independent. They are scheduled independently, started and end independently, and if preempted, the other pods are unaffected. The scheduling behavior for individual pods are exactly as described in the Scheduler Details section above for Jobs.
For more information on Hyperparameter Optimization in Run:AI see here