Design docummentation for supporting gpu-memory on ray core resource scheduler.
Currently, ray supports num_gpus to scheduler resource to tasks/actors, which then assign either num_gpus number of gpu ids to be used by the tasks/actors. Additionally, ray provides support fractional gpus allocation by specifying num_gpus < 1 so a single GPU can be used to run multiple tasks. This works well if the cluster only has a single type of GPUs. However, imagining a cluster has both A100 40GB and A100 80GB GPUs, setting num_gpus to a fixed number doesn’t work that well: if we set to 0.1 then we will get 4GB if the scheduler picks A100 40GB but 8GB if the scheduler picks A100 80GB which is a waste of resource. We can also set accelerator_type to A100_40GB and num_gpus to 0.1 to make sure we get the exact amount of GPU memory we need but then the task cannot run on A100 80GB even if it’s free.
This new scheduler design will allows users to directly schedule fractional gpu resources by amount of memory. In our example, if user specify _gpu_memory = 20GB, then ray automatically convert the value to num_gpus depending on which nodes the request is assigned too. As example, if it's scheduled on A100 40GB node, then num_gpus = 0.5, otherwise if it's scheduled on A100 80GB node, then num_gpus = 0.25. As a result, user can schedule a fixed amount of GPU resources without depending on which types of GPUs the tasks/actos are scheduled to.
... issue with num gpus
Ray community users’s demand:
https://discuss.ray.io/t/gpu-memory-aware-scheduling/2922/5
https://discuss.ray.io/t/automatic-calculation-of-a-value-for-the-num-gpu-param/7844/4
Inside ray project since we want to add new parameter _gpu_memory to the ray remote function.
The proposal will be open to the public, but please suggest a few experienced Ray contributors in this technical domain whose comments will help this proposal. Ideally, the list should include Ray committers.
@pcmoritz, @jjyao, @scv119
To make the review process more productive, the owner of each proposal should identify a shepherd (should be a senior Ray committer). The shepherd is responsible for working with the owner and making sure the proposal is in good shape (with necessary information) before marking it as ready for broader review.
@jjyao
Users will be able to specify amount of gpu memory to their ray tasks/actors using gpu_memory on ray.remote. The specified gpu_memory will be the amount of gpu resources from a single gpu that will be allocated to users ray tasks/actors.
# Request a fractional GPU with specified gpu_memory in bytes.
# Mutually exclusive with num_gpus.
@ray.remote(_gpu_memory=1024 * 1024 * 1024) # 1 mb request
def task:
…When a Ray node is started, Ray will auto detect the number of GPUs and GPU memory of each GPU and set the GPU and _gpu_memory resources accordingly. Users also have the option to manually specify _gpu_memory resources as the sum of total gpu memory across all gpus in the node. The default value is num_gpus * total memory of the gpu type in the node.
ray start # auto detection
ray start –num_gpus=3 –gpu_memory=3000 * 1024 * 1024 * 1024 # manual override, each gpu has 1000mb total memoryNote that GPU memory and compute unit is 1-1 conversion, means 20GB of gpu memory is equivalent to 0.5 fractional value of an A100_40GB gpu. So, for simplicity and consistency, ray doesn't allow users to specify both num_gpus and _gpu_memory in a single ray task/actor.
# Request a fractional GPU both num_gpus and _gpu_memory is not allowed
@ray.remote(_gpu_memory=1024 * 1024 * 1024, num_gpus=0.5) # raise ValueError exception
def not_allowed_task:
…
# Request a fractional GPU with specified num_gpus.
@ray.remote(num_gpus=0.5)
def num_gpus_task:
…
# Request a fractional GPU with specified _gpu_memory.
@ray.remote(_gpu_memory=1024 * 1024 * 1024)
def gpu_memory_task:
…Additionally, users can still specify which gpu type they want to use by specifying accelerator_type.
# Request a fractional of A100 GPU with specified _gpu_memory
@ray.remote(_gpu_memory=1024 * 1024 * 1024 * 1024, accelerator_type="NVIDIA_A100")
def nvidia_a100_gpu_task:
…
# Requesting 30GB of gpu memory from a A10 GPU with 24GB of memory.
# Task won't be able to be scheduled.
@ray.remote(_gpu_memory=30 * 1024 * 1024 * 1024 * 1024, accelerator_type="NVIDIA_TESLA_A10G")
def nvidia_a10_gpu_task:
…TBD
We introduce a new ResourceID named GPU_Memory (gpu_memory in string) to specify the amount of GPU memory resources. GPU_Memory is treated as a GPU resource with a distinct representation, where the relationship is defined as GPU equals to GPU_Memory divided by the total memory of a single GPU in the node as GPU resources in the node is homogeneous. Despite their distinct representations, both GPU and GPU_Memory signify the same underlying resource, with the caveat that Ray currently only supports homogeneous GPU type for each node.
There are two pivotal options in how we store GPU_Memory in NodeResources:
We opt to store only GPU in NodeResource and convert gpu_memory to GPU means gpu_memory is ResourceRequest only resource. This implementation involves saving metadata of the single node's total_single_gpu_memory. ConvertRelativeResource converts gpu_memory in ResourceRequest to num_gpus based on the single node's total_single_gpu_memory during node feasibility checks and resource allocation.
# Suppose we have two nodes with GPU type A100 40GB and A100 80gb respectively
NodeResource(available={"GPU": [1,1,1]}, label={"gpu_memory_per_gpu": "40GB"})
NodeResource(available={"GPU": [1,1,1]}, label={"gpu_memory_per_gpu": "80GB"})
# gpu_memory request
task.options(_gpu_memory="10GB")
# equivalent resource request when scheduled in Node 1
ResourceRequest({"GPU": 0.25})
# remaining resources in Node 1, check using ray.available_resources()
NodeResource(available={"GPU": [0.75,1,1]})
# equivalent resource request when scheduled in Node 2
ResourceRequest({"GPU": 0.125})
# remaining resources in Node 2, check using ray.available_resources()
NodeResource(available={"GPU": [0.875,1,1]})
# Roundup gpu_memory request
task.options(_gpu_memory="10MB")
# equivalent resource request when scheduled in Node 1
ResourceRequest({"GPU": 0.00025})
# round up to nearest 10^{-4}
ResourceRequest({"GPU": 0.0003})
# remaining resources in Node 1, check using ray.available_resources()
NodeResource(available={"GPU": [0.9997,1,1]})
# equivalent resource request when scheduled in Node 2
ResourceRequest({"GPU": 0.000125})
# round up to nearest 10^{-4}
ResourceRequest({"GPU": 0.0002})
# remaining resources in Node 2, check using ray.available_resources()
NodeResource(available={"GPU": [0.9998,1,1]})Pros:
- Simplified Resource Model: Better emphasizes to new users that Ray only have
GPUto represent GPU resource, simplifying the resource model. - Straightforward Conversion:
GPU_Memoryis converted to GPU based on the single node's total_single_gpu_memory during node feasibility checks and resource allocation with the roundup logic applied underhood.
Cons:
- Limited Observability:
ray.available_resources()only displays remaining GPU resources in terms of percentage, without specific amounts forGPU_Memory. - Incompatibility with Heterogeneous GPUs: Doesn't work for heterogeneous GPUs in a single node, a limitation existing in Ray's current support.
We store GPU_Memory as part of the NodeResource. This implementation ensures synchronization between GPU and GPU_Memory. During node feasibility checks and resource allocation, the ConvertRelativeResource function performs two conversions: calculating gpu_memory if num_gpus is specified and vice versa.
# Suppose we have two nodes with GPU type A100 40GB and A100 80gb respectively
NodeResource(available={"GPU": [1,1,1], "gpu_memory": ["40GB", "40GB", "40GB"]})
NodeResource(available={"GPU": [1,1,1], "gpu_memory": ["80GB", "80GB", "80GB"]})
# gpu_memory request
task.options(_gpu_memory="10GB")
# equivalent resource request when scheduled in Node 1
ResourceRequest({"GPU": 0.25, "gpu_memory": "10GB"})
# remaining resources in Node 1, check using ray.available_resources()
NodeResource(available={"GPU": [0.75,1,1], "gpu_memory": ["30GB", "80GB", "80GB"]})
# equivalent resource request when scheduled in Node 2
ResourceRequest({"GPU": 0.125, "gpu_memory": "10GB"})
# remaining resources in Node 2, check using ray.available_resources()
NodeResource(available={"GPU": [0.875,1,1], "gpu_memory": ["70GB", "80GB", "80GB"]})
# Roundup gpu_memory request
task.options(_gpu_memory="10MB")
# equivalent resource request when scheduled in Node 1
ResourceRequest({"GPU": 0.00025, "gpu_memory": "10MB"})
# round up to nearest 10^{-4}
ResourceRequest({"GPU": 0.0003, "gpu_memory": "12MB"})
# remaining resources in Node 1, check using ray.available_resources()
NodeResource(available={"GPU": [0.9997,1,1], "gpu_memory": ["39.988GB", "40GB", "40GB"]})
# equivalent resource request when scheduled in Node 2
ResourceRequest({"GPU": 0.000125, "gpu_memory": "10MB"})
# round up to nearest 10^{-4}
ResourceRequest({"GPU": 0.0002, "gpu_memory": "16MB"})
# remaining resources in Node 2, check using ray.available_resources()
NodeResource(available={"GPU": [0.9998,1,1], "gpu_memory": ["79.984GB", "80GB", "80GB"]})Pros:
- Enhanced Observability: Users can see the remaining GPU memory resources after roundup allocation, providing detailed insights.
Cons:
- Synchronization Overhead: Requires synchronization for both
GPUandGPU_Memory, introducing an additional layer of complexity by updating bothGPUandGPU_Memoryfor rounding up. - Resource Duality: Users need to grasp that both resources,
GPUandGPU_Memory, essentially denote the same underlying resource.
The primary implementation entails the automatic detection of GPU memory during the initialization of a Ray cluster.
class AcceleratorManager:
# return 0 if accelerator is not GPU,
# else return total GPU memory of a single GPU
def get_current_node_gpu_memory(self):
... Subsequently, we added another change within the scheduler to convert gpu_memory to num_gpus depending on which node the request got assigned to check if resource request is feasible in the node and allocate the resource.
for node in node_list:
++ def convert_relative_resources(resource_request, node): # option 1 implementation
++ if gpu_memory in resource_request:
++ resource_request.num_gpus = roundup(gpu_memory / node.label["gpu_memory_per_gpu"] , 0.0001)
++ resource_request.gpu_memory = 0
++ return resource_request
++ convert_relative_resources(resource_request, node)
if check_is_feasible(resource_request):
allocation = TryAllocate(resource_request)
For the first option, we require to node label to store gpu_memory_per_gpu. However, currently label is not included in autoscaler as label-based node scheduling is yet to be supported for autoscaler. Therefore, for current solution, we introduce a new private resources node:gpu_memory_per_gpu__ as a constant label value representing gpu_memory_per_gpu node label. Next, we also add convert_relative_resources function before in _fits and _inplace_subtract in resource_demand_scheduler.py
def _convert_relative_resources(node, resources):
adjusted_resources = resources.copy()
if "gpu_memory" in resources:
if "node:gpu_memory_per_gpu" not in node or node["node:gpu_memory_per_gpu"] == 0:
return None
adjusted_resources["GPU"] = (
resources["gpu_memory"] / node["node:gpu_memory_per_gpu"]
)
del adjusted_resources["gpu_memory"]
return adjusted_resources
def _fits(node: ResourceDict, resources: ResourceDict) -> bool:
adjusted_resources = _convert_relative_resources(node, resources)
if adjusted_resources is None:
return False
for k, v in adjusted_resources.items():
...
def _inplace_subtract(node: ResourceDict, resources: ResourceDict) -> None:
adjusted_resources = _convert_relative_resources(node, resources)
if adjusted_resources is None:
return
for k, v in adjusted_resources.items():
...For the second option, there's no required addition of node:gpu_memory_per_gpu_ since GPU_Memory is part of resources, but the _convert_relative_resources still required.
TBD