Accelerator Resource Orchestration via Communication Abstraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Accelerator devices in data centers face limitations in accessing and cooperating with a large percentage of other accelerator devices due to their limited communication faculties, restricting their ability to execute workloads across diverse physical communication paths.
Innovation Solution
The implementation of a system that manages and allocates disaggregated resources across multiple sleds within a data center, utilizing an orchestrator server to dynamically allocate and deallocate resources, and employing communication abstraction logic units to establish logical communication paths between accelerator devices, enabling efficient workload distribution and resource utilization across heterogeneous devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If accelerator devices are configured to establish connections with other accelerator devices through physical communication paths, then communication between accelerator devices is enabled, but the ability to access a large percentage of accelerator devices in the data center is limited due to restricted communication faculties
Solution Approach 1:
The patent introduces a communication abstraction layer that acts as an intermediary between accelerator devices and the physical communication infrastructure. This abstraction layer translates diverse physical communication paths (PCIe, CXL, NVLink, etc.) into a unified logical communication interface, enabling accelerator devices to access a broader range of devices without direct configuration of each physical path. The intermediary handles protocol translation and path selection, resolving the contradiction between expanded access capability and reduced communication complexity.
2Productivity
If accelerator devices are designed to perform operations efficiently, then workload execution performance is improved, but communication with accelerator devices of other types through different physical communication paths is restricted
Solution Approach 1:
The patent implements a universal communication abstraction layer that enables a single accelerator device to communicate with multiple types of accelerator devices through different physical communication paths. The abstraction layer provides multi-functionality by supporting various protocols (PCIe, CXL, NVLink, HSA) through a unified interface, allowing workload execution efficiency to be maintained while expanding communication compatibility across heterogeneous devices.
3Ease of operation
If communication protocols are explicitly defined in the kernel by the developer, then communication functionality is established, but the communication aspect is limited to a single type of accelerator device kernel
Solution Approach 1:
The patent employs protocol translation and abstraction where the communication abstraction layer creates logical copies or representations of physical communication paths. Instead of requiring the kernel to directly configure each physical protocol, the abstraction layer translates high-level communication requests into appropriate physical protocol configurations. This copying mechanism maintains ease of operation for developers while enabling compatibility with multiple accelerator device types through protocol translation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Technologies for managing accelerator resources in a computing environment include an orchestrator having circuitry. According to one embodiment, the circuitry is to monitor resource usage of an accelerator kernel configured on a source accelerator device. The circuitry is to determine whether the resource usage exceeds a threshold specified in one or more policies. Upon a determination that the resource usage exceeds the threshold, the circuitry is to identify a target accelerator device to which to migrate the accelerator kernel. The circuitry migrates the accelerator kernel from the source accelerator device to the target accelerator device.