Disaggregated Accelerator Resource Allocation for Shared Memory Bottlenecks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing accelerator devices face inefficiencies in dynamically reallocating shared resources like high bandwidth memory among logic portions, leading to delays and suboptimal resource utilization due to static allocation methods, which hinder performance and increase costs.
Innovation Solution
Implementing a system with an orchestrator server and dynamic resource allocation logic within accelerators that monitor and adjust resource utilization thresholds in real-time, allowing for dynamic allocation of resources among logic portions without stopping workload execution, thereby optimizing resource use and reducing the need for additional resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If static allocation of shared resources is used among logic portions, then resource allocation is simple and stable, but resource utilization efficiency deteriorates when workload needs change
Solution Approach 1:
The patent implements dynamic resource allocation where the resource allocation manager continuously monitors workload requirements and adjusts the allocation of shared high bandwidth memory to logic portions in real-time based on current needs, transforming the static allocation system into a dynamic one that adapts to changing conditions without requiring full reconfiguration of the logic portions
Solution Approach 2:
The system changes the allocation parameters of shared resources (specifically high bandwidth memory allocation ratios) dynamically based on workload characteristics, allowing each logic portion to receive appropriate memory resources according to its current computational demands rather than fixed predetermined allocations
2Adaptability or versatility
If reconfiguration of logic portions is performed to reallocate shared resources, then resource allocation adapts to workload needs, but execution time increases due to stopping and resuming workloads
Solution Approach 1:
The resource allocation manager performs preliminary monitoring of workload requirements and proactively adjusts resource allocations before performance degradation occurs, allowing smooth transitions without interrupting logic portion execution or requiring workload restarts
Solution Approach 2:
The patent introduces a resource allocation manager as an intermediary layer between the logic portions and shared high bandwidth memory, which handles all resource allocation decisions and adjustments without requiring the logic portions themselves to be reconfigured or restarted, thus decoupling resource management from logic portion execution
3Quantity of substance
If static resource allocation is used, then system operation is simple and stable, but additional resources must be provisioned to handle peak demands, increasing costs
Solution Approach 1:
The shared high bandwidth memory serves multiple logic portions simultaneously with dynamic allocation, allowing the same physical memory resources to be reused across different workloads and logic portions based on current demands, eliminating the need for each logic portion to have dedicated static memory allocations for peak loads
Data Source
AI summary
Technologies for dynamically managing resources in disaggregated accelerators include an accelerator. The accelerator includes acceleration circuitry with multiple logic portions, each capable of executing a different workload. Additionally, the accelerator includes communication circuitry to receive a workload to be executed by a logic portion of the accelerator and a dynamic resource allocation logic unit to identify a resource utilization threshold associated with one or more shared resources of the accelerator to be used by a logic portion in the execution of the workload, limit, as a function of the resource utilization threshold, the utilization of the one or more shared resources by the logic portion as the logic portion executes the workload, and subsequently adjust the resource utilization threshold as the workload is executed. Other embodiments are also described and claimed.


