Weight Generating Module for Disaggregated Host Machine Allocation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In disaggregated hardware systems, the non-uniform access latencies between CPU and memory pools lead to unpredictable performance of host machines, as resources are allocated unevenly, resulting in varying performance even for machines with identical specifications.
Innovation Solution
A Weight Generating Module calculates a set of weights based on user-defined policies and allocation weights to distribute CPU-memory pairs, ensuring predictable and uniform performance by optimizing resource allocation according to expected demands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If resources are allocated from pools in a disaggregated hardware system, then resource flexibility and scalability are improved, but access latency becomes non-uniform leading to unpredictable performance
Solution Approach 1:
The patent applies local quality by assigning different weight values to different resource pairs based on their specific latency characteristics. Each CPU-memory pair is evaluated individually and assigned weights that reflect its local performance characteristics, allowing the system to account for non-uniform access latencies while maintaining overall resource flexibility.
Solution Approach 2:
The patent changes the parameter of resource allocation by introducing weight values that modify how resources are selected from pools. Instead of uniform allocation, the system uses latency-based weights to adjust the probability of selecting specific CPU-memory pairs, thereby transforming the allocation behavior to achieve predictable performance while maintaining flexibility.
2Quantity of substance
If host machines are allocated with identical numbers of CPUs and memory units, then hardware specifications are uniform, but performance varies due to different access latencies
Solution Approach 1:
The patent applies local quality by recognizing that not all CPU-memory pairs are equal even when the quantities are identical. Each pair is assigned a weight based on its specific latency characteristics, allowing the system to differentiate between locally optimal and suboptimal resource combinations while maintaining uniform hardware specifications.
Solution Approach 2:
The patent introduces asymmetry in the resource allocation process by using non-uniform weight distributions for CPU-memory pairs. This asymmetric weighting scheme allows the system to favor certain pairs with lower latencies while still allocating the same number of resources, thereby achieving performance consistency without changing hardware quantities.
3Device complexity
If memory units are allocated without considering access latency, then allocation process is simple, but performance becomes unpredictable
Solution Approach 1:
The patent applies preliminary action by pre-calculating weight values for all CPU-memory pairs based on their latency characteristics before actual resource allocation occurs. This preliminary weighting step creates a lookup structure that guides subsequent allocation decisions, making the performance prediction possible without adding significant complexity to the real-time allocation process.
Solution Approach 2:
The patent introduces weights as an intermediary element between the raw latency measurements and the resource allocation decision. These weights serve as a mediating parameter that translates latency characteristics into allocation probabilities, simplifying the overall process while ensuring performance predictability.
Data Source
AI summary
A Weight Generating Module for generating a respective set of weights representing a respective policy of a plurality of policies. The respective set of weights defines a respective host machine. The Module calculates a respective user-defined number of pairs for each collection of a number of collections by distributing a total number of pairs among the number of collections for each respective policy based on a respective set of user-defined weights and a respective allocation weight. The Module selects, for said each collection, the respective user-defined number of pairs in increasing order with respect to the latencies to obtain a respective sub-set of collection weights relating to latencies associated with said each collection. The Module determines, for said each respective policy, the respective set of weights representing the respective policy based on a set of collection weights. The Module provides, for said each respective policy, the respective set of weights.