Modeling Heterogeneous Memory Deployment on Server Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In data center operations, the shuffle stage of big data analytical workloads often leads to memory overflow, resulting in disk spill, which significantly increases query execution time due to the performance gap between disk and memory storage, necessitating larger cluster sizes and higher memory capacities, but existing solutions lack effective methods to evaluate the Total Cost of Ownership (TCO) implications of heterogeneous memory deployments.
Innovation Solution
A computing system that models heterogeneous memory deployment on a server network by processing circuitry, which receives user inputs including memory ratios and throughput parameters to determine the optimal number of servers and calculate TCO savings, generating a server network design that balances memory capacity, server count, and network architecture while ensuring equivalent data throughput performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If memory capacity is increased to prevent disk spill and improve query performance, then query execution time is reduced, but hardware cost and TCO increase
Solution Approach 1:
The patent transitions from a single-dimension local memory architecture to a multi-dimensional memory hierarchy by introducing heterogeneous memory (CXL-attached DDR5) as a remote memory layer. This allows the system to access additional memory capacity beyond physical server boundaries while maintaining performance closer to local memory, effectively resolving the contradiction between memory capacity and execution time without linearly increasing hardware cost.
Solution Approach 2:
The patent creates a universal memory pool that can be dynamically shared across multiple servers in the cluster. The heterogeneous memory infrastructure serves multiple functions: it acts as extended memory for individual servers, provides a shared memory pool for cluster-wide operations, and enables flexible memory allocation based on workload demands, thereby improving query performance without requiring each server to have dedicated high-capacity memory.
2Productivity
If cluster size is increased to provide sufficient memory capacity, then performance targets are met, but hardware cost and infrastructure complexity increase
Solution Approach 1:
The patent merges multiple servers' memory resources into a unified heterogeneous memory pool accessible by all cluster nodes. By combining memory capacity across servers and enabling shared access through CXL interconnects, the system achieves the required memory capacity for high query throughput without proportionally increasing the number of physical servers, thus improving productivity while controlling infrastructure scale.
3Reliability
If local DDR5 memory deployment is increased to meet performance requirements, then memory capacity is sufficient, but TCO increases
Solution Approach 1:
The patent introduces CXL-attached heterogeneous memory as an intermediary layer between local memory and disk storage. This mediator provides a buffer that captures frequently accessed data that would otherwise require slow disk access, thereby maintaining performance reliability without requiring proportional increases in expensive local DDR5 memory capacity. The intermediary handles memory capacity demands that would otherwise require additional local memory hardware.
Data Source
AI summary
A computing system is provided for modeling a deployment of heterogeneous memory on a server network. The computing system receives a user input of parameters including a memory ratio of local memory to heterogeneous memory in each server, a first relative throughput when an entire dataset is in local memory on a server, and a second relative throughput when the entire dataset is in heterogeneous memory on the server. Based on these parameters, the system determines a server ratio of a number of servers in an enhanced cluster with heterogeneous memory to a number of servers in a baseline cluster without heterogeneous memory, where the enhanced cluster and the baseline cluster deliver equivalent data throughput performance. Based on the parameters and the server ratio, a server network design is generated and outputted.