CXL Workload Distribution via Switch-Based Capacity Balancing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In heterogeneous computing systems, managing cache-coherence and efficiently distributing workloads across multiple CXL devices with varying memory capacities and processing rates is challenging, leading to potential bottlenecks and inefficiencies.
Innovation Solution
An apparatus and method that utilize a switch to connect CXL devices, allowing a first device to select and distribute work based on the usable capacity and processing rate of other devices, ensuring optimal workload distribution and reducing data skew by calculating the distribution amount to balance processing times across devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If works are distributed to multiple CXL devices with varying memory capacities and processing rates, then system throughput is improved, but workload imbalance and data skew occur
Solution Approach 1:
The patent applies local quality by customizing the workload distribution strategy according to the specific characteristics of each CXL device. The host device determines distribution targets and amounts based on individual device attributes including memory capacity, processing rate, and current workload status. This ensures that each device receives workloads appropriate to its local capabilities, preventing both overload and underutilization.
Solution Approach 2:
The patent implements dynamics by making the workload distribution adaptive and changeable. The distribution strategy is not fixed but dynamically adjusted based on real-time device status information. The host device can modify distribution targets and amounts as device conditions change, allowing the system to respond to varying workloads and device performance states.
2Quantity of substance
If work distribution is based on memory capacity alone, then devices with larger memory receive more work, but devices with faster processing rates are underutilized
Solution Approach 1:
The patent applies parameter changes by incorporating multiple device parameters into the distribution decision-making process. Instead of relying solely on memory capacity, the system considers both memory capacity and processing rate as key parameters. The host device uses these parameters to calculate appropriate distribution amounts that balance both storage and processing capabilities, ensuring optimal utilization of all device resources.
3Device complexity
If centralized work distribution is implemented by one device, then coordination is simplified, but the distributing device becomes a bottleneck
Solution Approach 1:
The patent applies self-service by enabling CXL devices to autonomously evaluate their own status and make informed decisions about workload acceptance and distribution. Each device can report its current state (memory usage, processing rate, workload status) to the host, and the host uses this self-reported information to determine optimal distribution strategies without requiring complex centralized coordination or intervention.
Data Source
AI summary
An apparatus including: a plurality of compute express link (CXL) devices each including a memory and a processor for processing works stored in the memory; and a switch configured to connect the CXL devices to each other, wherein a first CXL device among the plurality of CXL devices selects at least one second CXL device from at least some CXL devices of the plurality of CXL devices to distribute works stored in a memory of the first CXL device based on a usable capacity of a memory of the at least some CXL devices.