Parallel Computer Job Migration Cache Index Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In parallel computing systems, memory relocation during job migration can lead to cache line overlap, resulting in frequent cache line rewriting and increased memory access latency, which prolongs job execution times due to the slashing effect in cache memory.
Innovation Solution
A parallel computer system that acquires index levels of memory areas accessed by a job on one node and relocates data to corresponding usable memory areas on another node to avoid cache line overlap, maintaining the relative position of index levels to prevent cache slashing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the relocation program relocates the memory area where the data accessed by the process of the job under the execution were memorized, then the job can be migrated to another node, but the address of the memory area for the access changes, causing cache line overlap and frequent cache line rewriting
Solution Approach 1:
The system performs preliminary analysis of cache index levels before memory relocation. By acquiring the group of index levels corresponding to memory areas accessed by the job on the source node, and predicting the group of index levels that would be accessed after relocation to the target node, the system can determine in advance whether cache line overlap will occur. This preliminary assessment enables informed decision-making about whether to proceed with relocation or select an alternative target node, thereby preventing cache slashing before it happens.
Solution Approach 2:
The system implements a feedback mechanism where the actual cache performance is monitored and used to guide future relocation decisions. By tracking memory access patterns and cache hit/miss ratios, the system can adjust its relocation strategy to avoid configurations that lead to cache line overlap. This feedback loop ensures that job migration decisions are continuously optimized based on actual system behavior rather than static predictions alone.
2Productivity
If the relocation program relocates the memory area, then the job can continue execution on a different node, but the cache line memorizing the data of plurality of memory areas may overlap, resulting in frequent cache line rewriting
Solution Approach 1:
The system applies preliminary anti-action by preventing cache line overlap before it can occur. Through cache index level analysis, the system identifies potential overlap scenarios in advance and takes corrective action by either selecting a different target node for relocation or adjusting the memory allocation strategy. This proactive approach counteracts the harmful cache slashing effect before it can degrade performance, ensuring smooth job migration without the penalty of frequent cache line rewriting.
3Ease of manufacture
If memory areas are relocated during job migration, then hardware resources can be reassigned efficiently, but the cache memory performance deteriorates due to overlapping cache lines and frequent rewriting
Solution Approach 1:
Before executing memory relocation, the system performs preliminary cache index level analysis to assess the impact on cache performance. By acquiring index levels from the source node and predicting index levels at the target node, the system can evaluate whether the relocation will cause cache line overlap. This preliminary assessment enables the system to choose relocation strategies that maintain both resource allocation efficiency and memory access performance, avoiding configurations that would degrade reliability.
Data Source
AI summary
A parallel computer includes a first node and a second node, each including a memory having a plurality of memory areas and a cache memory, and a processing unit that acquires a first group of index levels of the cache memory, the first group of index levels corresponding with addresses of first plurality of memory areas storing data accessed by a job in the first node, when continuing an execution of the job by migrating the job carried out on the first node to the second node, judges whether or not the second node has second plurality of memory areas that are a usable state corresponding to a second group of index levels that has same as or relative position relation with the first group of index levels, and relocates the data to the second plurality of memory areas when the second node has the second plurality of memory areas.