Compression Ratio Selection for AI Data Staging Bottlenecks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current data compression techniques for parallel learning in AI systems face challenges in determining the optimal compression ratio to balance decompression cost and bandwidth, leading to inefficiencies in staging time and processing efficiency.

Innovation Solution

A computer-readable recording medium stores an information processing program that dynamically adjusts compression ratio settings based on actual compression ratios and decompression rates, determining the optimal setting to minimize staging time by considering the maximum total bandwidth of storage devices and the number of servers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If compression ratio is increased to reduce data transfer size, then bandwidth saturation is suppressed and staging time is reduced, but decompression processing time increases and processing efficiency deteriorates

Engineering Contradiction:
Improvestaging timeVSAvoidprocessing efficiency
Core Design Contradiction:
Loss of timeVSProductivity

Solution Approach 1:

The patent dynamically adjusts compression ratio parameters based on data characteristics and system conditions. The information processing apparatus selects optimal compression ratios by analyzing data types, access patterns, and current bandwidth availability, thereby balancing the trade-off between reducing transfer time and maintaining decompression efficiency.

Inventive Principle:
Principle #35Parameter changes

2Quantity of substance

If compression ratio is increased to reduce data transfer size, then bandwidth usage is optimized, but decompression cost increases

Engineering Contradiction:
Improvedata transfer sizeVSAvoiddecompression cost
Core Design Contradiction:
Quantity of substanceVSLoss of energy

Solution Approach 1:

The system dynamically changes compression parameters based on data characteristics and selects optimal compression algorithms. By analyzing data types and access patterns, the information processing apparatus adjusts compression intensity to minimize both transfer size and decompression energy consumption.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements feedback mechanisms where the system monitors decompression performance and bandwidth usage in real-time. Based on this feedback, the information processing apparatus adjusts compression ratio settings to optimize the balance between reduced data transfer and controlled decompression costs.

Inventive Principle:
Principle #23Feedback

3Loss of time

If data pieces are compressed to suppress bandwidth saturation, then staging time is reduced, but system complexity increases due to compression management

Engineering Contradiction:
Improvestaging timeVSAvoidcompression management complexity
Core Design Contradiction:
Loss of timeVSDevice complexity

Solution Approach 1:

The information processing apparatus automatically performs compression ratio selection and adjustment without requiring manual intervention. The system self-monitors bandwidth usage, data characteristics, and performance metrics, then autonomously optimizes compression settings, thereby reducing operational complexity while maintaining staging time benefits.

Inventive Principle:
Principle #25Self-service

Data Source

PatentUS11960449B2Computer-readable recording medium storing information processing program, information processing method, and information processing apparatus
Publication Date: 2024.04.16 FUJITSU LTD
  • US11960449B2 patent drawing
  • US11960449B2 patent drawing
  • US11960449B2 patent drawing

AI summary

A recording medium stores an information processing program for managing a plurality of storage devices and a plurality of servers. The program causes a computer to execute a process including: while changing a compression ratio setting, obtaining an actual compression ratio by using some of data pieces to be used by the plurality of servers and a decompression rate at which the servers decompress a compressed dataset in which the some data pieces are compressed; and determining the compression ratio setting to be used based on a maximum total bandwidth of the plurality of storage devices and a number of the plurality of servers by using the obtained actual compression ratio and the decompression rate for each of the compression ratio settings.