Comparative Performance Analysis of Modern NoSQL Data Technologies: Redis, Aerospike, and Dragonfly

Basic Information

• Paper ID: 2510.08863
• Title: Comparative Performance Analysis of Modern NoSQL Data Technologies: Redis, Aerospike, and Dragonfly
• Authors: Deep Bodra (Harrisburg University of Science and Technology), Sushil Khairnar (Virginia Tech)
• Classification: cs.DB cs.DC
• Published Journal: Journal of Research, Innovation and Technologies, Volume IV, Issue 2(8), 2025
• Paper Link: https://doi.org/10.57017/jorit.v4.2(8).05

Abstract

With the emergence of distributed applications and cloud computing, the demand for scalable, high-performance key-value storage systems has grown substantially. This paper evaluates the performance of three major NoSQL key-value storage systems—Redis, Aerospike, and Dragonfly—using the Yahoo! Cloud Serving Benchmark (YCSB) framework. The study conducted extensive experiments across three distinct workload patterns (read-intensive, write-intensive, and balanced), systematically varying client concurrency from 1 to 32 connections. The evaluation methodology captures latency, throughput, and memory characteristics under realistic operational conditions, providing profound insights into the performance trade-offs and scalability behaviors of each system.

Research Background and Motivation

Problem Definition

1. Modern Application Requirements Challenges: The contemporary digital environment involves massive data creation and utilization. The rapid expansion of web applications, mobile technologies, and IoT devices presents new challenges to database systems.
2. Limitations of Traditional Databases: While traditional relational database management systems are feature-rich, they encounter difficulties in meeting the performance and scalability requirements of modern applications, particularly those demanding sub-millisecond response times and processing millions of operations per second.
3. Rise of NoSQL Databases: NoSQL databases, particularly key-value stores, overcome these challenges by emphasizing performance and scalability.

Research Significance

• Practical Value: Provides practical guidance for system architects in selecting appropriate key-value storage solutions.
• Academic Value: Fills the gap in systematic comparative evaluation of Redis, Aerospike, and Dragonfly systems.
• Technical Value: Reveals performance characteristics of each system through systematic evaluation across different workload patterns and concurrency levels.

Limitations of Existing Research

Although these systems are widely used, there is a lack of comprehensive comparative studies systematically evaluating their performance characteristics across various workload patterns and concurrency levels.

Core Contributions

1. Comprehensive Performance Comparison: Provides complete performance comparative analysis including latency and throughput metrics.
2. Memory Consumption Characteristics Analysis: Conducts in-depth analysis of memory usage patterns and efficiency across the three systems.
3. Multi-Workload Evaluation: Performs systematic evaluation under read-intensive, write-intensive, and balanced workload scenarios.
4. Scalability Analysis: Reveals scaling characteristics of each system through testing with 1-32 concurrent clients.
5. Practical Guidance: Provides actionable guidance for system architects in selecting appropriate key-value storage solutions.

Methodology Details

System Overview

Redis:

• Open-source in-memory data structure store, developed in 2009
• Single-threaded architecture, eliminating complex locking mechanisms but limiting multi-core scalability
• Supports multiple data structures: strings, hashes, lists, sets, sorted sets, etc.
• Implements persistence through periodic snapshots or append-only files

Aerospike:

• Distributed NoSQL database, founded in 2009
• Hybrid memory architecture: DRAM for index storage, SSD for data storage
• Shared-nothing architecture with independent node operation
• Provides strong consistency and automatic failover capabilities

Dragonfly:

• In-memory data store launched in 2022, designed as a direct Redis replacement
• Multi-threaded, shared-nothing architecture capable of leveraging multiple CPU cores
• Protocol-compatible with Redis
• Implements sophisticated memory management and lock-free data structures

Experimental Setup

Hardware Environment:

• System: Mac OS with Apple M3 Pro chip
• Configuration: 12 cores, 36GB RAM, macOS Sequoia
• Deployment: Docker containers ensure consistent and isolated environments

Benchmark Framework:

• Yahoo! Cloud Serving Benchmark (YCSB)
• Two-phase approach: loading phase populates initial data, run phase executes benchmark operations
• Concurrency levels: 1, 2, 4, 8, 16, 32 concurrent clients
• Key selection distribution: Zipfian distribution, simulating realistic non-uniform access patterns

Workload Configuration

Read-Intensive Workload:

• 95% read, 5% update operations
• 1KB data per record (10 fields, 100 bytes each)
• Load 1,474,560 records
• Simulates caching scenarios, content delivery systems, etc.

Balanced Workload:

• 50% read, 50% update operations
• Same 1KB record structure
• Represents mixed access patterns in social media platforms, collaborative applications, etc.

Write-Intensive Workload:

• 10% read, 90% insert operations
• Time-series data with 64 fields, 8 characters per field
• Run phase executes 2,949,120 insert operations
• Simulates high-throughput data ingestion scenarios in IoT applications, monitoring systems, etc.

Experimental Results

Read-Intensive Workload Performance

Aerospike Performs Optimally:

• P99 latency: 436ms (single client) to 2,979ms (32 clients)
• Throughput: 3,348 ops/s to 32,592 ops/s
• Performance advantage stems from hybrid memory architecture and shared-nothing design

Redis Shows Moderate Performance:

• P99 latency: 862ms to 4,447ms
• Throughput: 1,656 to 17,158 ops/s
• Single-threaded architecture becomes a performance bottleneck under high concurrency

Dragonfly Exhibits Highest Latency:

• P99 latency: 1,137ms to 4,883ms
• Throughput: 1,371 to 16,328 ops/s
• Multi-threaded coordination overhead offsets parallel processing advantages

Balanced Workload Performance

Performance Hierarchy Remains Consistent:

• Aerospike: P99 latency 441ms-2,409ms, throughput 3,372-33,741 ops/s
• Redis: P99 latency 874ms-4,017ms, throughput 1,664-17,004 ops/s
• Dragonfly: P99 latency 1,187ms-4,631ms, throughput 1,278-16,497 ops/s

Write-Intensive Workload Performance

All Systems Perform Optimally:

• Aerospike: P99 latency 410ms-2,233ms, throughput 3,562-34,896 ops/s
• Redis: P99 latency 808ms-3,547ms, throughput 1,757-17,170 ops/s
• Dragonfly: P99 latency 1,124ms-3,859ms, throughput 1,331-16,925 ops/s

Memory Consumption Analysis

Key Findings:

• Aerospike demonstrates highest memory efficiency, benefiting from hybrid storage model
• Redis incurs largest memory overhead, reflecting limitations of single-node in-memory storage
• Dragonfly falls between the two, with additional overhead from multi-threaded coordination structures

Scalability Analysis

Throughput Scaling Characteristics:

• Aerospike: Near-linear scaling with 9-10x improvement
• Redis: 10-11x improvement, but with more pronounced latency growth
• Dragonfly: 12-13x improvement, but with lower baseline performance

Related Work

The paper cites multiple relevant studies:

1. Benchmark Frameworks: Cooper et al. (2010) YCSB framework establishes the foundation for cloud service system benchmarking
2. NoSQL Comparative Studies: Anthony & Rao's empirical comparison of key-value stores
3. System-Specific Research: Volminger (2021) Aerospike research, Charan et al. Redis analysis
4. Recent Developments: Mohan et al. (2024) NoSQL evaluation for OLAP workloads

Conclusions and Discussion

Main Conclusions

1. Aerospike Leads Comprehensively: Demonstrates optimal performance across all workloads and concurrency levels, with superior throughput scalability and relatively low latency.
2. Redis Remains Stable and Reliable: Exhibits consistent and predictable performance across all workload patterns, though constrained by single-threaded architecture.
3. Dragonfly Shows Promise and Challenges: Despite modern design, exhibits poor latency performance but demonstrates potential in write-intensive scenarios.
4. Workload Impact is Significant: All databases perform optimally under write-intensive conditions.

Practical Guidance

• Maximum Performance Requirements: Select Aerospike
• Operational Simplicity Priority: Redis suffices for most requirements
• Redis Compatibility Requirements: Dragonfly is an interesting option, but requires careful evaluation for latency-sensitive applications

Limitations

1. Single-Machine Test Environment: All tests conducted on a single machine, failing to fully demonstrate distributed system advantages.
2. Limited Network Conditions: Does not account for network latency and partition effects on performance.
3. Uniform Data Distribution: Uses only Zipfian distribution; real applications may exhibit different patterns.
4. Absence of Cluster Mode: Does not test genuine distributed deployment scenarios.

Future Directions

1. Production Environment Testing: Evaluate system performance under genuine production conditions.
2. Distributed Scenarios: Test true distributed scalability in cluster mode.
3. Consistency Model Research: Investigate CAP theorem implications on system design.
4. Fault Tolerance Mechanisms: Evaluate fault tolerance mechanisms during node failures.
5. Cross-Datacenter Replication: Assess data consistency and replication latency under network partitions.

In-Depth Evaluation

Strengths

1. Rigorous Methodology: Uses standard YCSB framework ensuring fair comparison.
2. Comprehensive Experiments: Covers multiple workloads and concurrency levels.
3. In-Depth Analysis: Provides not only performance data but also architectural explanations.
4. High Practical Value: Offers clear guidance for actual system selection.
5. Clear Writing: Well-structured with accurate technical descriptions.

Weaknesses

1. Environmental Limitations: Single-machine Docker environment fails to fully demonstrate distributed system advantages.
2. Uniform Configuration: Does not test impact of different configuration parameters on performance.
3. Persistence Omission: Lacks detailed evaluation of persistence mechanisms' performance impact.
4. Missing Cost Analysis: Does not consider hardware costs and operational complexity.
5. Long-Term Stability: Lacks long-duration stability testing.

Impact

1. Academic Value: Provides systematic methodology for NoSQL database performance research.
2. Practical Value: Offers reference for industrial selection of appropriate key-value storage systems.
3. Methodological Contribution: Demonstrates systematic approach to comparing NoSQL system performance.
4. Reproducibility: Detailed experimental setup descriptions facilitate reproduction and extension.

Applicable Scenarios

1. System Selection: Provides reference for projects requiring key-value storage system selection.
2. Performance Optimization: Offers benchmarks for existing system performance tuning.
3. Architecture Design: Provides basis for large-scale distributed system architecture design.
4. Academic Research: Provides foundational data and methodological reference for related research.

References

The paper cites multiple important references, including:

• Cooper, B.F. et al. (2010). Benchmarking cloud serving systems with YCSB
• Anthony, A., & Rao, Y. N. M. Memcached, Redis, and Aerospike Key-Value Stores Empirical Comparison
• Mohan, R. K. et al. (2024). Evaluating NoSQL Databases for OLAP Workloads
• Official documentation and technical materials for respective database systems

This paper makes valuable contributions to the NoSQL database performance evaluation field. Through systematic experimental design and in-depth analysis, it provides important reference for understanding performance characteristics of modern key-value storage systems and selecting appropriate technological solutions.