## Bar Chart: Throughput Comparison Across Caching Strategies
### Overview
The image presents a grouped bar chart comparing throughput (tasks/second) for three caching strategies - ForkKV (ours), Prefix Caching (SGLang), and Prefix Caching (vLLM) - across two applications (ReAct and MapReduce) at different numbers of concurrent workflows (4, 8, 16). The chart uses three distinct colors for the strategies: green for ForkKV, teal for SGLang, and dark blue for vLLM.
### Components/Axes
- **X-axis**: Number of Concurrent Workflows (4, 8, 16)
- **Y-axis**: Throughput (tasks/second) with different scales:
- ReAct: 0.00-0.36
- MapReduce: 0.00-0.06
- **Legend**: Located at the top, with color-coded labels:
- Green: ForkKV (ours)
- Teal: Prefix Caching (SGLang)
- Dark Blue: Prefix Caching (vLLM)
### Detailed Analysis
#### ReAct Application
- **4 workflows**:
- ForkKV: ~0.18 tasks/s
- SGLang: ~0.35 tasks/s
- vLLM: ~0.34 tasks/s
- **8 workflows**:
- ForkKV: ~0.12 tasks/s
- SGLang: ~0.06 tasks/s
- vLLM: ~0.05 tasks/s
- **16 workflows**:
- ForkKV: ~0.09 tasks/s
- SGLang: ~0.04 tasks/s
- vLLM: ~0.03 tasks/s
#### MapReduce Application
- **4 workflows**:
- ForkKV: ~0.07 tasks/s
- SGLang: ~0.025 tasks/s
- vLLM: ~0.025 tasks/s
- **8 workflows**:
- ForkKV: ~0.045 tasks/s
- SGLang: ~0.015 tasks/s
- vLLM: ~0.012 tasks/s
- **16 workflows**:
- ForkKV: ~0.015 tasks/s
- SGLang: ~0.01 tasks/s
- vLLM: ~0.01 tasks/s
### Key Observations
1. **ForkKV Dominance**: ForkKV consistently outperforms both SGLang and vLLM across all workflow counts in both applications.
2. **Scalability Pattern**: Throughput decreases as workflow count increases, but ForkKV maintains a higher throughput margin.
3. **SGLang vs vLLM**: The two prefix caching strategies show similar performance, with SGLang slightly outperforming vLLM in ReAct at 4 workflows.
4. **Application-Specific Performance**: ReAct shows significantly higher throughput values than MapReduce across all configurations.
### Interpretation
The data suggests ForkKV demonstrates superior efficiency in handling concurrent workflows compared to traditional prefix caching strategies. The performance gap between ForkKV and others widens as workflow count increases, indicating better scalability. The minimal difference between SGLang and vLLM suggests similar underlying limitations in prefix caching approaches. The application-specific performance disparity (ReAct vs MapReduce) may reflect differences in task complexity or resource requirements between the two frameworks.
