## Line Charts: Throughput Comparison of ForkKV vs Prefix Caching
### Overview
The image contains two line charts comparing the throughput performance of two systems: ForkKV (green line) and Prefix Caching (blue line). Chart (a) examines performance across LoRA Rank values, while chart (b) analyzes performance across Output Length values. Both charts use throughput (tokens/second) as the y-axis metric.
### Components/Axes
**Chart (a): LoRA Rank**
- X-axis: LoRA Rank (values: 16, 32)
- Y-axis: Throughput (tokens/s) (0-40 scale)
- Legend: Top position, labels "ForkKV (ours)" (green) and "Prefix Caching (SGLang)" (blue)
- Data points: Circular markers connected by lines
**Chart (b): Output Length**
- X-axis: Output Length (values: 512, 1024)
- Y-axis: Throughput (tokens/s) (0-80 scale)
- Legend: Top position, same color coding as chart (a)
- Data points: Circular markers connected by lines
### Detailed Analysis
**Chart (a) Trends:**
1. **ForkKV (green):**
- At LoRA Rank 16: ~30 tokens/s
- At LoRA Rank 32: ~25 tokens/s
- Trend: Gradual decline (slope: -1.25 tokens/s per rank unit)
2. **Prefix Caching (blue):**
- Constant at ~10 tokens/s across both ranks
- Trend: Flat line (no change)
**Chart (b) Trends:**
1. **ForkKV (green):**
- At Output Length 512: ~20 tokens/s
- At Output Length 1024: ~70 tokens/s
- Trend: Linear increase (slope: +0.033 tokens/s per token)
2. **Prefix Caching (blue):**
- At Output Length 512: ~15 tokens/s
- At Output Length 1024: ~25 tokens/s
- Trend: Gradual increase (slope: +0.05 tokens/s per token)
### Key Observations
1. ForkKV demonstrates superior scalability with longer output lengths (chart b), achieving 3.5x higher throughput at 1024 tokens vs 512 tokens
2. Prefix Caching maintains consistent performance across LoRA ranks but shows diminishing returns with longer outputs
3. ForkKV's throughput advantage grows significantly with increased output length (chart b)
4. Both systems show stable performance at lower LoRA ranks (chart a)
### Interpretation
The data suggests ForkKV's architecture is more efficient for longer output generation tasks, with throughput increasing linearly with output length. In contrast, Prefix Caching maintains stable but lower performance across both tested parameters. The performance gap between the two systems widens dramatically at longer output lengths (chart b), indicating ForkKV's potential advantage in handling complex generation tasks. The stable performance of Prefix Caching across LoRA ranks suggests it may be more suitable for simpler, shorter generation tasks where output length is less variable.
