Hugging Face's logo

Datasets:
kshitijthakkar
/
mcp-server-bench-gradio-optimized-full-bench

Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
< 1K
Tags:
mcp
benchmark
gradio
fastmcp
api-performance
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
kshitijthakkar commited on
Commit
35d0a77
·
verified ·
1 Parent(s): b5e700b

Update benchmark report, charts, and detailed results

Browse files
Files changed (6) hide show
  1. README.md +173 -61
  2. detailed_results.json +0 -0
  3. plots/gradio_cl_scaling.png +3 -0
  4. plots/latency_comparison.png +3 -0
  5. plots/protocol_overhead.png +3 -0
  6. plots/throughput_comparison.png +3 -0
README.md CHANGED
@@ -1,63 +1,175 @@
1
  ---
2
- dataset_info:
3
- features:
4
- - name: scenario_id
5
- dtype: large_string
6
- - name: server
7
- dtype: large_string
8
- - name: protocol
9
- dtype: large_string
10
- - name: tool
11
- dtype: large_string
12
- - name: virtual_users
13
- dtype: int64
14
- - name: concurrency_limit
15
- dtype: float64
16
- - name: duration_s
17
- dtype: float64
18
- - name: total_requests
19
- dtype: int64
20
- - name: successful_requests
21
- dtype: int64
22
- - name: failed_requests
23
- dtype: int64
24
- - name: error_rate_pct
25
- dtype: float64
26
- - name: throughput_rps
27
- dtype: float64
28
- - name: latency_min_ms
29
- dtype: float64
30
- - name: latency_p50_ms
31
- dtype: float64
32
- - name: latency_p75_ms
33
- dtype: float64
34
- - name: latency_p90_ms
35
- dtype: float64
36
- - name: latency_p95_ms
37
- dtype: float64
38
- - name: latency_p99_ms
39
- dtype: float64
40
- - name: latency_max_ms
41
- dtype: float64
42
- - name: latency_mean_ms
43
- dtype: float64
44
- - name: latency_stddev_ms
45
- dtype: float64
46
- - name: avg_cpu_pct
47
- dtype: float64
48
- - name: avg_memory_mb
49
- dtype: float64
50
- - name: peak_memory_mb
51
- dtype: float64
52
- splits:
53
- - name: train
54
- num_bytes: 89349
55
- num_examples: 337
56
- download_size: 43894
57
- dataset_size: 89349
58
- configs:
59
- - config_name: default
60
- data_files:
61
- - split: train
62
- path: data/train-*
63
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ language:
3
+ - en
4
+ license: mit
5
+ tags:
6
+ - mcp
7
+ - benchmark
8
+ - gradio
9
+ - fastmcp
10
+ - api-performance
11
+ pretty_name: MCP Server Benchmark Results
12
+ size_categories:
13
+ - n<1K
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  ---
15
+
16
+ # 🔬 Gradio vs FastMCP Benchmark Report
17
+
18
+ **Generated:** 2026-03-02T21:04:48.460108
19
+
20
+ **Total scenarios:** 337
21
+
22
+ ## Executive Summary
23
+
24
+ - **echo**: Fastmcp wins (100.9 vs 189.9 RPS, 1.88x difference)
25
+ - Gradio best config: concurrency_limit=1.0
26
+ - **fibonacci**: Fastmcp wins (45.5 vs 55.0 RPS, 1.21x difference)
27
+ - Gradio best config: concurrency_limit=nan
28
+ - **json_transform**: Fastmcp wins (94.5 vs 165.4 RPS, 1.75x difference)
29
+ - Gradio best config: concurrency_limit=5.0
30
+ - **async_sleep**: Fastmcp wins (94.0 vs 104.3 RPS, 1.11x difference)
31
+ - Gradio best config: concurrency_limit=10.0
32
+ - **payload_echo**: Fastmcp wins (97.4 vs 172.0 RPS, 1.77x difference)
33
+ - Gradio best config: concurrency_limit=10.0
34
+
35
+ ## Throughput (Requests/Second)
36
+
37
+ | | ('fastmcp', 'http_api') | ('fastmcp', 'mcp_streamable') | ('gradio', 'http_api') | ('gradio', 'mcp_streamable') |
38
+ |:-----------------------|--------------------------:|--------------------------------:|-------------------------:|-------------------------------:|
39
+ | ('async_sleep', 1) | 16.15 | 12.27 | 8.72 | 14.2 |
40
+ | ('async_sleep', 10) | 104.32 | 0 | 60.07 | 94.02 |
41
+ | ('async_sleep', 25) | 54.87 | 0 | 38.35 | 59.43 |
42
+ | ('async_sleep', 50) | 49.07 | 14.67 | 34.16 | 53.27 |
43
+ | ('echo', 1) | 189.87 | 80.53 | 11.45 | 77.57 |
44
+ | ('echo', 10) | 79.67 | 50.13 | 53.64 | 100.86 |
45
+ | ('echo', 25) | 78.61 | 38.44 | 39.15 | 71.79 |
46
+ | ('echo', 50) | 63.15 | 32.53 | 34.48 | 66.26 |
47
+ | ('fibonacci', 1) | 43.7 | 31.06 | 12.08 | 36.39 |
48
+ | ('fibonacci', 10) | 54.99 | 43.91 | 38.65 | 45.5 |
49
+ | ('fibonacci', 25) | 53.34 | 30.89 | 34.5 | 45.02 |
50
+ | ('fibonacci', 50) | 45.36 | 26.2 | 31.99 | 43.81 |
51
+ | ('json_transform', 1) | 165.42 | 61.54 | 11.58 | 76.65 |
52
+ | ('json_transform', 10) | 85.96 | 35.91 | 54.97 | 94.48 |
53
+ | ('json_transform', 25) | 50.54 | 34.85 | 40.27 | 68.97 |
54
+ | ('json_transform', 50) | 60.12 | 29.89 | 34.32 | 65.31 |
55
+ | ('payload_echo', 1) | 172 | 41.58 | 11.45 | 73.3 |
56
+ | ('payload_echo', 10) | 77.79 | 0 | 51.81 | 97.35 |
57
+ | ('payload_echo', 25) | 68.79 | 31.16 | 39.1 | 71.09 |
58
+ | ('payload_echo', 50) | 61.18 | 31.2 | 34.25 | 67.32 |
59
+
60
+ ## Latency p50 (ms)
61
+
62
+ | | ('fastmcp', 'http_api') | ('fastmcp', 'mcp_streamable') | ('gradio', 'http_api') | ('gradio', 'mcp_streamable') |
63
+ |:-----------------------|--------------------------:|--------------------------------:|-------------------------:|-------------------------------:|
64
+ | ('async_sleep', 1) | 62.532 | 78.089 | 111.556 | 66.027 |
65
+ | ('async_sleep', 10) | 85.062 | 0 | 150.179 | 100.587 |
66
+ | ('async_sleep', 25) | 340.339 | 0 | 527.091 | 320.992 |
67
+ | ('async_sleep', 50) | 718.018 | 2516.25 | 1171.91 | 695.627 |
68
+ | ('echo', 1) | 5 | 11.973 | 81.08 | 12.232 |
69
+ | ('echo', 10) | 86.744 | 174.638 | 162.622 | 91.59 |
70
+ | ('echo', 25) | 235.09 | 542.953 | 519.866 | 262.988 |
71
+ | ('echo', 50) | 519.902 | 1329.38 | 1151.61 | 515.254 |
72
+ | ('fibonacci', 1) | 21.503 | 30.181 | 66.38 | 26.53 |
73
+ | ('fibonacci', 10) | 171.886 | 205.011 | 258.26 | 215.707 |
74
+ | ('fibonacci', 25) | 441.508 | 701.476 | 611.039 | 550.476 |
75
+ | ('fibonacci', 50) | 768.812 | 1612.96 | 1281.01 | 1147.82 |
76
+ | ('json_transform', 1) | 5.996 | 15.004 | 79.591 | 12.565 |
77
+ | ('json_transform', 10) | 62.526 | 234.142 | 157.188 | 95.055 |
78
+ | ('json_transform', 25) | 322.956 | 607.396 | 502.211 | 270.443 |
79
+ | ('json_transform', 50) | 536.465 | 1419.95 | 1190.15 | 507.173 |
80
+ | ('payload_echo', 1) | 5.266 | 22.11 | 80.573 | 13.012 |
81
+ | ('payload_echo', 10) | 90.521 | 0 | 167.601 | 94.079 |
82
+ | ('payload_echo', 25) | 214.665 | 650.125 | 526.828 | 269.352 |
83
+ | ('payload_echo', 50) | 523.993 | 1344.04 | 1184.89 | 501.297 |
84
+
85
+ ## Gradio concurrency_limit Scaling
86
+
87
+ How does Gradio's throughput change as concurrency_limit increases?
88
+
89
+ | | 1.0 | 5.0 | 10.0 |
90
+ |:-----------------------|--------:|--------:|--------:|
91
+ | ('async_sleep', 1) | 10.405 | 11.27 | 11.2825 |
92
+ | ('async_sleep', 10) | 30.7 | 72.29 | 71.0275 |
93
+ | ('async_sleep', 25) | 24.375 | 45.8375 | 46.4825 |
94
+ | ('async_sleep', 50) | 30.3925 | 41.5775 | 40.7275 |
95
+ | ('echo', 1) | 43.7875 | 42.05 | 33.84 |
96
+ | ('echo', 10) | 59.0175 | 73.0875 | 59.8225 |
97
+ | ('echo', 25) | 45.0575 | 51.41 | 44.3075 |
98
+ | ('echo', 50) | 42.3375 | 45.29 | 38.755 |
99
+ | ('fibonacci', 1) | 23.95 | 23.1425 | 20.555 |
100
+ | ('fibonacci', 10) | 31.4525 | 40.265 | 37.3475 |
101
+ | ('fibonacci', 25) | 31.0675 | 38.22 | 35.7975 |
102
+ | ('fibonacci', 50) | 30.54 | 36.47 | 31.64 |
103
+ | ('json_transform', 1) | 43.485 | 42.26 | 36.2125 |
104
+ | ('json_transform', 10) | 55.41 | 74.06 | 66.2875 |
105
+ | ('json_transform', 25) | 40.025 | 53.8475 | 49.5975 |
106
+ | ('json_transform', 50) | 34.0875 | 49.08 | 45.295 |
107
+ | ('payload_echo', 1) | 36.205 | 38.2725 | 41.5925 |
108
+ | ('payload_echo', 10) | 48.8175 | 66.65 | 67.315 |
109
+ | ('payload_echo', 25) | 38.595 | 51.07 | 50.155 |
110
+ | ('payload_echo', 50) | 39.38 | 46.105 | 45.0975 |
111
+
112
+ ## Protocol Overhead: HTTP API vs MCP
113
+
114
+ Comparing latency of the same tool called via REST API vs MCP protocol:
115
+
116
+ | | http_api | mcp_streamable |
117
+ |:------------------------------|-----------:|-----------------:|
118
+ | ('fastmcp', 'async_sleep') | 301.488 | 648.586 |
119
+ | ('fastmcp', 'echo') | 211.684 | 514.737 |
120
+ | ('fastmcp', 'fibonacci') | 350.927 | 637.407 |
121
+ | ('fastmcp', 'json_transform') | 231.986 | 569.122 |
122
+ | ('fastmcp', 'payload_echo') | 208.611 | 504.068 |
123
+ | ('gradio', 'async_sleep') | 920.038 | 332.935 |
124
+ | ('gradio', 'echo') | 693.001 | 246.16 |
125
+ | ('gradio', 'fibonacci') | 763.248 | 506.714 |
126
+ | ('gradio', 'json_transform') | 690.337 | 234.987 |
127
+ | ('gradio', 'payload_echo') | 710.962 | 239.838 |
128
+
129
+ ## Error Rates
130
+
131
+ | scenario_id | error_rate_pct |
132
+ |:------------------------------------------------------|-----------------:|
133
+ | fastmcp__http_api__fibonacci__vu50__clunlimited | 0.037 |
134
+ | gradio__mcp_streamable__echo__vu25__cl1__noq | 0.023 |
135
+ | gradio__mcp_streamable__payload_echo__vu25__cl1__noq | 0.028 |
136
+ | gradio__http_api__echo__vu25__cl5 | 0.05 |
137
+ | gradio__http_api__fibonacci__vu25__cl5 | 0.05 |
138
+ | gradio__http_api__payload_echo__vu50__cl5__noq | 0.05 |
139
+ | gradio__mcp_streamable__async_sleep__vu25__cl5__noq | 0.03 |
140
+ | gradio__http_api__echo__vu25__cl10 | 0.048 |
141
+ | gradio__http_api__fibonacci__vu25__cl10__noq | 0.056 |
142
+ | gradio__http_api__fibonacci__vu50__cl10 | 0.066 |
143
+ | gradio__http_api__payload_echo__vu25__cl10__noq | 0.056 |
144
+ | gradio__mcp_streamable__async_sleep__vu25__cl10__noq | 0.028 |
145
+ | gradio__http_api__async_sleep__vu25__clunlimited__noq | 0.047 |
146
+
147
+ ## Resource Usage
148
+
149
+ | server | ('mean', 'avg_cpu_pct') | ('mean', 'avg_memory_mb') | ('mean', 'peak_memory_mb') | ('max', 'avg_cpu_pct') | ('max', 'avg_memory_mb') | ('max', 'peak_memory_mb') |
150
+ |:---------|--------------------------:|----------------------------:|-----------------------------:|-------------------------:|---------------------------:|----------------------------:|
151
+ | fastmcp | 0 | 5.684 | 5.684 | 0 | 6.15 | 6.15 |
152
+ | gradio | 0 | 6.14953 | 6.14953 | 0 | 6.16 | 6.16 |
153
+
154
+ ## Benchmark Charts
155
+
156
+ ![Throughput Comparison](plots/throughput_comparison.png)
157
+
158
+ ![Latency Comparison](plots/latency_comparison.png)
159
+
160
+ ![Gradio Concurrency Limit Scaling](plots/gradio_cl_scaling.png)
161
+
162
+ ![Protocol Overhead](plots/protocol_overhead.png)
163
+
164
+ ## Methodology
165
+
166
+ - Both servers use identical tool implementations (imported from shared_tools.py)
167
+ - Each scenario runs in an isolated server subprocess
168
+ - Warmup period excluded from measurements
169
+ - Load generated by async httpx workers (not external tools)
170
+ - MCP tests use full protocol lifecycle (initialize → call_tool)
171
+ - System metrics sampled every 1s via psutil
172
+
173
+ ---
174
+
175
+ *Benchmarks generated by [mcp-server-bench](https://github.com/Mandark-droid/mcp-server-bench)*
detailed_results.json ADDED
The diff for this file is too large to render. See raw diff
 
plots/gradio_cl_scaling.png ADDED

Git LFS Details

  • SHA256: f3bdfe433cc3309f086b7d838e704c1e37f987857b3c86bbb252245666926bfa
  • Pointer size: 131 Bytes
  • Size of remote file: 404 kB
plots/latency_comparison.png ADDED

Git LFS Details

  • SHA256: ffd1fcf826c4831bafc78c0343659f8ddfa1c6905b2a4d7cfbe7ab334a8f307f
  • Pointer size: 130 Bytes
  • Size of remote file: 46.4 kB
plots/protocol_overhead.png ADDED

Git LFS Details

  • SHA256: 630be674d32bd71ea63bd8be3c1b14ec36489db26a7ab6874febd35c2fd8dfd8
  • Pointer size: 130 Bytes
  • Size of remote file: 48.9 kB
plots/throughput_comparison.png ADDED

Git LFS Details

  • SHA256: 8aa847ab1c99c8dfd309810d30e00cdc0e164eb3534364cacbbf786506b567be
  • Pointer size: 131 Bytes
  • Size of remote file: 210 kB