Skip to content

Commit a3be24c

Browse files
Docs: the performance page leads with ranked charts, with a data page and CSVs generated from the benchmark results (#164)
* Docs: the Performance page leads with generated SVG charts; the numbers live in CSV files and a data page The page now opens with a chart and each section is chart, explanation, then a link to its data: headline ratios, reader-open time, total time, memory, cost against columns, the one-time cost of a map, real databases (time and memory) and sync against async reads. The general text (what is measured, limits) moved to the end. Numbers are in docsite/perf-data/*.csv, one per chart set, each with a source file naming the benchmark, build, run date, machine, conditions and the BenchmarkDotNet result file of every number. The SVGs and the new data page (performance-data.md, one table per chart plus provenance) are generated from those CSVs by docsite/tools/perf/New-PerfCharts.ps1; Extract-PerfData.ps1 builds the CSVs from a run's own BenchmarkDotNet exports. The README beside them lists how to update the charts. DocFX publishes the CSVs (resource entry added). The data comes from the runs of 2026-11-14, before rc.6, one run per case. It is to be replaced by a fresh run on current master before this merges. Docs and docs tooling only: no Source or PublicAPI change. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * Docs: performance charts use solid family fills, and scroll sideways on a phone Bars are solid in the family colors of the benchmark ranking page (its light and dark sets), rounded at the data end, with no outline. Each chart sits in a container that scrolls horizontally, with the SVG at a minimum width of 560 px (its own width), so its text stays at 11.5 px or more on a narrow screen. The SVGs still draw their own background card. Regenerated with the committed scripts; the CSV files are unchanged. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
1 parent 6618ab5 commit a3be24c

27 files changed

Lines changed: 2642 additions & 37 deletions
Lines changed: 204 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,204 @@
1+
<!-- Generated by docsite/tools/perf/New-PerfCharts.ps1 from docsite/perf-data/*.csv. Do not edit by hand. -->
2+
# Performance data
3+
4+
The numbers behind every chart on [Performance](performance.md), as tables. Each section gives the same numbers as the CSV file next to it, which you can download, and a source file that names the benchmark, the build, the date, the machine and the conditions, and the result file each number came from. The tables and the charts are generated from the CSV files by one script, so nothing here is typed by hand.
5+
6+
## How fresh the data is
7+
8+
Each data set comes from one BenchmarkDotNet run (or a few runs of the same build). The build is the commit of the benchmark project that was run; the date is when the run finished.
9+
10+
| Data | Charts | Build | Run date | Machine | Runtime |
11+
|---|---|---|---|---|---|
12+
| `synthetic-overhead.csv` | headline ratios, reader-open time, total time, memory | `b27a02f` | 2026-11-14 | AMD Ryzen 9 7900 3.70GHz, 1 CPU, 24 logical and 12 physical cores | .NET 11.0.0 (11.0.0, 11.0.26.53018), X64 RyuJIT x86-64-v4 |
13+
| `width-sweep.csv` | cost grows with columns | `b27a02f` | 2026-11-14 | AMD Ryzen 9 7900 3.70GHz, 1 CPU, 24 logical and 12 physical cores | .NET 11.0.0 (11.0.0, 11.0.26.53018), X64 RyuJIT x86-64-v4 |
14+
| `map-cost.csv` | what building a map costs | `544cc01` | 2026-11-14 | AMD Ryzen 9 7900 3.70GHz, 1 CPU, 24 logical and 12 physical cores | .NET 11.0.0 (11.0.0, 11.0.26.53018), X64 RyuJIT x86-64-v4 |
15+
| `real-providers.csv` | real databases | `fbb13d7` | 2026-11-14 | AMD Ryzen 9 7900 3.70GHz, 1 CPU, 24 logical and 12 physical cores | .NET 11.0.0 (11.0.0, 11.0.26.53018), X64 RyuJIT x86-64-v4 |
16+
| `sync-async.csv` | sync against async reads | `ea5f096` | 2026-11-14 | AMD Ryzen 9 7900 3.70GHz, 1 CPU, 24 logical and 12 physical cores | .NET 11.0.0 (11.0.0, 11.0.26.53018), X64 RyuJIT x86-64-v4 |
17+
18+
<a id="headline"></a>
19+
## Headline ratios
20+
21+
[Back to the chart](performance.md#headline). Each ratio divides two numbers of the table [Reader-open time](#open-time) or [Memory allocated](#memory).
22+
23+
| Ratio | Value | Computed from |
24+
|---|---:|---|
25+
| Dapper reader-open time over the RowMap plan | 2.48× | 202.4 µs / 81.6 µs |
26+
| Dapper allocation over the RowMap plan | 2.41× | 548.1 KB / 227.4 KB |
27+
| object[] rows reader-open time over the RowMap plan | 2.19× | 178.4 µs / 81.6 µs |
28+
| RowMap plan reader-open time over the hand-written selector | 1.74× | 81.6 µs / 46.8 µs |
29+
30+
Data: [synthetic-overhead.csv](../perf-data/synthetic-overhead.csv) ([source and conditions](../perf-data/synthetic-overhead.source.txt)).
31+
32+
<a id="open-time"></a>
33+
## Reader-open time
34+
35+
[Back to the chart](performance.md#open-time). Mean time from ExecuteReader until the reader closes, 20 columns, 1,000 rows, synthetic reader. Build b27a02f, 2026-11-14.
36+
37+
| Rank | Variant | Family | Reader open (µs) | Against the best |
38+
|---:|---|---|---:|---:|
39+
| 1 | Map(selector), hand-written | ODX: one object per row | 46.8 | best |
40+
| 2 | Selector, reused row, Select | ODX: one object per row | 58.8 | +25.6% |
41+
| 3 | Selector, pooled row, Select | ODX: one object per row | 62.8 | +34.2% |
42+
| 4 | Shaping inside the selector | ODX: one object per row | 67.6 | +44.4% |
43+
| 5 | Map&lt;T&gt;(), the RowMap plan | ODX: one object per row | 81.6 | +74.3% |
44+
| 6 | Row class, close, then Select | ODX: row class, then Select | 82.1 | +75.5% |
45+
| 7 | Row class, shaping after close | ODX: row class, then Select | 82.3 | +75.7% |
46+
| 8 | Row class, inline Select | ODX: row class, then Select | 106.4 | +127.3% |
47+
| 9 | Row class, shaping in Select | ODX: row class, then Select | 138.9 | +196.7% |
48+
| 10 | object[] rows, objects later | ODX: raw object[] rows | 170.1 | +263.4% |
49+
| 11 | object[] rows (baseline) | ODX: raw object[] rows | 178.4 | +281.0% |
50+
| 12 | Dapper, then Select after | Dapper 2.1.89 | 197.5 | +321.8% |
51+
| 13 | Dapper, shaping after | Dapper 2.1.89 | 200.5 | +328.2% |
52+
| 14 | Dapper Query&lt;T&gt; | Dapper 2.1.89 | 202.4 | +332.3% |
53+
| 15 | Dapper, inline Select | Dapper 2.1.89 | 214.0 | +357.0% |
54+
| 16 | Dapper, shaping in Select | Dapper 2.1.89 | 257.6 | +450.2% |
55+
56+
Data: [synthetic-overhead.csv](../perf-data/synthetic-overhead.csv) ([source and conditions](../perf-data/synthetic-overhead.source.txt)).
57+
58+
<a id="total-time"></a>
59+
## Total time per call
60+
61+
[Back to the chart](performance.md#total-time). Mean time of the whole call, same runs as above. Build b27a02f, 2026-11-14.
62+
63+
| Rank | Variant | Family | Total (µs) | Against the best |
64+
|---:|---|---|---:|---:|
65+
| 1 | Map(selector), hand-written | ODX: one object per row | 47.0 | best |
66+
| 2 | Selector, reused row, Select | ODX: one object per row | 59.0 | +25.7% |
67+
| 3 | Selector, pooled row, Select | ODX: one object per row | 63.1 | +34.3% |
68+
| 4 | Shaping inside the selector | ODX: one object per row | 67.8 | +44.3% |
69+
| 5 | Map&lt;T&gt;(), the RowMap plan | ODX: one object per row | 81.7 | +74.1% |
70+
| 6 | Row class, close, then Select | ODX: row class, then Select | 97.8 | +108.3% |
71+
| 7 | Row class, inline Select | ODX: row class, then Select | 106.7 | +127.1% |
72+
| 8 | Row class, shaping after close | ODX: row class, then Select | 134.5 | +186.4% |
73+
| 9 | Row class, shaping in Select | ODX: row class, then Select | 139.2 | +196.4% |
74+
| 10 | object[] rows (baseline) | ODX: raw object[] rows | 178.5 | +280.1% |
75+
| 11 | Dapper Query&lt;T&gt; | Dapper 2.1.89 | 202.7 | +331.6% |
76+
| 12 | Dapper, inline Select | Dapper 2.1.89 | 214.3 | +356.3% |
77+
| 13 | Dapper, then Select after | Dapper 2.1.89 | 217.1 | +362.4% |
78+
| 14 | object[] rows, objects later | ODX: raw object[] rows | 221.0 | +370.7% |
79+
| 15 | Dapper, shaping after | Dapper 2.1.89 | 256.2 | +445.6% |
80+
| 16 | Dapper, shaping in Select | Dapper 2.1.89 | 257.9 | +449.2% |
81+
82+
Data: [synthetic-overhead.csv](../perf-data/synthetic-overhead.csv) ([source and conditions](../perf-data/synthetic-overhead.source.txt)).
83+
84+
<a id="memory"></a>
85+
## Memory allocated per call
86+
87+
[Back to the chart](performance.md#memory). Bytes allocated per call divided by 1,024, same runs as above. Build b27a02f, 2026-11-14.
88+
89+
| Rank | Variant | Family | Allocated (KB) | Against the best |
90+
|---:|---|---|---:|---:|
91+
| 1 | Map&lt;T&gt;(), the RowMap plan | ODX: one object per row | 227.4 | best |
92+
| 2 | Map(selector), hand-written | ODX: one object per row | 227.4 | +0.0% |
93+
| 3 | Selector, reused row, Select | ODX: one object per row | 228.1 | +0.3% |
94+
| 4 | Selector, pooled row, Select | ODX: one object per row | 228.2 | +0.4% |
95+
| 5 | Shaping inside the selector | ODX: one object per row | 297.7 | +30.9% |
96+
| 6 | Row class, inline Select | ODX: row class, then Select | 438.7 | +92.9% |
97+
| 7 | Row class, close, then Select | ODX: row class, then Select | 446.3 | +96.3% |
98+
| 8 | Row class, shaping in Select | ODX: row class, then Select | 509.1 | +123.9% |
99+
| 9 | object[] rows (baseline) | ODX: raw object[] rows | 516.6 | +127.2% |
100+
| 10 | Row class, shaping after close | ODX: row class, then Select | 516.7 | +127.2% |
101+
| 11 | Dapper Query&lt;T&gt; | Dapper 2.1.89 | 548.1 | +141.0% |
102+
| 12 | object[] rows, objects later | ODX: raw object[] rows | 735.4 | +223.4% |
103+
| 13 | Dapper, inline Select | Dapper 2.1.89 | 759.0 | +233.8% |
104+
| 14 | Dapper, then Select after | Dapper 2.1.89 | 767.0 | +237.3% |
105+
| 15 | Dapper, shaping in Select | Dapper 2.1.89 | 829.3 | +264.7% |
106+
| 16 | Dapper, shaping after | Dapper 2.1.89 | 837.4 | +268.2% |
107+
108+
Data: [synthetic-overhead.csv](../perf-data/synthetic-overhead.csv) ([source and conditions](../perf-data/synthetic-overhead.source.txt)).
109+
110+
<a id="width"></a>
111+
## Cost grows with the number of columns
112+
113+
[Back to the chart](performance.md#width). Total time per call (µs) at 1,000 rows. The last column is the cost of one more cell: the time at the widest result minus the time at the narrowest, divided by the cells added. Build b27a02f, 2026-11-14.
114+
115+
| Approach | 1 columns | 3 columns | 5 columns | 10 columns | 20 columns | Per added cell (ns) |
116+
|---|---:|---:|---:|---:|---:|---:|
117+
| Map(selector), hand-written | 10.7 | 11.9 | 16.8 | 31.2 | 47.0 | 1.91 |
118+
| Map&lt;T&gt;(), the RowMap plan | 12.4 | 18.3 | 25.2 | 49.1 | 81.7 | 3.65 |
119+
| Row class, inline Select | 28.3 | 32.4 | 41.4 | 67.2 | 106.7 | 4.12 |
120+
| object[] rows (baseline) | 9.5 | 20.0 | 35.9 | 75.7 | 178.5 | 8.90 |
121+
| Dapper Query&lt;T&gt; | 18.5 | 31.1 | 44.6 | 96.2 | 202.7 | 9.70 |
122+
123+
Data: [width-sweep.csv](../perf-data/width-sweep.csv) ([source and conditions](../perf-data/width-sweep.source.txt)).
124+
125+
<a id="map-cost"></a>
126+
## What building a map costs
127+
128+
[Back to the chart](performance.md#map-cost). One row per call. The one-time cost is the time (and memory) of a call that builds and compiles a new map, minus a call that reuses the cached one. Build 544cc01, 2026-11-14.
129+
130+
| Columns | Reused map (µs) | Map built every call (ms) | One-time cost (ms) | Reused map (KB) | Built every call (KB) | One-time memory (KB) |
131+
|---:|---:|---:|---:|---:|---:|---:|
132+
| 1 | 0.17 | 0.19 | 0.19 | 0.55 | 15.6 | 15.1 |
133+
| 5 | 0.19 | 0.33 | 0.33 | 0.55 | 22.9 | 22.4 |
134+
| 20 | 0.27 | 0.88 | 0.87 | 0.55 | 52.0 | 51.5 |
135+
136+
Data: [map-cost.csv](../perf-data/map-cost.csv) ([source and conditions](../perf-data/map-cost.source.txt)).
137+
138+
<a id="real-time"></a>
139+
## Real databases: time per call
140+
141+
[Back to the chart](performance.md#real-time). 20 columns, 10,000 rows. The CSV also holds the 5-column runs of the same benchmark. Build fbb13d7, 2026-11-14.
142+
143+
### SQL Server LocalDB
144+
145+
| Rank | Variant | Family | Time (ms) | Against the best |
146+
|---:|---|---|---:|---:|
147+
| 1 | Map(selector), hand-written | ODX: one object per row | 25.18 | best |
148+
| 2 | Hand loop (reference) | Hand loop on the provider reader (reference) | 25.59 | +1.7% |
149+
| 3 | Map&lt;T&gt;(), the RowMap plan | ODX: one object per row | 25.90 | +2.9% |
150+
| 4 | Row class, inline Select | ODX: row class, then Select | 27.24 | +8.2% |
151+
| 5 | Dapper Query&lt;T&gt; | Dapper 2.1.89 | 27.57 | +9.5% |
152+
153+
### SQLite file
154+
155+
| Rank | Variant | Family | Time (ms) | Against the best |
156+
|---:|---|---|---:|---:|
157+
| 1 | Map(selector), hand-written | ODX: one object per row | 25.90 | best |
158+
| 2 | Hand loop (reference) | Hand loop on the provider reader (reference) | 25.92 | +0.1% |
159+
| 3 | Dapper Query&lt;T&gt; | Dapper 2.1.89 | 29.78 | +15.0% |
160+
| 4 | Row class, inline Select | ODX: row class, then Select | 30.90 | +19.3% |
161+
| 5 | Map&lt;T&gt;(), the RowMap plan | ODX: one object per row | 31.30 | +20.9% |
162+
163+
Data: [real-providers.csv](../perf-data/real-providers.csv) ([source and conditions](../perf-data/real-providers.source.txt)).
164+
165+
<a id="real-memory"></a>
166+
## Real databases: memory per call
167+
168+
[Back to the chart](performance.md#real-memory). 20 columns, 10,000 rows. The CSV also holds the 5-column runs of the same benchmark. Build fbb13d7, 2026-11-14.
169+
170+
### SQL Server LocalDB
171+
172+
| Rank | Variant | Family | Allocated (KB) | Against the best |
173+
|---:|---|---|---:|---:|
174+
| 1 | Hand loop (reference) | Hand loop on the provider reader (reference) | 9,708.2 | best |
175+
| 2 | Map&lt;T&gt;(), the RowMap plan | ODX: one object per row | 9,708.4 | +0.0% |
176+
| 3 | Map(selector), hand-written | ODX: one object per row | 9,708.4 | +0.0% |
177+
| 4 | Row class, inline Select | ODX: row class, then Select | 11,818.1 | +21.7% |
178+
| 5 | Dapper Query&lt;T&gt; | Dapper 2.1.89 | 12,561.5 | +29.4% |
179+
180+
### SQLite file
181+
182+
| Rank | Variant | Family | Allocated (KB) | Against the best |
183+
|---:|---|---|---:|---:|
184+
| 1 | Hand loop (reference) | Hand loop on the provider reader (reference) | 12,250.3 | best |
185+
| 2 | Map(selector), hand-written | ODX: one object per row | 12,250.7 | +0.0% |
186+
| 3 | Map&lt;T&gt;(), the RowMap plan | ODX: one object per row | 14,448.5 | +17.9% |
187+
| 4 | Dapper Query&lt;T&gt; | Dapper 2.1.89 | 15,144.2 | +23.6% |
188+
| 5 | Row class, inline Select | ODX: row class, then Select | 16,558.3 | +35.2% |
189+
190+
Data: [real-providers.csv](../perf-data/real-providers.csv) ([source and conditions](../perf-data/real-providers.source.txt)).
191+
192+
<a id="sync-async"></a>
193+
## Sync reads against async reads
194+
195+
[Back to the chart](performance.md#sync-async). Reader-open time (µs), RowMap plan, 20 columns, 1,000 rows. Buffer refills: every 100 rows the reader waits 20 µs. Busy pool: 8 work items kept queued on the thread pool. Build ea5f096, 2026-11-14.
196+
197+
| Condition | Sync | Async execute | Async read | Async read against sync |
198+
|---|---:|---:|---:|---:|
199+
| Neither | 77.3 µs | 80.9 µs | 82.3 µs | 1.07× |
200+
| Busy thread pool | 94.4 µs | 97.4 µs | 102.4 µs | 1.08× |
201+
| Buffer refills | 257.9 µs | 261.2 µs | 314.3 µs | 1.22× |
202+
| Refills and busy pool | 279.3 µs | 279.0 µs | 11.2 ms | 40.01× |
203+
204+
Data: [sync-async.csv](../perf-data/sync-async.csv) ([source and conditions](../perf-data/sync-async.source.txt)).

0 commit comments

Comments
 (0)