• Home
Name Date Size #Lines LOC

..--

Android.bpD03-May-20241.5 KiB6258

README.mdD03-May-20242.5 KiB7361

bpf_benchmark.cppD03-May-20242.9 KiB8552

connect_benchmark.cppD03-May-20248.3 KiB245149

dns_benchmark.cppD03-May-20243.7 KiB11164

main.cppD03-May-2024673 202

README.md

1# NetD benchmarks
2
3These are benchmarks for libc **connect** and **gethostbyname** functions as hooked by netd.
4
5## Infrastructure
6
7The benchmarks are powered by [google-benchmark](https://github.com/google/benchmark), which records
8four statistics:
9
10<table>
11  <tr>
12    <th>Statistic</th>
13    <th>Meaning</th>
14  </tr>
15  <tr>
16    <th>real_time</th>
17    <td>Mean wall-clock time taken for each iteration of the function under test to complete. This
18        includes any time spent in other processes.
19        <p>
20        Some of the **connect** benchmarks override this by implementing their own timings, but
21        the semantics are roughly the same. See connect_benchmark.cpp for details.</td>
22  </tr>
23  <tr>
24    <th>cpu_time</th>
25    <td>Mean time spent in the current thread during each iteration of the function under test. This
26        does not include any time spent in other processes or blocking on I/O.
27        <p>
28        Since almost all of the functionality we are testing here is in the netd server, not the
29        client, these timings are not relevant and should be ignored for most purposes.</td>
30  </tr>
31  <tr>
32    <th>iterations</th>
33    <td>The number of runs of the test function that were executed, across all threads, in order for
34        the test to reach the minimum amount of time run. Most tests here define 'minimum time run'
35        in terms of real time or 'manual time', rather than CPU time, for the reasons above.
36        <p>
37        This should be roughly inversely proportional to **real_time** for single-threaded
38        benchmarks and also somewhat proportional to the number of threads used.</td>
39  </tr>
40  <tr>
41    <th>label</th>
42    <td>Field that the benchmark can fill with an arbitrary string.
43        <p>
44        For graphing purposes, a few of the benchmarks in this directory store the 90th-percentile
45        latency in this field for comparison against the mean latency
46        (it's only possible to plot one other value apart from mean).
47        <p>
48        If the graphing ecosystem ever becomes advanced enough to plot multiple outputs from one
49        label, it might be more useful to log every decile to get a better picture of how spiky
50        things are.</td>
51</table>
52
53# Methods currently being benchmarked
54
55## connect()
56
57- Documented in [connect\_benchmark.cpp](connect_benchmark.cpp)
58
59## getaddrinfo()
60
61- Documented in [dns\_benchmark.cpp](dns_benchmark.cpp)
62
63
64<style type="text/css">
65  tr:nth-child(2n+1) {
66    background: lightgrey;
67  }
68  td,th {
69    text-align: left;
70    vertical-align: top;
71  }
72</style>
73