Test send and receive throughput for Mozilla's Servo ipc-channel between 2 processes
cargo benchIt appears I was missing to tell serde how to work with bytes, following fix makes performance sending with bincode serde comparable to sending raw bytes array. Thanks Dr-Ermann for pointing out:
#[derive(Serialize, Deserialize, Clone)]
struct Message {
pub topic: u32,
#[serde(with = "serde_bytes")]
pub data: Vec<u8>
}IpcSender / IpcReceiver
sends_custom/1024 time: [1.7104 us 1.7234 us 1.7407 us]
thrpt: [561.02 MiB/s 566.63 MiB/s 570.95 MiB/s]
sends_custom/10240 time: [4.4657 us 4.5620 us 4.6565 us]
thrpt: [2.0480 GiB/s 2.0905 GiB/s 2.1355 GiB/s]
sends_custom/51200 time: [43.689 us 44.632 us 45.778 us]
thrpt: [1.0416 GiB/s 1.0684 GiB/s 1.0914 GiB/s]
sends_custom/102400 time: [69.929 us 70.925 us 72.257 us]
thrpt: [1.3198 GiB/s 1.3446 GiB/s 1.3638 GiB/s]
receives_custom/1024 time: [1.7672 us 1.8164 us 1.8696 us]
thrpt: [522.35 MiB/s 537.63 MiB/s 552.60 MiB/s]
receives_custom/10240 time: [4.3900 us 4.4939 us 4.6213 us]
thrpt: [2.0636 GiB/s 2.1221 GiB/s 2.1724 GiB/s]
receives_custom/51200 time: [42.596 us 42.876 us 43.201 us]
thrpt: [1.1038 GiB/s 1.1121 GiB/s 1.1194 GiB/s]
receives_custom/102400 time: [70.224 us 70.906 us 71.863 us]
thrpt: [1.3271 GiB/s 1.3450 GiB/s 1.3580 GiB/s]
Test results with IpcBytesSender/IpcBytesReceiver
sends_custom_bytes/1024 time: [1.5317 us 1.5523 us 1.5793 us]
thrpt: [618.36 MiB/s 629.09 MiB/s 637.55 MiB/s]
sends_custom_bytes/10240
time: [3.6246 us 3.6457 us 3.6703 us]
thrpt: [2.5983 GiB/s 2.6159 GiB/s 2.6311 GiB/s]
sends_custom_bytes/51200
time: [40.257 us 40.851 us 41.614 us]
thrpt: [1.1459 GiB/s 1.1673 GiB/s 1.1845 GiB/s]
sends_custom_bytes/102400
time: [60.408 us 61.474 us 62.816 us]
thrpt: [1.5182 GiB/s 1.5514 GiB/s 1.5787 GiB/s]
receives_custom_bytes/1024
time: [1.4920 us 1.5064 us 1.5213 us]
thrpt: [641.93 MiB/s 648.26 MiB/s 654.51 MiB/s]
receives_custom_bytes/10240
time: [3.6613 us 3.7150 us 3.7777 us]
thrpt: [2.5245 GiB/s 2.5671 GiB/s 2.6047 GiB/s]
receives_custom_bytes/51200
time: [41.050 us 42.098 us 43.319 us]
thrpt: [1.1008 GiB/s 1.1327 GiB/s 1.1616 GiB/s]
receives_custom_bytes/102400
time: [57.150 us 57.577 us 58.095 us]
thrpt: [1.6416 GiB/s 1.6563 GiB/s 1.6687 GiB/s]
Benches compare sending of byte buffer of a given size either within a Serialize derived structure vs plain buffer.
It appears that sending and receiving bytes has about 5-10 times higher throughput than sending serializable object. Surprisingly sending 50KiB is more than 2 times slower than sending 10KiB and 100KiB - should be very platform specific. Sending/receiving bytes buf throughput has high variance. Sending/receiving serializable object has low variance.
Tests performed on MacBook Pro '13 (which sucks at perf tests), tests named by pattern type/message size.
Structure of a form, where size of Message::data variable between tests:
#[derive(Serialize, Deserialize)]
pub struct Message {
//pub topic: String,
pub topic: u32,
pub data: Vec<u8>
}Send Message.clone() via IpcSender (Serialized via bincode)
sends/1024 time: [7.8511 us 7.9644 us 8.0932 us]
thrpt: [120.66 MiB/s 122.62 MiB/s 124.39 MiB/s]
sends/10240 time: [61.313 us 61.641 us 61.968 us]
thrpt: [157.59 MiB/s 158.43 MiB/s 159.28 MiB/s]
sends/51200 time: [310.77 us 314.08 us 318.15 us]
thrpt: [153.47 MiB/s 155.46 MiB/s 157.12 MiB/s]
sends/102400 time: [600.69 us 605.17 us 610.21 us]
thrpt: [160.04 MiB/s 161.37 MiB/s 162.57 MiB/s]
Receive Message via IpcReceiver (Deserialized via bincode)
receives/1024 time: [7.6571 us 7.6989 us 7.7505 us]
thrpt: [126.00 MiB/s 126.84 MiB/s 127.54 MiB/s]
receives/10240 time: [60.936 us 61.342 us 61.843 us]
thrpt: [157.91 MiB/s 159.20 MiB/s 160.26 MiB/s]
receives/51200 time: [311.62 us 316.45 us 322.98 us]
thrpt: [151.18 MiB/s 154.30 MiB/s 156.69 MiB/s]
receives/102400 time: [661.13 us 668.65 us 676.87 us]
thrpt: [144.28 MiB/s 146.05 MiB/s 147.71 MiB/s]
type Message = Vec<u8>Sending bytes buf Vec<u8>.as_slice()
bytes_sends/1024 time: [1.4149 us 1.4572 us 1.5078 us]
thrpt: [647.69 MiB/s 670.16 MiB/s 690.22 MiB/s]
bytes_sends/10240 time: [4.1461 us 4.2852 us 4.4799 us]
thrpt: [2.1288 GiB/s 2.2255 GiB/s 2.3002 GiB/s]
bytes_sends/51200 time: [57.111 us 58.914 us 60.991 us]
thrpt: [800.58 MiB/s 828.80 MiB/s 854.97 MiB/s]
bytes_sends/102400 time: [59.126 us 60.230 us 61.648 us]
thrpt: [1.5470 GiB/s 1.5834 GiB/s 1.6129 GiB/s]
Sending cloned bytes buf Vec<u8>.clone().as_slice()
bytes_sends_cloned/1024 time: [1.6410 us 1.6559 us 1.6758 us]
thrpt: [582.73 MiB/s 589.75 MiB/s 595.09 MiB/s]
bytes_sends_cloned/10240
time: [4.0405 us 4.0725 us 4.1074 us]
thrpt: [2.3219 GiB/s 2.3417 GiB/s 2.3603 GiB/s]
bytes_sends_cloned/51200
time: [43.689 us 44.761 us 45.886 us]
thrpt: [1.0392 GiB/s 1.0653 GiB/s 1.0914 GiB/s]
bytes_sends_cloned/102400
time: [70.887 us 72.588 us 74.637 us]
thrpt: [1.2778 GiB/s 1.3138 GiB/s 1.3454 GiB/s]
Receiving bytes buf as Vec<u8>
bytes_receives/1024 time: [1.3999 us 1.4303 us 1.4700 us]
thrpt: [664.33 MiB/s 682.75 MiB/s 697.62 MiB/s]
bytes_receives/10240 time: [4.1569 us 4.2248 us 4.3060 us]
thrpt: [2.2147 GiB/s 2.2573 GiB/s 2.2942 GiB/s]
bytes_receives/51200 time: [38.904 us 42.204 us 45.431 us]
thrpt: [1.0496 GiB/s 1.1298 GiB/s 1.2257 GiB/s]
bytes_receives/102400 time: [60.481 us 62.754 us 65.247 us]
thrpt: [1.4616 GiB/s 1.5197 GiB/s 1.5768 GiB/s]
It does seem that main reason of slow send is Serialization/Deserialization, let's see bincode for
#[derive(Serialize, Deserialize)]
pub struct Message {
//pub topic: String,
pub topic: u32,
pub data: Vec<u8>
}Encode
bincode_encode/1024 time: [8.8087 us 9.0109 us 9.2546 us]
thrpt: [105.52 MiB/s 108.38 MiB/s 110.86 MiB/s]
bincode_encode/10240 time: [87.656 us 89.375 us 91.548 us]
thrpt: [106.67 MiB/s 109.27 MiB/s 111.41 MiB/s]
bincode_encode/51200 time: [433.69 us 439.43 us 447.46 us]
thrpt: [109.12 MiB/s 111.12 MiB/s 112.59 MiB/s]
bincode_encode/102400 time: [865.30 us 879.94 us 899.56 us]
thrpt: [108.56 MiB/s 110.98 MiB/s 112.86 MiB/s]
Decode
bincode_decode/1024 time: [2.9798 us 3.1397 us 3.3105 us]
thrpt: [294.99 MiB/s 311.03 MiB/s 327.73 MiB/s]
bincode_decode/10240 time: [29.083 us 31.390 us 34.096 us]
thrpt: [286.42 MiB/s 311.10 MiB/s 335.78 MiB/s]
bincode_decode/51200 time: [133.91 us 147.23 us 164.21 us]
thrpt: [297.34 MiB/s 331.64 MiB/s 364.64 MiB/s]
bincode_decode/102400 time: [333.02 us 359.75 us 393.40 us]
thrpt: [248.24 MiB/s 271.45 MiB/s 293.24 MiB/s]
ipc_channel_custom.rs takes it even further, using the same underlying data layer (just a Vec<u8>) with
channels instantiated either ipc::channel or ipc::bytes_channel, when sending bytes it uses custom tailored serialization.
As it can be seen custom tailored Serialization/Deserialization implementation working up to 10x times faster than bincode's.
sends_custom/1024 time: [7.1496 us 7.1883 us 7.2374 us]
thrpt: [134.93 MiB/s 135.85 MiB/s 136.59 MiB/s]
sends_custom/10240 time: [56.718 us 56.893 us 57.076 us]
thrpt: [171.10 MiB/s 171.65 MiB/s 172.18 MiB/s]
sends_custom/51200 time: [295.99 us 298.58 us 301.73 us]
thrpt: [161.83 MiB/s 163.53 MiB/s 164.97 MiB/s]
sends_custom/102400 time: [581.99 us 583.39 us 585.09 us]
thrpt: [166.91 MiB/s 167.39 MiB/s 167.80 MiB/s]
receives_custom/1024 time: [7.1560 us 7.2338 us 7.3347 us]
thrpt: [133.14 MiB/s 135.00 MiB/s 136.47 MiB/s]
receives_custom/10240 time: [56.625 us 56.965 us 57.326 us]
thrpt: [170.35 MiB/s 171.43 MiB/s 172.46 MiB/s]
receives_custom/51200 time: [296.42 us 296.94 us 297.62 us]
thrpt: [164.06 MiB/s 164.44 MiB/s 164.73 MiB/s]
receives_custom/102400 time: [588.19 us 590.89 us 593.70 us]
thrpt: [164.49 MiB/s 165.27 MiB/s 166.03 MiB/s]
sends_custom_bytes/1024 time: [1.5784 us 1.6231 us 1.6922 us]
thrpt: [577.10 MiB/s 601.66 MiB/s 618.70 MiB/s]
sends_custom_bytes/10240
time: [3.6300 us 3.6490 us 3.6735 us]
thrpt: [2.5961 GiB/s 2.6135 GiB/s 2.6272 GiB/s]
sends_custom_bytes/51200
time: [29.473 us 29.964 us 30.615 us]
thrpt: [1.5575 GiB/s 1.5914 GiB/s 1.6179 GiB/s]
sends_custom_bytes/102400
time: [51.452 us 52.001 us 52.597 us]
thrpt: [1.8132 GiB/s 1.8339 GiB/s 1.8535 GiB/s]
receives_custom_bytes/1024
time: [1.5848 us 1.6060 us 1.6289 us]
thrpt: [599.51 MiB/s 608.05 MiB/s 616.20 MiB/s]
receives_custom_bytes/10240
time: [3.6046 us 3.6162 us 3.6305 us]
thrpt: [2.6269 GiB/s 2.6372 GiB/s 2.6457 GiB/s]
receives_custom_bytes/51200
time: [30.061 us 31.198 us 32.522 us]
thrpt: [1.4662 GiB/s 1.5284 GiB/s 1.5862 GiB/s]
receives_custom_bytes/102400
time: [53.772 us 55.318 us 57.163 us]
thrpt: [1.6683 GiB/s 1.7240 GiB/s 1.7735 GiB/s]
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent