aquatic_ws: deserialize InMessage with simd-json for performance

From 18% to 13% CPU time during load test run.

Criterion benchmark of deserialization got about 90% improvement.

Cargo.lock

@@ -65,7 +65,7 @@ dependencies = [
  "aquatic_http_protocol",
  "crossbeam-channel",
  "either",
- "hashbrown",
+ "hashbrown 0.8.1",
  "indexmap",
  "itoa",
  "log",
@@ -90,7 +90,7 @@ dependencies = [
  "anyhow",
  "aquatic_cli_helpers",
  "aquatic_http_protocol",
- "hashbrown",
+ "hashbrown 0.8.1",
  "mimalloc",
  "mio",
  "quickcheck",
@@ -107,7 +107,7 @@ dependencies = [
  "anyhow",
  "bendy",
  "criterion",
- "hashbrown",
+ "hashbrown 0.8.1",
  "hex",
  "httparse",
  "itoa",
@@ -130,7 +130,7 @@ dependencies = [
  "aquatic_common",
  "aquatic_udp_protocol",
  "crossbeam-channel",
- "hashbrown",
+ "hashbrown 0.8.1",
  "histogram",
  "indexmap",
  "mimalloc",
@@ -168,7 +168,7 @@ dependencies = [
  "aquatic_cli_helpers",
  "aquatic_udp_protocol",
  "crossbeam-channel",
- "hashbrown",
+ "hashbrown 0.8.1",
  "mimalloc",
  "mio",
  "parking_lot",
@@ -199,7 +199,7 @@ dependencies = [
  "aquatic_ws_protocol",
  "crossbeam-channel",
  "either",
- "hashbrown",
+ "hashbrown 0.8.1",
  "indexmap",
  "log",
  "mimalloc",
@@ -222,7 +222,7 @@ dependencies = [
  "anyhow",
  "aquatic_cli_helpers",
  "aquatic_ws_protocol",
- "hashbrown",
+ "hashbrown 0.8.1",
  "mimalloc",
  "mio",
  "quickcheck",
@@ -241,11 +241,12 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "criterion",
- "hashbrown",
+ "hashbrown 0.8.1",
  "quickcheck",
  "quickcheck_macros",
  "serde",
  "serde_json",
+ "simd-json",
  "tungstenite",
 ]
 
@@ -681,6 +682,15 @@ version = "0.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e88a8acf291dafb59c2d96e8f59828f3838bb1a70398823ade51a84de6a6deed"
 
+[[package]]
+name = "float-cmp"
+version = "0.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "499a1bfa03d254b047e7e5c1fc8dd23a8cf6b344a8eb7e622ae4bc76bfac8e68"
+dependencies = [
+ "num-traits",
+]
+
 [[package]]
 name = "fnv"
 version = "1.0.7"
@@ -754,6 +764,26 @@ version = "1.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d36fab90f82edc3c747f9d438e06cf0a491055896f2a279638bb5beed6c40177"
 
+[[package]]
+name = "halfbrown"
+version = "0.1.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c12499524b5585419ab2f51545a19b842263a373580a83c0eb98a0142a260a10"
+dependencies = [
+ "hashbrown 0.7.2",
+ "serde",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.7.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "96282e96bfcd3da0d3aa9938bedf1e50df3269b6db08b4876d2da0bb1a0841cf"
+dependencies = [
+ "ahash",
+ "autocfg",
+]
+
 [[package]]
 name = "hashbrown"
 version = "0.8.1"
@@ -827,7 +857,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5b88cd59ee5f71fea89a62248fc8f387d44400cefe05ef548466d61ced9029a7"
 dependencies = [
  "autocfg",
- "hashbrown",
+ "hashbrown 0.8.1",
 ]
 
 [[package]]
@@ -1602,6 +1632,18 @@ dependencies = [
  "opaque-debug",
 ]
 
+[[package]]
+name = "simd-json"
+version = "0.3.16"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0696059565e9aac60099f9ab388869551643fc2d04bf1535554e6ee157821a2f"
+dependencies = [
+ "halfbrown",
+ "serde",
+ "serde_json",
+ "value-trait",
+]
+
 [[package]]
 name = "simplelog"
 version = "0.8.0"
@@ -1849,6 +1891,18 @@ version = "0.7.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "05e42f7c18b8f902290b009cde6d651262f956c98bc51bca4cd1d511c9cd85c7"
 
+[[package]]
+name = "value-trait"
+version = "0.1.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3384a4788567e35113300281d737ab1b77917bd35cb99ffd8a4283345da9a825"
+dependencies = [
+ "float-cmp",
+ "halfbrown",
+ "itoa",
+ "ryu",
+]
+
 [[package]]
 name = "vcpkg"
 version = "0.2.10"

TODO.md

@@ -2,6 +2,7 @@
 
 ## General
 
+* rustflags for scripts in own .env file or similar
 * automatic tests running real clients in container?
 
 ## aquatic_http_load_test
@@ -18,8 +19,6 @@
   can distribute them to different workers)
 
 ## aquatic_http
-* array buffer for EstablishedConnection.send_response, there is a lot of
-  allocating and deallocating now
 * test torrent transfer with real clients
   * test tls
   * scrape: does it work (serialization etc), and with multiple hashes?
@@ -27,6 +26,7 @@
     positive number.
 * compact=0 should result in error response
 * config: multiple request workers
+* actually delete old benchmark
 
 ## aquatic_ws_load_test
 * still maybe too few answers received with aquatic_ws
@@ -38,8 +38,6 @@
 
 ## aquatic_ws
 * config: multiple request workers
-* create criterion benchmarks, then try out simd_json. deserializing InMessages
-  takes about 18% CPU, serializing OutMessages takes about 13.5% CPU
 * test transfer again with changes made:
   * crossbeam-channel
   * ipv6/ipv4 mapping
@@ -70,6 +68,8 @@
   if that is the since since it means a panic occured
 
 ## aquatic_http
+* array buffer for EstablishedConnection.send_response? there is a lot of
+  allocating and deallocating now. Doesn't seem to help performance a lot.
 * request parsing:
   * smartstring: maybe use for keys? maybe use less? needs benchmarking
 * use fastrand instead of rand? (also for ws and udp then I guess because of

aquatic_ws/src/lib/network/mod.rs

@@ -195,8 +195,8 @@ pub fn run_handshakes_and_read_messages(
             use ::tungstenite::Error::Io;
 
             match established_ws.ws.read_message(){
-                Ok(ws_message) => {
+                Ok(mut ws_message) => {
-                    if let Ok(in_message) = InMessage::from_ws_message(&ws_message){
+                    if let Ok(in_message) = InMessage::from_ws_message(&mut ws_message){
                         let naive_peer_addr = established_ws.peer_addr;
                         let converted_peer_ip = convert_ipv4_mapped_ipv6(
                             naive_peer_addr.ip()

aquatic_ws_protocol/Cargo.toml

@@ -18,6 +18,7 @@ anyhow = "1"
 hashbrown = { version = "0.8", features = ["serde"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+simd-json = "0.3"
 tungstenite = "0.11"
 
 [dev-dependencies]

aquatic_ws_protocol/benches/bench_deserialize_announce_request.rs

@@ -37,10 +37,10 @@ pub fn bench(c: &mut Criterion) {
         offer_id: Some(OfferId(info_hash.0))
     });
 
-    let ws_message = request.to_ws_message();
+    let mut ws_message = request.to_ws_message();
 
     c.bench_function("deserialize-announce-request", |b| b.iter(||
-        InMessage::from_ws_message(black_box(&ws_message))
+        InMessage::from_ws_message(black_box(&mut ws_message))
     ));
 }
 

aquatic_ws_protocol/src/lib.rs

@@ -258,16 +258,18 @@ pub enum InMessage {
 
 impl InMessage {
     #[inline]
-    pub fn from_ws_message(ws_message: &tungstenite::Message) -> ::anyhow::Result<Self> {
+    pub fn from_ws_message(
-        use tungstenite::Message::{Text, Binary};
+        ws_message: &mut tungstenite::Message
+    ) -> ::anyhow::Result<Self> {
+        use tungstenite::Message::Text;
 
-        let text = match ws_message {
+        let text: &mut str = if let Text(text) = ws_message {
-            Text(text) => text,
+            text
-            Binary(bytes) => ::std::str::from_utf8(bytes)?,
+        } else {
-            _ => return Err(anyhow::anyhow!("Message is neither text nor bytes")),
+            return Err(anyhow::anyhow!("Message is not text"));
         };
 
-        ::serde_json::from_str(text).context("serialize with serde")
+        ::simd_json::serde::from_str(text).context("deserialize with serde")
     }
 
     pub fn to_ws_message(&self) -> ::tungstenite::Message {
@@ -521,9 +523,9 @@ mod tests {
 
     #[quickcheck]
     fn quickcheck_serde_identity_in_message(in_message_1: InMessage) -> bool {
-        let ws_message = in_message_1.to_ws_message();
+        let mut ws_message = in_message_1.to_ws_message();
 
-        let in_message_2 = InMessage::from_ws_message(&ws_message).unwrap();
+        let in_message_2 = InMessage::from_ws_message(&mut ws_message).unwrap();
 
         let success = in_message_1 ==  in_message_2;
 
@@ -660,4 +662,4 @@ mod tests {
 
         success
     }
-}
+}

aquatic_ws_protocol/target/criterion/deserialize-announce-request/latest/estimates.json

@@ -1 +1 @@
-{"mean":{"confidence_interval":{"confidence_level":0.95,"lower_bound":18703.904042751496,"upper_bound":18792.03205243071},"point_estimate":18746.072800173675,"standard_error":22.574908262590952},"median":{"confidence_interval":{"confidence_level":0.95,"lower_bound":18547.912025316455,"upper_bound":18605.846380105744},"point_estimate":18576.281397429004,"standard_error":15.686062550253332},"median_abs_dev":{"confidence_interval":{"confidence_level":0.95,"lower_bound":306.28255537969375,"upper_bound":380.480602345445},"point_estimate":338.0236020241898,"standard_error":19.790518378924638},"slope":{"confidence_interval":{"confidence_level":0.95,"lower_bound":18658.020869749234,"upper_bound":18761.405070764362},"point_estimate":18706.576690779686,"standard_error":26.604673332450176},"std_dev":{"confidence_interval":{"confidence_level":0.95,"lower_bound":553.3799052716993,"upper_bound":868.3032654487926},"point_estimate":715.2362082741504,"standard_error":80.39906836878366}}
+{"mean":{"confidence_interval":{"confidence_level":0.95,"lower_bound":2094.926862805663,"upper_bound":2106.4267028892173},"point_estimate":2100.4015350320765,"standard_error":2.9392946950512737},"median":{"confidence_interval":{"confidence_level":0.95,"lower_bound":2085.6324189161965,"upper_bound":2087.229197620325},"point_estimate":2086.413383946413,"standard_error":0.4169721177244358},"median_abs_dev":{"confidence_interval":{"confidence_level":0.95,"lower_bound":16.490233538836364,"upper_bound":21.81096312849954},"point_estimate":18.680703664212437,"standard_error":1.3102508160911694},"slope":{"confidence_interval":{"confidence_level":0.95,"lower_bound":2085.6033212698335,"upper_bound":2094.8775402138917},"point_estimate":2089.9275901786054,"standard_error":2.361080413460781},"std_dev":{"confidence_interval":{"confidence_level":0.95,"lower_bound":73.7250326777399,"upper_bound":111.45831108639615},"point_estimate":93.04040248967267,"standard_error":9.710703224470429}}

aquatic_ws_protocol/target/criterion/deserialize-announce-request/latest/tukey.json

@@ -1 +1 @@
-[16630.874818319186,17501.98981446029,19824.96313750323,20696.078133644332]
+[1994.3859525202693,2034.7983478769509,2142.5647354947687,2182.97713085145]

scripts/test.sh

@@ -1,7 +1,17 @@
 #!/bin/sh
 
+# Test in release mode to avoid quickcheck tests taking forever
+
+# Compile with target-cpu=native but without AVX512 features, since they
+# decrease performance.
+
+DISABLE_AVX512=$(rustc --print target-features | grep "    avx512" |
+    awk '{print $1}'  | sed 's/^/-C target-feature=-/' | xargs)
+
+export RUSTFLAGS="-C target-cpu=native $DISABLE_AVX512"
+
 # Not chosen for exact values, only to be larger than defaults
 export QUICKCHECK_TESTS=2000
 export QUICKCHECK_GENERATOR_SIZE=1000
 
-cargo test
+cargo test --all