pushkarnk
commented
5 years ago Investigation revealed unmasking to be a possible bottleneck, specifically the following function from NIOWebSocket:
mutating func webSocketMask(_ maskingKey: WebSocketMaskingKey, indexOffset: Int = 0) {
self.withUnsafeMutableReadableBytes {
for (index, byte) in $0.enumerated() {
$0[index] = byte ^ maskingKey[(index + indexOffset) % 4]
}
}
}Flamegraphs show UnsafeMutableRawBufferPointer.enumerated() running for ~75% of the time. Rewriting this function with a simpler loop gives a good speedup:
mutating func webSocketMask(_ maskingKey: WebSocketMaskingKey, indexOffset: Int = 0) {
self.withUnsafeMutableReadableBytes {
var index = 0
for byte in $0 {
$0[index] = byte ^ maskingKey[(index + indexOffset) % 4]
index += 1
}
}
}This seems to be directly associated with the performance of Sequence.enumerated().
For example, on Ubuntu 14.04 image running on my machine, with 1M elements in array, this loop finishes in ~0.6s
var index = 0
for a in array {
_ = a
index += 1
}.. while this loop takes ~1.7s on the same array
for (index, a) in array.enumerated() {
_ = a
_ = index
}
Autobahn has 54 performance tests.
websocket-nioperforms better thanmasteronly on 16/54 tests. On the rest, the performance gap is also low as 4% and as high as 200%.