very experimental ledbat++ -ish cca

ledbat.hpp

@@ -0,0 +1,115 @@
+#pragma once
+
+#include <chrono>
+#include <deque>
+#include <vector>
+#include <cstdint>
+
+// LEDBAT: https://www.rfc-editor.org/rfc/rfc6817
+// LEDBAT++: https://www.ietf.org/archive/id/draft-irtf-iccrg-ledbat-plus-plus-01.txt
+
+// LEDBAT++ implementation
+struct LEDBAT {
+	public: // config
+		using SeqIDType = std::pair<uint8_t, uint16_t>; // tf_id, seq_id
+
+		static constexpr size_t IPV4_HEADER_SIZE {20};
+		static constexpr size_t IPV6_HEADER_SIZE {40}; // bru
+		static constexpr size_t UDP_HEADER_SIZE {8};
+
+		// TODO: tcp AND IPv6 will be different
+		static constexpr size_t segment_overhead {
+			4+ // ft overhead
+			46+ // tox?
+			UDP_HEADER_SIZE+
+			IPV4_HEADER_SIZE
+		};
+
+		static constexpr size_t maximum_segment_size {496 + segment_overhead}; // tox 500 - 4 from ft
+		static_assert(maximum_segment_size == 574); // mesured in wireshark
+
+		// ledbat++ says 60ms, we might need other values if relayed
+		const float target_delay {0.060f};
+		//const float target_delay {0.030f};
+		//const float target_delay {0.120f}; // 2x if relayed?
+
+		// TODO: use a factor for multiple of rtt
+		static constexpr size_t current_delay_filter_window {16*4*2};
+
+		//static constexpr size_t rtt_buffer_size_max {2000};
+
+		float max_byterate_allowed {10*1024*1024}; // 10MiB/s
+		//float max_byterate_allowed {2*1024*1024};
+
+	public:
+		LEDBAT(void);
+
+		// return the current believed window in bytes of how much data can be inflight,
+		// without overstepping the delay requirement
+		float getCWnD(void) const {
+			return _cwnd;
+		}
+
+		// TODO: api for how much data we should send
+		// take time since last sent into account
+		// respect max_byterate_allowed
+		size_t canSend(void) const;
+
+	public: // callbacks
+		// data size is without overhead
+		void onSent(SeqIDType seq, size_t data_size);
+
+		void onAck(std::vector<SeqIDType> seqs);
+
+		// if discard, not resent, not inflight
+		void onLoss(SeqIDType seq, bool discard);
+
+	private:
+		using clock = std::chrono::steady_clock;
+
+		// make values relative to algo start for readability (and precision)
+		// get timestamp in seconds
+		float getTimeNow(void) const {
+			return std::chrono::duration<float>{clock::now() - _time_start_offset}.count();
+		}
+
+		// moving avg over the last few delay samples
+		// VERY sensitive to bundling acks
+		float getCurrentDelay(void) const;
+
+		void addRTT(float new_delay);
+
+		void updateWindows(void);
+
+	private: // state
+		//float _cto {2.f}; // congestion timeout value in seconds
+
+		float _cwnd {2.f * maximum_segment_size}; // in bytes
+		float _base_delay {2.f}; // lowest mesured delay in _rtt_buffer in seconds
+
+		float _last_cwnd {0.f}; // timepoint of last cwnd correction
+		int64_t _recently_acked_data {0}; // reset on _last_cwnd
+		bool _recently_lost_data {false};
+		int64_t _recently_sent_bytes {0};
+
+		// initialize to low value, will get corrected very fast
+		float _fwnd {0.01f * max_byterate_allowed}; // in bytes
+
+
+		// ssthresh
+
+		// spec recomends 10min
+		// TODO: optimize and devide into spans of 1min (spec recom)
+		std::deque<float> _tmp_rtt_buffer;
+		std::deque<std::pair<float, float>> _rtt_buffer; // timepoint, delay
+		std::deque<float> _rtt_buffer_minutes;
+
+		// list of sequence ids and timestamps of when they where sent
+		std::deque<std::tuple<SeqIDType, float, size_t>> _in_flight;
+
+		int64_t _in_flight_bytes {0};
+
+	private: // helper
+		clock::time_point _time_start_offset;
+};
+