1 #ifndef _QUEUE_LENGTH_POLICY_H
2 #define _QUEUE_LENGTH_POLICY_H 1
15 // A class to estimate the future jitter. Used in QueueLengthPolicy (see below).
17 // There are many ways to estimate jitter; I've tested a few ones (and also
18 // some algorithms that don't explicitly model jitter) with different
19 // parameters on some real-life data in experiments/queue_drop_policy.cpp.
20 // This is one based on simple order statistics where I've added some margin in
21 // the number of starvation events; I believe that about one every hour would
22 // probably be acceptable, but this one typically goes lower than that, at the
23 // cost of 2–3 ms extra latency. (If the queue is hard-limited to one frame, it's
24 // possible to get ~10 ms further down, but this would mean framedrops every
25 // second or so.) The general strategy is: Take the 99.9-percentile jitter over
26 // last 5000 frames, multiply by two, and that's our worst-case jitter
27 // estimate. The fact that we're not using the max value means that we could
28 // actually even throw away very late frames immediately, which means we only
29 // get one user-visible event instead of seeing something both when the frame
30 // arrives late (duplicate frame) and then again when we drop.
33 static constexpr size_t history_length = 5000;
34 static constexpr double percentile = 0.999;
35 static constexpr double multiplier = 2.0;
38 void register_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
39 void unregister_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
44 expected_timestamp = std::chrono::steady_clock::time_point::min();
46 void frame_arrived(std::chrono::steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames);
47 std::chrono::steady_clock::time_point get_expected_next_frame() const { return expected_timestamp; }
48 double estimate_max_jitter() const;
51 // A simple O(k) based algorithm for getting the k-th largest or
52 // smallest element from our window; we simply keep the multiset
53 // ordered (insertions and deletions are O(n) as always) and then
54 // iterate from one of the sides. If we had larger values of k,
55 // we could go for a more complicated setup with two sets or heaps
56 // (one increasing and one decreasing) that we keep balanced around
57 // the point, or it is possible to reimplement std::set with
58 // counts in each node. However, since k=5, we don't need this.
59 std::multiset<double> orders;
60 std::deque<std::multiset<double>::iterator> history;
62 std::chrono::steady_clock::time_point expected_timestamp = std::chrono::steady_clock::time_point::min();
63 int64_t last_duration = 0;
65 // Metrics. There are no direct summaries for jitter, since we already have latency summaries.
66 std::atomic<int64_t> metric_input_underestimated_jitter_frames{0};
67 std::atomic<double> metric_input_estimated_max_jitter_seconds{0.0 / 0.0};
70 // For any card that's not the master (where we pick out the frames as they
71 // come, as fast as we can process), there's going to be a queue. The question
72 // is when we should drop frames from that queue (apart from the obvious
73 // dropping if the 16-frame queue should become full), especially given that
74 // the frame rate could be lower or higher than the master (either subtly or
75 // dramatically). We have two (conflicting) demands:
77 // 1. We want to avoid starving the queue.
78 // 2. We don't want to add more delay than is needed.
80 // Our general strategy is to drop as many frames as we can (helping for #2)
81 // that we think is safe for #1 given jitter. To this end, we measure the
82 // deviation from the expected arrival time for all cards, and use that for
83 // continuous jitter estimation.
85 // We then drop everything from the queue that we're sure we won't need to
86 // serve the output in the time before the next frame arrives. Typically,
87 // this means the queue will contain 0 or 1 frames, although more is also
88 // possible if the jitter is very high.
89 class QueueLengthPolicy {
91 QueueLengthPolicy() {}
93 void register_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
94 void unregister_metrics(const std::vector<std::pair<std::string, std::string>> &labels);
96 // Call after picking out a frame, so 0 means starvation.
97 // Note that the policy has no memory; everything is given in as parameters.
98 void update_policy(std::chrono::steady_clock::time_point now,
99 std::chrono::steady_clock::time_point expected_next_input_frame,
100 int64_t input_frame_duration,
101 int64_t master_frame_duration,
102 double max_input_card_jitter_seconds,
103 double max_master_card_jitter_seconds);
104 unsigned get_safe_queue_length() const { return safe_queue_length; }
107 unsigned safe_queue_length = 0; // Can never go below zero.
110 std::atomic<int64_t> metric_input_queue_safe_length_frames{1};
113 #endif // !defined(_QUEUE_LENGTH_POLICY_H)