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