feat(quic): replace probe UDP data plane with QUIC datagrams

This commit is contained in:
justme committed 2026-10-10 02:59:56 +08:00
1 parent 4609048865
commit b964dc2bbc
53 files changed
+2413 -411

No files matched your search

@@ -26,10 +26,18 @@ public static class LinkMetrics
"jitter.down", s_udpMsg, Ops.OwdVar(Dir.Down), "Jitter (down)", "One-way delay variation, offset and drift removed.", Unit.Ms, [Stat.P95], good: 3, poor: 20);
public static readonly MetricDef LossRun = Metric.RunLength(
"loss.run", s_udpMsg, Ops.LossRuns, "Loss runs", "Lengths of consecutive lost packets.",
"loss.run", s_udpMsg, Ops.LossRuns(Dir.Down), "Loss runs (legacy downstream alias)", "Legacy downstream-only consecutive loss metric.",
new MetricExport("loss.run_max", Stat.Max, Unit.Packets, 2, 10),
new MetricExport("loss.run_mean", Stat.Mean, Unit.Packets));
public static readonly MetricDef LossRunUp = Metric.Max(
"loss.run.up", s_udpMsg, Ops.LossRunMax(Dir.Up), "Upstream longest loss run", "Longest consecutive Probe-to-Agent Datagram loss run in each chunk; zero means no loss.",
Unit.Packets, "loss.run.up_max", good: 0, poor: 10);
public static readonly MetricDef LossRunDown = Metric.Max(
"loss.run.down", s_udpMsg, Ops.LossRunMax(Dir.Down), "Downstream longest loss run", "Longest consecutive Agent-to-Probe Datagram loss run in each chunk, conditioned on the request reaching the Agent; zero means no loss.",
Unit.Packets, "loss.run.down_max", good: 0, poor: 10);
public static readonly MetricDef Reorder = Metric.Counter(
"reorder", s_udpMsg, Ops.Reorder, "Reordered", "Replies that arrived after a higher sequence number.", new MetricExport("reorder", Stat.Count, Unit.Count));
@@ -11,7 +11,7 @@ public static class MetricRegistry
public static readonly IReadOnlyList<MetricDef> All =
[
LinkMetrics.Rtt, LinkMetrics.LossUp, LinkMetrics.LossDown, LinkMetrics.LossTotal,
LinkMetrics.JitterUp, LinkMetrics.JitterDown, LinkMetrics.LossRun, LinkMetrics.Reorder, LinkMetrics.Dup, LinkMetrics.SamplesUdp,
LinkMetrics.JitterUp, LinkMetrics.JitterDown, LinkMetrics.LossRun, LinkMetrics.LossRunUp, LinkMetrics.LossRunDown, LinkMetrics.Reorder, LinkMetrics.Dup, LinkMetrics.SamplesUdp,
TcpMetrics.Rtt, TcpMetrics.Slow, TcpMetrics.Stall, TcpMetrics.StallMax, TcpMetrics.Retrans, TcpMetrics.Srtt, TcpMetrics.BulkRate, TcpMetrics.BulkPeak, TcpMetrics.SamplesTcp,
SessionMetrics.ConnectSuccess, SessionMetrics.ConnectTcp, SessionMetrics.ConnectTls,
SessionMetrics.FailTimeout, SessionMetrics.FailReset, SessionMetrics.FailTls,
@@ -22,7 +22,8 @@ public static class MetricRegistry
/// <summary>指标分组:前端按分组展示,新增指标只需要把它加进分组。</summary>
public static readonly IReadOnlyList<PresetDef> Presets =
[
new("udp", "UDP", ["loss.up", "loss.down", "loss.total", "rtt.p50", "rtt.p95", "rtt.p99", "jitter.up.p95", "jitter.down.p95", "loss.run_max", "reorder", "dup"]),
new("quic", "QUIC Datagram", ["loss.up", "loss.down", "loss.total", "rtt.p50", "rtt.p95", "rtt.p99", "jitter.up.p95", "jitter.down.p95", "loss.run.up_max", "loss.run.down_max", "reorder", "dup"]),
new("udp", "QUIC Datagram (legacy alias)", ["loss.up", "loss.down", "loss.total", "rtt.p50", "rtt.p95", "rtt.p99", "jitter.up.p95", "jitter.down.p95", "loss.run_max", "reorder", "dup"]),
new("tcp", "TCP", ["tcp.rtt.p50", "tcp.rtt.p95", "tcp.rtt.p99", "tcp.rtt.p999", "tcp.rtt.max", "tcp.rtt.tail", "tcp.slow", "tcp.stall.per_min", "tcp.stall_max", "tcp.retrans", "tcp.srtt.p95"]),
new("connect", "Connect", ["connect.success", "connect.tcp_ms.p95", "connect.tls_ms.p95", "connect.fail.timeout", "connect.fail.rst", "connect.fail.tls"]),
new("session", "Session", ["hold.resets.per_h", "hold.lifetime_max"]),
@@ -1,6 +1,6 @@
namespace MoShiro.Backend.Analysis;
/// <summary>连接建立、长连接、UDP 压测(udp_train)的指标。</summary>
/// <summary>TCP 建连、长连接、QUIC Datagram 压测(legacy kind udp_train)的指标。</summary>
public static class SessionMetrics
{
private static readonly string[] s_conn = ["connect.conn"];
@@ -33,13 +33,13 @@ public static class SessionMetrics
"hold.lifetime", s_hold, Ops.HoldLifetime, "Longest connection lifetime", "Longest observed lifetime of the held connection (ms).", Unit.Ms, "hold.lifetime_max", Better.Higher);
public static readonly MetricDef TrainLoss = Metric.Ratio(
"train.loss", s_train, Ops.TrainLoss, "Train loss", "Loss of a fixed-rate UDP stream.", good: 1, poor: 10);
"train.loss", s_train, Ops.TrainLoss, "Datagram train loss", "Loss of a fixed-rate QUIC Datagram stream.", good: 1, poor: 10);
public static readonly MetricDef TrainJitter = Metric.Histogram(
"train.jitter", s_train, Ops.TrainJitter, "Train jitter", "One-way delay variation of a fixed-rate UDP stream.", Unit.Ms, [Stat.P95], good: 3, poor: 20);
"train.jitter", s_train, Ops.TrainJitter, "Datagram train delay variation", "One-way delay variation of a fixed-rate QUIC Datagram stream.", Unit.Ms, [Stat.P95], good: 3, poor: 20);
public static readonly MetricDef TrainRate = Metric.Histogram(
"train.rate", s_train, Ops.TrainRate, "Train rate", "Delivered rate of a fixed-rate UDP stream per second.", Unit.Mbps, [Stat.P50, Stat.Mean], Better.Higher);
"train.rate", s_train, Ops.TrainRate, "Datagram train rate", "Delivered rate of a fixed-rate QUIC Datagram stream per second.", Unit.Mbps, [Stat.P50, Stat.Mean], Better.Higher);
public static readonly MetricDef SamplesConnect = Metric.Counter(
"n.connect", s_conn, Ops.RowCount, "connect samples", "Number of connection attempts.", new MetricExport("n.connect", Stat.Count, Unit.Count), Better.None);
@@ -12,9 +12,9 @@ public sealed class BaseModel : ServiceModelBase
public override string Title => "Line quality";
public override int Version => 1;
public override int Version => 2;
public override IReadOnlyList<string> Needs { get; } = ["rtt.p95", "loss.total", "jitter.down.p95", "jitter.up.p95", "loss.run_max"];
public override IReadOnlyList<string> Needs { get; } = ["rtt.p95", "loss.total", "jitter.down.p95", "jitter.up.p95", "loss.run.up_max", "loss.run.down_max"];
public override ModelScore Score(MetricView view) =>
Combine(
@@ -24,7 +24,8 @@ public sealed class BaseModel : ServiceModelBase
Dim(view, "loss.total", "Packet loss", 4, s_loss),
Dim(view, "jitter.down.p95", "Jitter (down)", 1.5, s_jitter),
Dim(view, "jitter.up.p95", "Jitter (up)", 1.5, s_jitter),
Dim(view, "loss.run_max", "Longest loss run", 1, s_run),
Dim(view, "loss.run.up_max", "Longest upstream loss run", 0.5, s_run),
Dim(view, "loss.run.down_max", "Longest downstream loss run", 0.5, s_run),
],
estimated: false);
}
@@ -273,7 +273,14 @@ internal static class JitterEmitter
/// <summary>连续丢失段的长度(按发送顺序)。</summary>
internal sealed class LossRunsOp : MetricOp
{
public override string Signature => "loss_runs/1";
private readonly Dir _direction;
public LossRunsOp(Dir direction)
{
_direction = direction;
}
public override string Signature => $"loss_runs/{_direction}/2";
public override void Evaluate(ChunkView view, RawStream stream, IObservationSink sink)
{
@@ -292,7 +299,13 @@ internal sealed class LossRunsOp : MetricOp
continue;
}
if (c.IsReplyLost(i))
bool lost = _direction switch
{
Dir.Up => c.IsUpLost(i),
Dir.Down => !c.IsUpLost(i) && c.IsReplyLost(i),
_ => c.IsReplyLost(i),
};
if (lost)
{
if (run == 0)
{
@@ -315,6 +328,59 @@ internal sealed class LossRunsOp : MetricOp
}
}
/// <summary>每个块一个最长连续丢包值;无丢包的有效块明确输出 0。</summary>
internal sealed class LossRunMaxOp(Dir direction) : MetricOp
{
public override string Signature => $"loss_run_max/{direction}/1";
public override void Evaluate(ChunkView view, RawStream stream, IObservationSink sink)
{
var c = new EchoColumns(stream);
if (!c.HasBasics || (direction != Dir.Total && c.T2 is null))
{
return;
}
int run = 0;
int maximum = 0;
bool hasSample = false;
long sampleAt = 0;
for (int i = 0; i < c.Count; i++)
{
if (!c.T1!.IsValid(i))
{
continue;
}
bool lost = direction switch
{
Dir.Up => c.IsUpLost(i),
Dir.Down => !c.IsUpLost(i) && c.IsReplyLost(i),
_ => c.IsReplyLost(i),
};
if (direction == Dir.Down && c.IsUpLost(i))
{
// A request lost before reaching the Agent is not a downstream sample.
run = 0;
continue;
}
if (!hasSample)
{
sampleAt = c.SentUtcMs(view, i);
hasSample = true;
}
run = lost ? run + 1 : 0;
maximum = Math.Max(maximum, run);
}
if (hasSample)
{
sink.Emit(sampleAt, maximum);
}
}
}
/// <summary>乱序:按到达顺序,序号比之前见过的最大序号还小。重复的序号不算乱序。</summary>
internal sealed class ReorderOp : MetricOp
{
@@ -11,7 +11,9 @@ public static class Ops
public static MetricOp OwdVar(Dir dir) => new OwdVarOp(dir);
public static MetricOp LossRuns { get; } = new LossRunsOp();
public static MetricOp LossRuns(Dir direction) => new LossRunsOp(direction);
public static MetricOp LossRunMax(Dir direction) => new LossRunMaxOp(direction);
public static MetricOp Reorder { get; } = new ReorderOp();