214 lines
8.3 KiB
Go
214 lines
8.3 KiB
Go
|
package congestion
|
||
|
|
|
||
|
|
import (
|
||
|
|
"math"
|
||
|
|
"time"
|
||
|
|
|
||
|
|
"github.com/metacubex/quic-go/congestion"
|
||
|
|
)
|
||
|
|
|
||
|
|
// This cubic implementation is based on the one found in Chromiums's QUIC
|
||
|
|
// implementation, in the files net/quic/congestion_control/cubic.{hh,cc}.
|
||
|
|
|
||
|
|
// Constants based on TCP defaults.
|
||
|
|
// The following constants are in 2^10 fractions of a second instead of ms to
|
||
|
|
// allow a 10 shift right to divide.
|
||
|
|
|
||
|
|
// 1024*1024^3 (first 1024 is from 0.100^3)
|
||
|
|
// where 0.100 is 100 ms which is the scaling round trip time.
|
||
|
|
const (
|
||
|
|
cubeScale = 40
|
||
|
|
cubeCongestionWindowScale = 410
|
||
|
|
cubeFactor congestion.ByteCount = 1 << cubeScale / cubeCongestionWindowScale / maxDatagramSize
|
||
|
|
// TODO: when re-enabling cubic, make sure to use the actual packet size here
|
||
|
|
maxDatagramSize = congestion.ByteCount(InitialPacketSizeIPv4)
|
||
|
|
)
|
||
|
|
|
||
|
|
const defaultNumConnections = 1
|
||
|
|
|
||
|
|
// Default Cubic backoff factor
|
||
|
|
const beta float32 = 0.7
|
||
|
|
|
||
|
|
// Additional backoff factor when loss occurs in the concave part of the Cubic
|
||
|
|
// curve. This additional backoff factor is expected to give up bandwidth to
|
||
|
|
// new concurrent flows and speed up convergence.
|
||
|
|
const betaLastMax float32 = 0.85
|
||
|
|
|
||
|
|
// Cubic implements the cubic algorithm from TCP
|
||
|
|
type Cubic struct {
|
||
|
|
clock Clock
|
||
|
|
|
||
|
|
// Number of connections to simulate.
|
||
|
|
numConnections int
|
||
|
|
|
||
|
|
// Time when this cycle started, after last loss event.
|
||
|
|
epoch time.Time
|
||
|
|
|
||
|
|
// Max congestion window used just before last loss event.
|
||
|
|
// Note: to improve fairness to other streams an additional back off is
|
||
|
|
// applied to this value if the new value is below our latest value.
|
||
|
|
lastMaxCongestionWindow congestion.ByteCount
|
||
|
|
|
||
|
|
// Number of acked bytes since the cycle started (epoch).
|
||
|
|
ackedBytesCount congestion.ByteCount
|
||
|
|
|
||
|
|
// TCP Reno equivalent congestion window in packets.
|
||
|
|
estimatedTCPcongestionWindow congestion.ByteCount
|
||
|
|
|
||
|
|
// Origin point of cubic function.
|
||
|
|
originPointCongestionWindow congestion.ByteCount
|
||
|
|
|
||
|
|
// Time to origin point of cubic function in 2^10 fractions of a second.
|
||
|
|
timeToOriginPoint uint32
|
||
|
|
|
||
|
|
// Last congestion window in packets computed by cubic function.
|
||
|
|
lastTargetCongestionWindow congestion.ByteCount
|
||
|
|
}
|
||
|
|
|
||
|
|
// NewCubic returns a new Cubic instance
|
||
|
|
func NewCubic(clock Clock) *Cubic {
|
||
|
|
c := &Cubic{
|
||
|
|
clock: clock,
|
||
|
|
numConnections: defaultNumConnections,
|
||
|
|
}
|
||
|
|
c.Reset()
|
||
|
|
return c
|
||
|
|
}
|
||
|
|
|
||
|
|
// Reset is called after a timeout to reset the cubic state
|
||
|
|
func (c *Cubic) Reset() {
|
||
|
|
c.epoch = time.Time{}
|
||
|
|
c.lastMaxCongestionWindow = 0
|
||
|
|
c.ackedBytesCount = 0
|
||
|
|
c.estimatedTCPcongestionWindow = 0
|
||
|
|
c.originPointCongestionWindow = 0
|
||
|
|
c.timeToOriginPoint = 0
|
||
|
|
c.lastTargetCongestionWindow = 0
|
||
|
|
}
|
||
|
|
|
||
|
|
func (c *Cubic) alpha() float32 {
|
||
|
|
// TCPFriendly alpha is described in Section 3.3 of the CUBIC paper. Note that
|
||
|
|
// beta here is a cwnd multiplier, and is equal to 1-beta from the paper.
|
||
|
|
// We derive the equivalent alpha for an N-connection emulation as:
|
||
|
|
b := c.beta()
|
||
|
|
return 3 * float32(c.numConnections) * float32(c.numConnections) * (1 - b) / (1 + b)
|
||
|
|
}
|
||
|
|
|
||
|
|
func (c *Cubic) beta() float32 {
|
||
|
|
// kNConnectionBeta is the backoff factor after loss for our N-connection
|
||
|
|
// emulation, which emulates the effective backoff of an ensemble of N
|
||
|
|
// TCP-Reno connections on a single loss event. The effective multiplier is
|
||
|
|
// computed as:
|
||
|
|
return (float32(c.numConnections) - 1 + beta) / float32(c.numConnections)
|
||
|
|
}
|
||
|
|
|
||
|
|
func (c *Cubic) betaLastMax() float32 {
|
||
|
|
// betaLastMax is the additional backoff factor after loss for our
|
||
|
|
// N-connection emulation, which emulates the additional backoff of
|
||
|
|
// an ensemble of N TCP-Reno connections on a single loss event. The
|
||
|
|
// effective multiplier is computed as:
|
||
|
|
return (float32(c.numConnections) - 1 + betaLastMax) / float32(c.numConnections)
|
||
|
|
}
|
||
|
|
|
||
|
|
// OnApplicationLimited is called on ack arrival when sender is unable to use
|
||
|
|
// the available congestion window. Resets Cubic state during quiescence.
|
||
|
|
func (c *Cubic) OnApplicationLimited() {
|
||
|
|
// When sender is not using the available congestion window, the window does
|
||
|
|
// not grow. But to be RTT-independent, Cubic assumes that the sender has been
|
||
|
|
// using the entire window during the time since the beginning of the current
|
||
|
|
// "epoch" (the end of the last loss recovery period). Since
|
||
|
|
// application-limited periods break this assumption, we reset the epoch when
|
||
|
|
// in such a period. This reset effectively freezes congestion window growth
|
||
|
|
// through application-limited periods and allows Cubic growth to continue
|
||
|
|
// when the entire window is being used.
|
||
|
|
c.epoch = time.Time{}
|
||
|
|
}
|
||
|
|
|
||
|
|
// CongestionWindowAfterPacketLoss computes a new congestion window to use after
|
||
|
|
// a loss event. Returns the new congestion window in packets. The new
|
||
|
|
// congestion window is a multiplicative decrease of our current window.
|
||
|
|
func (c *Cubic) CongestionWindowAfterPacketLoss(currentCongestionWindow congestion.ByteCount) congestion.ByteCount {
|
||
|
|
if currentCongestionWindow+maxDatagramSize < c.lastMaxCongestionWindow {
|
||
|
|
// We never reached the old max, so assume we are competing with another
|
||
|
|
// flow. Use our extra back off factor to allow the other flow to go up.
|
||
|
|
c.lastMaxCongestionWindow = congestion.ByteCount(c.betaLastMax() * float32(currentCongestionWindow))
|
||
|
|
} else {
|
||
|
|
c.lastMaxCongestionWindow = currentCongestionWindow
|
||
|
|
}
|
||
|
|
c.epoch = time.Time{} // Reset time.
|
||
|
|
return congestion.ByteCount(float32(currentCongestionWindow) * c.beta())
|
||
|
|
}
|
||
|
|
|
||
|
|
// CongestionWindowAfterAck computes a new congestion window to use after a received ACK.
|
||
|
|
// Returns the new congestion window in packets. The new congestion window
|
||
|
|
// follows a cubic function that depends on the time passed since last
|
||
|
|
// packet loss.
|
||
|
|
func (c *Cubic) CongestionWindowAfterAck(
|
||
|
|
ackedBytes congestion.ByteCount,
|
||
|
|
currentCongestionWindow congestion.ByteCount,
|
||
|
|
delayMin time.Duration,
|
||
|
|
eventTime time.Time,
|
||
|
|
) congestion.ByteCount {
|
||
|
|
c.ackedBytesCount += ackedBytes
|
||
|
|
|
||
|
|
if c.epoch.IsZero() {
|
||
|
|
// First ACK after a loss event.
|
||
|
|
c.epoch = eventTime // Start of epoch.
|
||
|
|
c.ackedBytesCount = ackedBytes // Reset count.
|
||
|
|
// Reset estimated_tcp_congestion_window_ to be in sync with cubic.
|
||
|
|
c.estimatedTCPcongestionWindow = currentCongestionWindow
|
||
|
|
if c.lastMaxCongestionWindow <= currentCongestionWindow {
|
||
|
|
c.timeToOriginPoint = 0
|
||
|
|
c.originPointCongestionWindow = currentCongestionWindow
|
||
|
|
} else {
|
||
|
|
c.timeToOriginPoint = uint32(math.Cbrt(float64(cubeFactor * (c.lastMaxCongestionWindow - currentCongestionWindow))))
|
||
|
|
c.originPointCongestionWindow = c.lastMaxCongestionWindow
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
// Change the time unit from microseconds to 2^10 fractions per second. Take
|
||
|
|
// the round trip time in account. This is done to allow us to use shift as a
|
||
|
|
// divide operator.
|
||
|
|
elapsedTime := int64(eventTime.Add(delayMin).Sub(c.epoch)/time.Microsecond) << 10 / (1000 * 1000)
|
||
|
|
|
||
|
|
// Right-shifts of negative, signed numbers have implementation-dependent
|
||
|
|
// behavior, so force the offset to be positive, as is done in the kernel.
|
||
|
|
offset := int64(c.timeToOriginPoint) - elapsedTime
|
||
|
|
if offset < 0 {
|
||
|
|
offset = -offset
|
||
|
|
}
|
||
|
|
|
||
|
|
deltaCongestionWindow := congestion.ByteCount(cubeCongestionWindowScale*offset*offset*offset) * maxDatagramSize >> cubeScale
|
||
|
|
var targetCongestionWindow congestion.ByteCount
|
||
|
|
if elapsedTime > int64(c.timeToOriginPoint) {
|
||
|
|
targetCongestionWindow = c.originPointCongestionWindow + deltaCongestionWindow
|
||
|
|
} else {
|
||
|
|
targetCongestionWindow = c.originPointCongestionWindow - deltaCongestionWindow
|
||
|
|
}
|
||
|
|
// Limit the CWND increase to half the acked bytes.
|
||
|
|
targetCongestionWindow = Min(targetCongestionWindow, currentCongestionWindow+c.ackedBytesCount/2)
|
||
|
|
|
||
|
|
// Increase the window by approximately Alpha * 1 MSS of bytes every
|
||
|
|
// time we ack an estimated tcp window of bytes. For small
|
||
|
|
// congestion windows (less than 25), the formula below will
|
||
|
|
// increase slightly slower than linearly per estimated tcp window
|
||
|
|
// of bytes.
|
||
|
|
c.estimatedTCPcongestionWindow += congestion.ByteCount(float32(c.ackedBytesCount) * c.alpha() * float32(maxDatagramSize) / float32(c.estimatedTCPcongestionWindow))
|
||
|
|
c.ackedBytesCount = 0
|
||
|
|
|
||
|
|
// We have a new cubic congestion window.
|
||
|
|
c.lastTargetCongestionWindow = targetCongestionWindow
|
||
|
|
|
||
|
|
// Compute target congestion_window based on cubic target and estimated TCP
|
||
|
|
// congestion_window, use highest (fastest).
|
||
|
|
if targetCongestionWindow < c.estimatedTCPcongestionWindow {
|
||
|
|
targetCongestionWindow = c.estimatedTCPcongestionWindow
|
||
|
|
}
|
||
|
|
return targetCongestionWindow
|
||
|
|
}
|
||
|
|
|
||
|
|
// SetNumConnections sets the number of emulated connections
|
||
|
|
func (c *Cubic) SetNumConnections(n int) {
|
||
|
|
c.numConnections = n
|
||
|
|
}
|