diff --git a/internal/ackhandler/sent_packet_handler.go b/internal/ackhandler/sent_packet_handler.go
index 72911e13..1aead119 100644
--- a/internal/ackhandler/sent_packet_handler.go
+++ b/internal/ackhandler/sent_packet_handler.go
@@ -220,11 +220,11 @@ func (h *sentPacketHandler) ReceivedAck(ackFrame *wire.AckFrame, withPacketNumbe
 		if p.largestAcked != 0 {
 			h.lowestPacketNotConfirmedAcked = utils.MaxPacketNumber(h.lowestPacketNotConfirmedAcked, p.largestAcked+1)
 		}
-		if err := h.onPacketAcked(p); err != nil {
+		if err := h.onPacketAcked(p, rcvTime); err != nil {
 			return err
 		}
 		if p.includedInBytesInFlight {
-			h.congestion.OnPacketAcked(p.PacketNumber, p.Length, priorInFlight)
+			h.congestion.OnPacketAcked(p.PacketNumber, p.Length, priorInFlight, rcvTime)
 		}
 	}
 
@@ -382,7 +382,7 @@ func (h *sentPacketHandler) GetAlarmTimeout() time.Time {
 	return h.alarm
 }
 
-func (h *sentPacketHandler) onPacketAcked(p *Packet) error {
+func (h *sentPacketHandler) onPacketAcked(p *Packet, rcvTime time.Time) error {
 	// This happens if a packet and its retransmissions is acked in the same ACK.
 	// As soon as we process the first one, this will remove all the retransmissions,
 	// so we won't find the retransmitted packet number later.
diff --git a/internal/ackhandler/sent_packet_handler_test.go b/internal/ackhandler/sent_packet_handler_test.go
index 37df5965..ecd608be 100644
--- a/internal/ackhandler/sent_packet_handler_test.go
+++ b/internal/ackhandler/sent_packet_handler_test.go
@@ -568,18 +568,19 @@ var _ = Describe("SentPacketHandler", func() {
 		})
 
 		It("should call MaybeExitSlowStart and OnPacketAcked", func() {
+			rcvTime := time.Now().Add(-5 * time.Second)
 			cong.EXPECT().OnPacketSent(gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any()).Times(3)
 			cong.EXPECT().TimeUntilSend(gomock.Any()).Times(3)
 			gomock.InOrder(
 				cong.EXPECT().MaybeExitSlowStart(), // must be called before packets are acked
-				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(1), protocol.ByteCount(1), protocol.ByteCount(3)),
-				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(2), protocol.ByteCount(1), protocol.ByteCount(3)),
+				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(1), protocol.ByteCount(1), protocol.ByteCount(3), rcvTime),
+				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(2), protocol.ByteCount(1), protocol.ByteCount(3), rcvTime),
 			)
 			handler.SentPacket(retransmittablePacket(&Packet{PacketNumber: 1}))
 			handler.SentPacket(retransmittablePacket(&Packet{PacketNumber: 2}))
 			handler.SentPacket(retransmittablePacket(&Packet{PacketNumber: 3}))
 			ack := &wire.AckFrame{AckRanges: []wire.AckRange{{Smallest: 1, Largest: 2}}}
-			err := handler.ReceivedAck(ack, 1, protocol.EncryptionForwardSecure, time.Now())
+			err := handler.ReceivedAck(ack, 1, protocol.EncryptionForwardSecure, rcvTime)
 			Expect(err).NotTo(HaveOccurred())
 		})
 
@@ -611,10 +612,11 @@ var _ = Describe("SentPacketHandler", func() {
 			Expect(handler.DequeuePacketForRetransmission()).ToNot(BeNil())
 			Expect(handler.DequeuePacketForRetransmission()).ToNot(BeNil())
 			// send one probe packet and receive an ACK for it
+			rcvTime := time.Now()
 			gomock.InOrder(
 				cong.EXPECT().MaybeExitSlowStart(),
 				cong.EXPECT().OnRetransmissionTimeout(true),
-				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(5), protocol.ByteCount(1), protocol.ByteCount(5)),
+				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(5), protocol.ByteCount(1), protocol.ByteCount(5), rcvTime),
 				cong.EXPECT().OnPacketLost(protocol.PacketNumber(1), protocol.ByteCount(1), protocol.ByteCount(5)),
 				cong.EXPECT().OnPacketLost(protocol.PacketNumber(2), protocol.ByteCount(1), protocol.ByteCount(5)),
 				cong.EXPECT().OnPacketLost(protocol.PacketNumber(3), protocol.ByteCount(1), protocol.ByteCount(5)),
@@ -622,7 +624,7 @@ var _ = Describe("SentPacketHandler", func() {
 			)
 			handler.SentPacket(retransmittablePacket(&Packet{PacketNumber: 5}))
 			ack := &wire.AckFrame{AckRanges: []wire.AckRange{{Smallest: 5, Largest: 5}}}
-			err := handler.ReceivedAck(ack, 1, protocol.EncryptionForwardSecure, time.Now())
+			err := handler.ReceivedAck(ack, 1, protocol.EncryptionForwardSecure, rcvTime)
 			Expect(err).ToNot(HaveOccurred())
 		})
 
@@ -643,7 +645,7 @@ var _ = Describe("SentPacketHandler", func() {
 			// don't EXPECT any call to OnRetransmissionTimeout
 			gomock.InOrder(
 				cong.EXPECT().MaybeExitSlowStart(),
-				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(2), protocol.ByteCount(1), protocol.ByteCount(3)),
+				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(2), protocol.ByteCount(1), protocol.ByteCount(3), gomock.Any()),
 				cong.EXPECT().OnPacketLost(protocol.PacketNumber(1), protocol.ByteCount(1), protocol.ByteCount(3)),
 			)
 			ack := &wire.AckFrame{AckRanges: []wire.AckRange{{Smallest: 2, Largest: 2}}}
@@ -659,7 +661,7 @@ var _ = Describe("SentPacketHandler", func() {
 			// lose packet 1
 			gomock.InOrder(
 				cong.EXPECT().MaybeExitSlowStart(),
-				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(2), protocol.ByteCount(1), protocol.ByteCount(2)),
+				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(2), protocol.ByteCount(1), protocol.ByteCount(2), gomock.Any()),
 				cong.EXPECT().OnPacketLost(protocol.PacketNumber(1), protocol.ByteCount(1), protocol.ByteCount(2)),
 			)
 			ack := &wire.AckFrame{AckRanges: []wire.AckRange{{Smallest: 2, Largest: 2}}}
@@ -681,7 +683,7 @@ var _ = Describe("SentPacketHandler", func() {
 			// receive the first ACK
 			gomock.InOrder(
 				cong.EXPECT().MaybeExitSlowStart(),
-				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(2), protocol.ByteCount(1), protocol.ByteCount(4)),
+				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(2), protocol.ByteCount(1), protocol.ByteCount(4), gomock.Any()),
 				cong.EXPECT().OnPacketLost(protocol.PacketNumber(1), protocol.ByteCount(1), protocol.ByteCount(4)),
 			)
 			ack := &wire.AckFrame{AckRanges: []wire.AckRange{{Smallest: 2, Largest: 2}}}
@@ -690,7 +692,7 @@ var _ = Describe("SentPacketHandler", func() {
 			// receive the second ACK
 			gomock.InOrder(
 				cong.EXPECT().MaybeExitSlowStart(),
-				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(4), protocol.ByteCount(1), protocol.ByteCount(2)),
+				cong.EXPECT().OnPacketAcked(protocol.PacketNumber(4), protocol.ByteCount(1), protocol.ByteCount(2), gomock.Any()),
 				cong.EXPECT().OnPacketLost(protocol.PacketNumber(3), protocol.ByteCount(1), protocol.ByteCount(2)),
 			)
 			ack = &wire.AckFrame{AckRanges: []wire.AckRange{{Smallest: 4, Largest: 4}}}
diff --git a/internal/congestion/cubic.go b/internal/congestion/cubic.go
index 3922f476..dcf91fc6 100644
--- a/internal/congestion/cubic.go
+++ b/internal/congestion/cubic.go
@@ -16,11 +16,10 @@ import (
 // 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.
+// where 0.100 is 100 ms which is the scaling round trip time.
 const cubeScale = 40
 const cubeCongestionWindowScale = 410
-const cubeFactor protocol.PacketNumber = 1 << cubeScale / cubeCongestionWindowScale
+const cubeFactor protocol.ByteCount = 1 << cubeScale / cubeCongestionWindowScale / protocol.DefaultTCPMSS
 
 const defaultNumConnections = 2
 
@@ -32,39 +31,35 @@ const beta float32 = 0.7
 // new concurrent flows and speed up convergence.
 const betaLastMax float32 = 0.85
 
-// If true, Cubic's epoch is shifted when the sender is application-limited.
-const shiftQuicCubicEpochWhenAppLimited = true
-
-const maxCubicTimeInterval = 30 * time.Millisecond
-
 // 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
-	// Time when sender went into application-limited period. Zero if not in
-	// application-limited period.
-	appLimitedStartTime time.Time
-	// Time when we updated last_congestion_window.
-	lastUpdateTime time.Time
-	// Last congestion window (in packets) used.
-	lastCongestionWindow protocol.PacketNumber
-	// Max congestion window (in packets) used just before last loss event.
+
+	// 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 protocol.PacketNumber
-	// Number of acked packets since the cycle started (epoch).
-	ackedPacketsCount protocol.PacketNumber
+	lastMaxCongestionWindow protocol.ByteCount
+
+	// Number of acked bytes since the cycle started (epoch).
+	ackedBytesCount protocol.ByteCount
+
 	// TCP Reno equivalent congestion window in packets.
-	estimatedTCPcongestionWindow protocol.PacketNumber
+	estimatedTCPcongestionWindow protocol.ByteCount
+
 	// Origin point of cubic function.
-	originPointCongestionWindow protocol.PacketNumber
+	originPointCongestionWindow protocol.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 protocol.PacketNumber
+	lastTargetCongestionWindow protocol.ByteCount
 }
 
 // NewCubic returns a new Cubic instance
@@ -80,11 +75,8 @@ func NewCubic(clock Clock) *Cubic {
 // Reset is called after a timeout to reset the cubic state
 func (c *Cubic) Reset() {
 	c.epoch = time.Time{}
-	c.appLimitedStartTime = time.Time{}
-	c.lastUpdateTime = time.Time{}
-	c.lastCongestionWindow = 0
 	c.lastMaxCongestionWindow = 0
-	c.ackedPacketsCount = 0
+	c.ackedBytesCount = 0
 	c.estimatedTCPcongestionWindow = 0
 	c.originPointCongestionWindow = 0
 	c.timeToOriginPoint = 0
@@ -107,57 +99,59 @@ func (c *Cubic) beta() float32 {
 	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() {
-	if shiftQuicCubicEpochWhenAppLimited {
-		// When sender is not using the available congestion window, Cubic's epoch
-		// should not continue growing. Record the time when sender goes into an
-		// app-limited period here, to compensate later when cwnd growth happens.
-		if c.appLimitedStartTime.IsZero() {
-			c.appLimitedStartTime = c.clock.Now()
-		}
-	} else {
-		// When sender is not using the available congestion window, Cubic's epoch
-		// should not continue growing. Reset the epoch when in such a period.
-		c.epoch = time.Time{}
-	}
+	// 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 protocol.PacketNumber) protocol.PacketNumber {
-	if currentCongestionWindow < c.lastMaxCongestionWindow {
+func (c *Cubic) CongestionWindowAfterPacketLoss(currentCongestionWindow protocol.ByteCount) protocol.ByteCount {
+	if currentCongestionWindow+protocol.DefaultTCPMSS < 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 = protocol.PacketNumber(betaLastMax * float32(currentCongestionWindow))
+		c.lastMaxCongestionWindow = protocol.ByteCount(c.betaLastMax() * float32(currentCongestionWindow))
 	} else {
 		c.lastMaxCongestionWindow = currentCongestionWindow
 	}
 	c.epoch = time.Time{} // Reset time.
-	return protocol.PacketNumber(float32(currentCongestionWindow) * c.beta())
+	return protocol.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(currentCongestionWindow protocol.PacketNumber, delayMin time.Duration) protocol.PacketNumber {
-	c.ackedPacketsCount++ // Packets acked.
-	currentTime := c.clock.Now()
-
-	// Cubic is "independent" of RTT, the update is limited by the time elapsed.
-	if c.lastCongestionWindow == currentCongestionWindow && (currentTime.Sub(c.lastUpdateTime) <= maxCubicTimeInterval) {
-		return utils.MaxPacketNumber(c.lastTargetCongestionWindow, c.estimatedTCPcongestionWindow)
-	}
-	c.lastCongestionWindow = currentCongestionWindow
-	c.lastUpdateTime = currentTime
+func (c *Cubic) CongestionWindowAfterAck(
+	ackedBytes protocol.ByteCount,
+	currentCongestionWindow protocol.ByteCount,
+	delayMin time.Duration,
+	eventTime time.Time,
+) protocol.ByteCount {
+	c.ackedBytesCount += ackedBytes
 
 	if c.epoch.IsZero() {
 		// First ACK after a loss event.
-		c.epoch = currentTime   // Start of epoch.
-		c.ackedPacketsCount = 1 // Reset count.
+		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 {
@@ -167,48 +161,37 @@ func (c *Cubic) CongestionWindowAfterAck(currentCongestionWindow protocol.Packet
 			c.timeToOriginPoint = uint32(math.Cbrt(float64(cubeFactor * (c.lastMaxCongestionWindow - currentCongestionWindow))))
 			c.originPointCongestionWindow = c.lastMaxCongestionWindow
 		}
-	} else {
-		// If sender was app-limited, then freeze congestion window growth during
-		// app-limited period. Continue growth now by shifting the epoch-start
-		// through the app-limited period.
-		if shiftQuicCubicEpochWhenAppLimited && !c.appLimitedStartTime.IsZero() {
-			shift := currentTime.Sub(c.appLimitedStartTime)
-			c.epoch = c.epoch.Add(shift)
-			c.appLimitedStartTime = time.Time{}
-		}
 	}
 
 	// 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((currentTime.Add(delayMin).Sub(c.epoch)/time.Microsecond)<<10) / 1000000
+	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
-	// Right-shifts of negative, signed numbers have
-	// implementation-dependent behavior.  Force the offset to be
-	// positive, similar to the kernel implementation.
 	if offset < 0 {
 		offset = -offset
 	}
-	deltaCongestionWindow := protocol.PacketNumber((cubeCongestionWindowScale * offset * offset * offset) >> cubeScale)
-	var targetCongestionWindow protocol.PacketNumber
+
+	deltaCongestionWindow := protocol.ByteCount(cubeCongestionWindowScale*offset*offset*offset) * protocol.DefaultTCPMSS >> cubeScale
+	var targetCongestionWindow protocol.ByteCount
 	if elapsedTime > int64(c.timeToOriginPoint) {
 		targetCongestionWindow = c.originPointCongestionWindow + deltaCongestionWindow
 	} else {
 		targetCongestionWindow = c.originPointCongestionWindow - deltaCongestionWindow
 	}
-	// With dynamic beta/alpha based on number of active streams, it is possible
-	// for the required_ack_count to become much lower than acked_packets_count_
-	// suddenly, leading to more than one iteration through the following loop.
-	for {
-		// Update estimated TCP congestion_window.
-		requiredAckCount := protocol.PacketNumber(float32(c.estimatedTCPcongestionWindow) / c.alpha())
-		if c.ackedPacketsCount < requiredAckCount {
-			break
-		}
-		c.ackedPacketsCount -= requiredAckCount
-		c.estimatedTCPcongestionWindow++
-	}
+	// Limit the CWND increase to half the acked bytes.
+	targetCongestionWindow = utils.MinByteCount(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 += protocol.ByteCount(float32(c.ackedBytesCount) * c.alpha() * float32(protocol.DefaultTCPMSS) / float32(c.estimatedTCPcongestionWindow))
+	c.ackedBytesCount = 0
 
 	// We have a new cubic congestion window.
 	c.lastTargetCongestionWindow = targetCongestionWindow
@@ -218,7 +201,6 @@ func (c *Cubic) CongestionWindowAfterAck(currentCongestionWindow protocol.Packet
 	if targetCongestionWindow < c.estimatedTCPcongestionWindow {
 		targetCongestionWindow = c.estimatedTCPcongestionWindow
 	}
-
 	return targetCongestionWindow
 }
 
diff --git a/internal/congestion/cubic_sender.go b/internal/congestion/cubic_sender.go
index 21f01942..b9f67e6c 100644
--- a/internal/congestion/cubic_sender.go
+++ b/internal/congestion/cubic_sender.go
@@ -8,9 +8,9 @@ import (
 )
 
 const (
-	maxBurstBytes                                        = 3 * protocol.DefaultTCPMSS
-	defaultMinimumCongestionWindow protocol.PacketNumber = 2
-	renoBeta                       float32               = 0.7 // Reno backoff factor.
+	maxBurstBytes                                     = 3 * protocol.DefaultTCPMSS
+	renoBeta                       float32            = 0.7 // Reno backoff factor.
+	defaultMinimumCongestionWindow protocol.ByteCount = 2 * protocol.DefaultTCPMSS
 )
 
 type cubicSender struct {
@@ -31,12 +31,6 @@ type cubicSender struct {
 	// Track the largest packet number outstanding when a CWND cutback occurs.
 	largestSentAtLastCutback protocol.PacketNumber
 
-	// Congestion window in packets.
-	congestionWindow protocol.PacketNumber
-
-	// Slow start congestion window in packets, aka ssthresh.
-	slowstartThreshold protocol.PacketNumber
-
 	// Whether the last loss event caused us to exit slowstart.
 	// Used for stats collection of slowstartPacketsLost
 	lastCutbackExitedSlowstart bool
@@ -44,24 +38,35 @@ type cubicSender struct {
 	// When true, exit slow start with large cutback of congestion window.
 	slowStartLargeReduction bool
 
-	// Minimum congestion window in packets.
-	minCongestionWindow protocol.PacketNumber
+	// Congestion window in packets.
+	congestionWindow protocol.ByteCount
 
-	// Maximum number of outstanding packets for tcp.
-	maxTCPCongestionWindow protocol.PacketNumber
+	// Minimum congestion window in packets.
+	minCongestionWindow protocol.ByteCount
+
+	// Maximum congestion window.
+	maxCongestionWindow protocol.ByteCount
+
+	// Slow start congestion window in bytes, aka ssthresh.
+	slowstartThreshold protocol.ByteCount
 
 	// Number of connections to simulate.
 	numConnections int
 
 	// ACK counter for the Reno implementation.
-	congestionWindowCount protocol.ByteCount
+	numAckedPackets uint64
 
-	initialCongestionWindow    protocol.PacketNumber
-	initialMaxCongestionWindow protocol.PacketNumber
+	initialCongestionWindow    protocol.ByteCount
+	initialMaxCongestionWindow protocol.ByteCount
+
+	minSlowStartExitWindow protocol.ByteCount
 }
 
+var _ SendAlgorithm = &cubicSender{}
+var _ SendAlgorithmWithDebugInfo = &cubicSender{}
+
 // NewCubicSender makes a new cubic sender
-func NewCubicSender(clock Clock, rttStats *RTTStats, reno bool, initialCongestionWindow, initialMaxCongestionWindow protocol.PacketNumber) SendAlgorithmWithDebugInfo {
+func NewCubicSender(clock Clock, rttStats *RTTStats, reno bool, initialCongestionWindow, initialMaxCongestionWindow protocol.ByteCount) SendAlgorithmWithDebugInfo {
 	return &cubicSender{
 		rttStats:                   rttStats,
 		initialCongestionWindow:    initialCongestionWindow,
@@ -69,7 +74,7 @@ func NewCubicSender(clock Clock, rttStats *RTTStats, reno bool, initialCongestio
 		congestionWindow:           initialCongestionWindow,
 		minCongestionWindow:        defaultMinimumCongestionWindow,
 		slowstartThreshold:         initialMaxCongestionWindow,
-		maxTCPCongestionWindow:     initialMaxCongestionWindow,
+		maxCongestionWindow:        initialMaxCongestionWindow,
 		numConnections:             defaultNumConnections,
 		cubic:                      NewCubic(clock),
 		reno:                       reno,
@@ -80,21 +85,26 @@ func NewCubicSender(clock Clock, rttStats *RTTStats, reno bool, initialCongestio
 func (c *cubicSender) TimeUntilSend(bytesInFlight protocol.ByteCount) time.Duration {
 	if c.InRecovery() {
 		// PRR is used when in recovery.
-		if c.prr.TimeUntilSend(c.GetCongestionWindow(), bytesInFlight, c.GetSlowStartThreshold()) == 0 {
+		if c.prr.CanSend(c.GetCongestionWindow(), bytesInFlight, c.GetSlowStartThreshold()) {
 			return 0
 		}
 	}
-	delay := c.rttStats.SmoothedRTT() / time.Duration(2*c.GetCongestionWindow()/protocol.DefaultTCPMSS)
+	delay := c.rttStats.SmoothedRTT() / time.Duration(2*c.GetCongestionWindow())
 	if !c.InSlowStart() { // adjust delay, such that it's 1.25*cwd/rtt
 		delay = delay * 8 / 5
 	}
 	return delay
 }
 
-func (c *cubicSender) OnPacketSent(sentTime time.Time, bytesInFlight protocol.ByteCount, packetNumber protocol.PacketNumber, bytes protocol.ByteCount, isRetransmittable bool) bool {
-	// Only update bytesInFlight for data packets.
+func (c *cubicSender) OnPacketSent(
+	sentTime time.Time,
+	bytesInFlight protocol.ByteCount,
+	packetNumber protocol.PacketNumber,
+	bytes protocol.ByteCount,
+	isRetransmittable bool,
+) {
 	if !isRetransmittable {
-		return false
+		return
 	}
 	if c.InRecovery() {
 		// PRR is used when in recovery.
@@ -102,7 +112,6 @@ func (c *cubicSender) OnPacketSent(sentTime time.Time, bytesInFlight protocol.By
 	}
 	c.largestSentPacketNumber = packetNumber
 	c.hybridSlowStart.OnPacketSent(packetNumber)
-	return true
 }
 
 func (c *cubicSender) InRecovery() bool {
@@ -114,18 +123,18 @@ func (c *cubicSender) InSlowStart() bool {
 }
 
 func (c *cubicSender) GetCongestionWindow() protocol.ByteCount {
-	return protocol.ByteCount(c.congestionWindow) * protocol.DefaultTCPMSS
+	return c.congestionWindow
 }
 
 func (c *cubicSender) GetSlowStartThreshold() protocol.ByteCount {
-	return protocol.ByteCount(c.slowstartThreshold) * protocol.DefaultTCPMSS
+	return c.slowstartThreshold
 }
 
 func (c *cubicSender) ExitSlowstart() {
 	c.slowstartThreshold = c.congestionWindow
 }
 
-func (c *cubicSender) SlowstartThreshold() protocol.PacketNumber {
+func (c *cubicSender) SlowstartThreshold() protocol.ByteCount {
 	return c.slowstartThreshold
 }
 
@@ -135,20 +144,29 @@ func (c *cubicSender) MaybeExitSlowStart() {
 	}
 }
 
-func (c *cubicSender) OnPacketAcked(ackedPacketNumber protocol.PacketNumber, ackedBytes protocol.ByteCount, bytesInFlight protocol.ByteCount) {
+func (c *cubicSender) OnPacketAcked(
+	ackedPacketNumber protocol.PacketNumber,
+	ackedBytes protocol.ByteCount,
+	priorInFlight protocol.ByteCount,
+	eventTime time.Time,
+) {
 	c.largestAckedPacketNumber = utils.MaxPacketNumber(ackedPacketNumber, c.largestAckedPacketNumber)
 	if c.InRecovery() {
 		// PRR is used when in recovery.
 		c.prr.OnPacketAcked(ackedBytes)
 		return
 	}
-	c.maybeIncreaseCwnd(ackedPacketNumber, ackedBytes, bytesInFlight)
+	c.maybeIncreaseCwnd(ackedPacketNumber, ackedBytes, priorInFlight, eventTime)
 	if c.InSlowStart() {
 		c.hybridSlowStart.OnPacketAcked(ackedPacketNumber)
 	}
 }
 
-func (c *cubicSender) OnPacketLost(packetNumber protocol.PacketNumber, lostBytes protocol.ByteCount, bytesInFlight protocol.ByteCount) {
+func (c *cubicSender) OnPacketLost(
+	packetNumber protocol.PacketNumber,
+	lostBytes protocol.ByteCount,
+	priorInFlight protocol.ByteCount,
+) {
 	// TCP NewReno (RFC6582) says that once a loss occurs, any losses in packets
 	// already sent should be treated as a single loss event, since it's expected.
 	if packetNumber <= c.largestSentAtLastCutback {
@@ -156,10 +174,8 @@ func (c *cubicSender) OnPacketLost(packetNumber protocol.PacketNumber, lostBytes
 			c.stats.slowstartPacketsLost++
 			c.stats.slowstartBytesLost += lostBytes
 			if c.slowStartLargeReduction {
-				if c.stats.slowstartPacketsLost == 1 || (c.stats.slowstartBytesLost/protocol.DefaultTCPMSS) > (c.stats.slowstartBytesLost-lostBytes)/protocol.DefaultTCPMSS {
-					// Reduce congestion window by 1 for every mss of bytes lost.
-					c.congestionWindow = utils.MaxPacketNumber(c.congestionWindow-1, c.minCongestionWindow)
-				}
+				// Reduce congestion window by lost_bytes for every loss.
+				c.congestionWindow = utils.MaxByteCount(c.congestionWindow-lostBytes, c.minSlowStartExitWindow)
 				c.slowstartThreshold = c.congestionWindow
 			}
 		}
@@ -170,17 +186,19 @@ func (c *cubicSender) OnPacketLost(packetNumber protocol.PacketNumber, lostBytes
 		c.stats.slowstartPacketsLost++
 	}
 
-	c.prr.OnPacketLost(bytesInFlight)
+	c.prr.OnPacketLost(priorInFlight)
 
 	// TODO(chromium): Separate out all of slow start into a separate class.
 	if c.slowStartLargeReduction && c.InSlowStart() {
-		c.congestionWindow = c.congestionWindow - 1
+		if c.congestionWindow >= 2*c.initialCongestionWindow {
+			c.minSlowStartExitWindow = c.congestionWindow / 2
+		}
+		c.congestionWindow = c.congestionWindow - protocol.DefaultTCPMSS
 	} else if c.reno {
-		c.congestionWindow = protocol.PacketNumber(float32(c.congestionWindow) * c.RenoBeta())
+		c.congestionWindow = protocol.ByteCount(float32(c.congestionWindow) * c.RenoBeta())
 	} else {
 		c.congestionWindow = c.cubic.CongestionWindowAfterPacketLoss(c.congestionWindow)
 	}
-	// Enforce a minimum congestion window.
 	if c.congestionWindow < c.minCongestionWindow {
 		c.congestionWindow = c.minCongestionWindow
 	}
@@ -188,7 +206,7 @@ func (c *cubicSender) OnPacketLost(packetNumber protocol.PacketNumber, lostBytes
 	c.largestSentAtLastCutback = c.largestSentPacketNumber
 	// reset packet count from congestion avoidance mode. We start
 	// counting again when we're out of recovery.
-	c.congestionWindowCount = 0
+	c.numAckedPackets = 0
 }
 
 func (c *cubicSender) RenoBeta() float32 {
@@ -201,32 +219,38 @@ func (c *cubicSender) RenoBeta() float32 {
 
 // Called when we receive an ack. Normal TCP tracks how many packets one ack
 // represents, but quic has a separate ack for each packet.
-func (c *cubicSender) maybeIncreaseCwnd(ackedPacketNumber protocol.PacketNumber, ackedBytes protocol.ByteCount, bytesInFlight protocol.ByteCount) {
+func (c *cubicSender) maybeIncreaseCwnd(
+	ackedPacketNumber protocol.PacketNumber,
+	ackedBytes protocol.ByteCount,
+	priorInFlight protocol.ByteCount,
+	eventTime time.Time,
+) {
 	// Do not increase the congestion window unless the sender is close to using
 	// the current window.
-	if !c.isCwndLimited(bytesInFlight) {
+	if !c.isCwndLimited(priorInFlight) {
 		c.cubic.OnApplicationLimited()
 		return
 	}
-	if c.congestionWindow >= c.maxTCPCongestionWindow {
+	if c.congestionWindow >= c.maxCongestionWindow {
 		return
 	}
 	if c.InSlowStart() {
 		// TCP slow start, exponential growth, increase by one for each ACK.
-		c.congestionWindow++
+		c.congestionWindow += protocol.DefaultTCPMSS
 		return
 	}
+	// Congestion avoidance
 	if c.reno {
 		// Classic Reno congestion avoidance.
-		c.congestionWindowCount++
+		c.numAckedPackets++
 		// Divide by num_connections to smoothly increase the CWND at a faster
 		// rate than conventional Reno.
-		if protocol.PacketNumber(c.congestionWindowCount*protocol.ByteCount(c.numConnections)) >= c.congestionWindow {
-			c.congestionWindow++
-			c.congestionWindowCount = 0
+		if c.numAckedPackets*uint64(c.numConnections) >= uint64(c.congestionWindow)/uint64(protocol.DefaultTCPMSS) {
+			c.congestionWindow += protocol.DefaultTCPMSS
+			c.numAckedPackets = 0
 		}
 	} else {
-		c.congestionWindow = utils.MinPacketNumber(c.maxTCPCongestionWindow, c.cubic.CongestionWindowAfterAck(c.congestionWindow, c.rttStats.MinRTT()))
+		c.congestionWindow = utils.MinByteCount(c.maxCongestionWindow, c.cubic.CongestionWindowAfterAck(ackedBytes, c.congestionWindow, c.rttStats.MinRTT(), eventTime))
 	}
 }
 
@@ -282,10 +306,10 @@ func (c *cubicSender) OnConnectionMigration() {
 	c.largestSentAtLastCutback = 0
 	c.lastCutbackExitedSlowstart = false
 	c.cubic.Reset()
-	c.congestionWindowCount = 0
+	c.numAckedPackets = 0
 	c.congestionWindow = c.initialCongestionWindow
 	c.slowstartThreshold = c.initialMaxCongestionWindow
-	c.maxTCPCongestionWindow = c.initialMaxCongestionWindow
+	c.maxCongestionWindow = c.initialMaxCongestionWindow
 }
 
 // SetSlowStartLargeReduction allows enabling the SSLR experiment
diff --git a/internal/congestion/cubic_sender_test.go b/internal/congestion/cubic_sender_test.go
index 4a07a405..de3b74fc 100644
--- a/internal/congestion/cubic_sender_test.go
+++ b/internal/congestion/cubic_sender_test.go
@@ -9,7 +9,7 @@ import (
 	. "github.com/onsi/gomega"
 )
 
-const initialCongestionWindowPackets protocol.PacketNumber = 10
+const initialCongestionWindowPackets = 10
 const defaultWindowTCP = protocol.ByteCount(initialCongestionWindowPackets) * protocol.DefaultTCPMSS
 
 type mockClock time.Time
@@ -22,7 +22,7 @@ func (c *mockClock) Advance(d time.Duration) {
 	*c = mockClock(time.Time(*c).Add(d))
 }
 
-const MaxCongestionWindow = protocol.PacketNumber(200)
+const MaxCongestionWindow protocol.ByteCount = 200 * protocol.DefaultTCPMSS
 
 var _ = Describe("Cubic Sender", func() {
 	var (
@@ -40,13 +40,16 @@ var _ = Describe("Cubic Sender", func() {
 		ackedPacketNumber = 0
 		clock = mockClock{}
 		rttStats = NewRTTStats()
-		sender = NewCubicSender(&clock, rttStats, true /*reno*/, initialCongestionWindowPackets, MaxCongestionWindow)
+		sender = NewCubicSender(&clock, rttStats, true /*reno*/, initialCongestionWindowPackets*protocol.DefaultTCPMSS, MaxCongestionWindow)
 	})
 
+	canSend := func() bool {
+		return bytesInFlight < sender.GetCongestionWindow()
+	}
+
 	SendAvailableSendWindowLen := func(packetLength protocol.ByteCount) int {
-		// Send as long as TimeUntilSend returns InfDuration.
 		packetsSent := 0
-		for bytesInFlight < sender.GetCongestionWindow() {
+		for canSend() {
 			sender.OnPacketSent(clock.Now(), bytesInFlight, packetNumber, packetLength, true)
 			packetNumber++
 			packetsSent++
@@ -56,14 +59,14 @@ var _ = Describe("Cubic Sender", func() {
 	}
 
 	// Normal is that TCP acks every other segment.
-	AckNPacketsLen := func(n int, packetLength protocol.ByteCount) {
+	AckNPackets := func(n int) {
 		rttStats.UpdateRTT(60*time.Millisecond, 0, clock.Now())
 		sender.MaybeExitSlowStart()
 		for i := 0; i < n; i++ {
 			ackedPacketNumber++
-			sender.OnPacketAcked(ackedPacketNumber, packetLength, bytesInFlight)
+			sender.OnPacketAcked(ackedPacketNumber, protocol.DefaultTCPMSS, bytesInFlight, clock.Now())
 		}
-		bytesInFlight -= protocol.ByteCount(n) * packetLength
+		bytesInFlight -= protocol.ByteCount(n) * protocol.DefaultTCPMSS
 		clock.Advance(time.Millisecond)
 	}
 
@@ -82,18 +85,20 @@ var _ = Describe("Cubic Sender", func() {
 	}
 
 	SendAvailableSendWindow := func() int { return SendAvailableSendWindowLen(protocol.DefaultTCPMSS) }
-	AckNPackets := func(n int) { AckNPacketsLen(n, protocol.DefaultTCPMSS) }
 	LoseNPackets := func(n int) { LoseNPacketsLen(n, protocol.DefaultTCPMSS) }
 
 	It("has the right values at startup", func() {
 		// At startup make sure we are at the default.
 		Expect(sender.GetCongestionWindow()).To(Equal(defaultWindowTCP))
-		// At startup make sure we can send.
-		Expect(sender.TimeUntilSend(0)).To(BeZero())
 		// Make sure we can send.
 		Expect(sender.TimeUntilSend(0)).To(BeZero())
+		Expect(canSend()).To(BeTrue())
 		// And that window is un-affected.
 		Expect(sender.GetCongestionWindow()).To(Equal(defaultWindowTCP))
+
+		// Fill the send window with data, then verify that we can't send.
+		SendAvailableSendWindow()
+		Expect(canSend()).To(BeFalse())
 	})
 
 	It("paces", func() {
@@ -384,8 +389,8 @@ var _ = Describe("Cubic Sender", func() {
 		// Expect the window to decrease to the minimum once the RTO fires
 		// and slow start threshold to be set to 1/2 of the CWND.
 		sender.OnRetransmissionTimeout(true)
-		Expect(sender.GetCongestionWindow()).To(Equal(protocol.ByteCount(2 * protocol.DefaultTCPMSS)))
-		Expect(sender.SlowstartThreshold()).To(Equal(protocol.PacketNumber(5)))
+		Expect(sender.GetCongestionWindow()).To(Equal(2 * protocol.DefaultTCPMSS))
+		Expect(sender.SlowstartThreshold()).To(Equal(5 * protocol.DefaultTCPMSS))
 	})
 
 	It("RTO congestion window no retransmission", func() {
@@ -397,73 +402,17 @@ var _ = Describe("Cubic Sender", func() {
 		Expect(sender.GetCongestionWindow()).To(Equal(defaultWindowTCP))
 	})
 
-	It("slow start max send window", func() {
-		const maxCongestionWindowTCP = 50
-		const numberOfAcks = 100
-		sender = NewCubicSender(&clock, rttStats, false, initialCongestionWindowPackets, maxCongestionWindowTCP)
-
-		for i := 0; i < numberOfAcks; i++ {
-			// Send our full send window.
-			SendAvailableSendWindow()
-			AckNPackets(2)
-		}
-		expectedSendWindow := maxCongestionWindowTCP * protocol.DefaultTCPMSS
-		Expect(sender.GetCongestionWindow()).To(Equal(protocol.ByteCount(expectedSendWindow)))
-	})
-
-	It("tcp reno max congestion window", func() {
-		const maxCongestionWindowTCP = 50
-		const numberOfAcks = 1000
-		sender = NewCubicSender(&clock, rttStats, false, initialCongestionWindowPackets, maxCongestionWindowTCP)
-
-		SendAvailableSendWindow()
-		AckNPackets(2)
-		// Make sure we fall out of slow start.
-		LoseNPackets(1)
-
-		for i := 0; i < numberOfAcks; i++ {
-			// Send our full send window.
-			SendAvailableSendWindow()
-			AckNPackets(2)
-		}
-
-		expectedSendWindow := maxCongestionWindowTCP * protocol.DefaultTCPMSS
-		Expect(sender.GetCongestionWindow()).To(Equal(protocol.ByteCount(expectedSendWindow)))
-	})
-
-	It("tcp cubic max congestion window", func() {
-		const maxCongestionWindowTCP = 50
-		// Set to 10000 to compensate for small cubic alpha.
-		const numberOfAcks = 10000
-
-		sender = NewCubicSender(&clock, rttStats, false, initialCongestionWindowPackets, maxCongestionWindowTCP)
-
-		SendAvailableSendWindow()
-		AckNPackets(2)
-		// Make sure we fall out of slow start.
-		LoseNPackets(1)
-
-		for i := 0; i < numberOfAcks; i++ {
-			// Send our full send window.
-			SendAvailableSendWindow()
-			AckNPackets(2)
-		}
-
-		expectedSendWindow := maxCongestionWindowTCP * protocol.DefaultTCPMSS
-		Expect(sender.GetCongestionWindow()).To(Equal(protocol.ByteCount(expectedSendWindow)))
-	})
-
 	It("tcp cubic reset epoch on quiescence", func() {
 		const maxCongestionWindow = 50
 		const maxCongestionWindowBytes = maxCongestionWindow * protocol.DefaultTCPMSS
-		sender = NewCubicSender(&clock, rttStats, false, initialCongestionWindowPackets, maxCongestionWindow)
+		sender = NewCubicSender(&clock, rttStats, false, initialCongestionWindowPackets*protocol.DefaultTCPMSS, maxCongestionWindowBytes)
 
 		numSent := SendAvailableSendWindow()
 
 		// Make sure we fall out of slow start.
-		saveCwnd := sender.GetCongestionWindow()
+		savedCwnd := sender.GetCongestionWindow()
 		LoseNPackets(1)
-		Expect(saveCwnd).To(BeNumerically(">", sender.GetCongestionWindow()))
+		Expect(savedCwnd).To(BeNumerically(">", sender.GetCongestionWindow()))
 
 		// Ack the rest of the outstanding packets to get out of recovery.
 		for i := 1; i < numSent; i++ {
@@ -472,12 +421,12 @@ var _ = Describe("Cubic Sender", func() {
 		Expect(bytesInFlight).To(BeZero())
 
 		// Send a new window of data and ack all; cubic growth should occur.
-		saveCwnd = sender.GetCongestionWindow()
+		savedCwnd = sender.GetCongestionWindow()
 		numSent = SendAvailableSendWindow()
 		for i := 0; i < numSent; i++ {
 			AckNPackets(1)
 		}
-		Expect(saveCwnd).To(BeNumerically("<", sender.GetCongestionWindow()))
+		Expect(savedCwnd).To(BeNumerically("<", sender.GetCongestionWindow()))
 		Expect(maxCongestionWindowBytes).To(BeNumerically(">", sender.GetCongestionWindow()))
 		Expect(bytesInFlight).To(BeZero())
 
@@ -486,50 +435,10 @@ var _ = Describe("Cubic Sender", func() {
 
 		// Send new window of data and ack one packet. Cubic epoch should have
 		// been reset; ensure cwnd increase is not dramatic.
-		saveCwnd = sender.GetCongestionWindow()
+		savedCwnd = sender.GetCongestionWindow()
 		SendAvailableSendWindow()
 		AckNPackets(1)
-		Expect(saveCwnd).To(BeNumerically("~", sender.GetCongestionWindow(), protocol.DefaultTCPMSS))
-		Expect(maxCongestionWindowBytes).To(BeNumerically(">", sender.GetCongestionWindow()))
-	})
-
-	It("tcp cubic shifted epoch on quiescence", func() {
-		const maxCongestionWindow = 50
-		const maxCongestionWindowBytes = maxCongestionWindow * protocol.DefaultTCPMSS
-		sender = NewCubicSender(&clock, rttStats, false, initialCongestionWindowPackets, maxCongestionWindow)
-
-		numSent := SendAvailableSendWindow()
-
-		// Make sure we fall out of slow start.
-		saveCwnd := sender.GetCongestionWindow()
-		LoseNPackets(1)
-		Expect(saveCwnd).To(BeNumerically(">", sender.GetCongestionWindow()))
-
-		// Ack the rest of the outstanding packets to get out of recovery.
-		for i := 1; i < numSent; i++ {
-			AckNPackets(1)
-		}
-		Expect(bytesInFlight).To(BeZero())
-
-		// Send a new window of data and ack all; cubic growth should occur.
-		saveCwnd = sender.GetCongestionWindow()
-		numSent = SendAvailableSendWindow()
-		for i := 0; i < numSent; i++ {
-			AckNPackets(1)
-		}
-		Expect(saveCwnd).To(BeNumerically("<", sender.GetCongestionWindow()))
-		Expect(maxCongestionWindowBytes).To(BeNumerically(">", sender.GetCongestionWindow()))
-		Expect(bytesInFlight).To(BeZero())
-
-		// Quiescent time of 100 seconds
-		clock.Advance(100 * time.Second)
-
-		// Send new window of data and ack one packet. Cubic epoch should have
-		// been reset; ensure cwnd increase is not dramatic.
-		saveCwnd = sender.GetCongestionWindow()
-		SendAvailableSendWindow()
-		AckNPackets(1)
-		Expect(saveCwnd).To(BeNumerically("~", sender.GetCongestionWindow(), protocol.DefaultTCPMSS))
+		Expect(savedCwnd).To(BeNumerically("~", sender.GetCongestionWindow(), protocol.DefaultTCPMSS))
 		Expect(maxCongestionWindowBytes).To(BeNumerically(">", sender.GetCongestionWindow()))
 	})
 
@@ -549,55 +458,6 @@ var _ = Describe("Cubic Sender", func() {
 		Expect(postLossWindow).To(BeNumerically(">", sender.GetCongestionWindow()))
 	})
 
-	It("don't track ack packets", func() {
-		// Send a packet with no retransmittable data, and ensure it's not tracked.
-		Expect(sender.OnPacketSent(clock.Now(), bytesInFlight, packetNumber, protocol.DefaultTCPMSS, false)).To(BeFalse())
-		packetNumber++
-
-		// Send a data packet with retransmittable data, and ensure it is tracked.
-		Expect(sender.OnPacketSent(clock.Now(), bytesInFlight, packetNumber, protocol.DefaultTCPMSS, true)).To(BeTrue())
-	})
-
-	// TEST_F(TcpCubicSenderPacketsTest, ConfigureInitialWindow) {
-	//   QuicConfig config;
-	//
-	//   QuicTagVector options;
-	//   options.push_back(kIW03);
-	//   QuicConfigPeer::SetReceivedConnectionOptions(&config, options);
-	//   sender.SetFromConfig(config, Perspective::IS_SERVER);
-	//   Expect( sender.congestion_window()).To(Equal(3u))
-	//
-	//   options.clear();
-	//   options.push_back(kIW10);
-	//   QuicConfigPeer::SetReceivedConnectionOptions(&config, options);
-	//   sender.SetFromConfig(config, Perspective::IS_SERVER);
-	//   Expect( sender.congestion_window()).To(Equal(10u))
-	//
-	//   options.clear();
-	//   options.push_back(kIW20);
-	//   QuicConfigPeer::SetReceivedConnectionOptions(&config, options);
-	//   sender.SetFromConfig(config, Perspective::IS_SERVER);
-	//   Expect( sender.congestion_window()).To(Equal(20u))
-	//
-	//   options.clear();
-	//   options.push_back(kIW50);
-	//   QuicConfigPeer::SetReceivedConnectionOptions(&config, options);
-	//   sender.SetFromConfig(config, Perspective::IS_SERVER);
-	//   Expect( sender.congestion_window()).To(Equal(50u))
-	// }
-	//
-	// TEST_F(TcpCubicSenderPacketsTest, ConfigureMinimumWindow) {
-	//   QuicConfig config;
-	//
-	//   // Verify that kCOPT: kMIN1 forces the min CWND to 1 packet.
-	//   QuicTagVector options;
-	//   options.push_back(kMIN1);
-	//   QuicConfigPeer::SetReceivedConnectionOptions(&config, options);
-	//   sender.SetFromConfig(config, Perspective::IS_SERVER);
-	//   sender.OnRetransmissionTimeout(true);
-	//   Expect( sender.congestion_window()).To(Equal(1u))
-	// }
-
 	It("2 connection congestion avoidance at end of recovery", func() {
 		sender.SetNumEmulatedConnections(2)
 		// Ack 10 packets in 5 acks to raise the CWND to 20.
@@ -698,67 +558,6 @@ var _ = Describe("Cubic Sender", func() {
 		Expect(sender.GetCongestionWindow()).To(Equal(expectedSendWindow))
 	})
 
-	// TEST_F(TcpCubicSenderPacketsTest, BandwidthResumption) {
-	//   // Test that when provided with CachedNetworkParameters and opted in to the
-	//   // bandwidth resumption experiment, that the TcpCubicSenderPackets sets
-	//   // initial CWND appropriately.
-	//
-	//   // Set some common values.
-	//   CachedNetworkParameters cached_network_params;
-	//   const QuicPacketCount kNumberOfPackets = 123;
-	//   const int kBandwidthEstimateBytesPerSecond =
-	//       kNumberOfPackets * protocol.DefaultTCPMSS;
-	//   cached_network_params.set_bandwidth_estimate_bytes_per_second(
-	//       kBandwidthEstimateBytesPerSecond);
-	//   cached_network_params.set_min_rtt_ms(1000);
-	//
-	//   // Make sure that a bandwidth estimate results in a changed CWND.
-	//   cached_network_params.set_timestamp(clock.WallNow().ToUNIXSeconds() -
-	//                                       (kNumSecondsPerHour - 1));
-	//   sender.ResumeConnectionState(cached_network_params, false);
-	//   Expect( sender.congestion_window()).To(Equal(kNumberOfPackets))
-	//
-	//   // Resumed CWND is limited to be in a sensible range.
-	//   cached_network_params.set_bandwidth_estimate_bytes_per_second(
-	//       (maxCongestionWindow + 1) * protocol.DefaultTCPMSS);
-	//   sender.ResumeConnectionState(cached_network_params, false);
-	//   Expect( sender.congestion_window()).To(Equal(maxCongestionWindow))
-	//
-	//   cached_network_params.set_bandwidth_estimate_bytes_per_second(
-	//       (kMinCongestionWindowForBandwidthResumption - 1) * protocol.DefaultTCPMSS);
-	//   sender.ResumeConnectionState(cached_network_params, false);
-	//   EXPECT_EQ(kMinCongestionWindowForBandwidthResumption,
-	//             sender.congestion_window());
-	//
-	//   // Resume to the max value.
-	//   cached_network_params.set_max_bandwidth_estimate_bytes_per_second(
-	//       (kMinCongestionWindowForBandwidthResumption + 10) * protocol.DefaultTCPMSS);
-	//   sender.ResumeConnectionState(cached_network_params, true);
-	//   EXPECT_EQ((kMinCongestionWindowForBandwidthResumption + 10) * protocol.DefaultTCPMSS,
-	//             sender.GetCongestionWindow());
-	// }
-	//
-	// TEST_F(TcpCubicSenderPacketsTest, PaceBelowCWND) {
-	//   QuicConfig config;
-	//
-	//   // Verify that kCOPT: kMIN4 forces the min CWND to 1 packet, but allows up
-	//   // to 4 to be sent.
-	//   QuicTagVector options;
-	//   options.push_back(kMIN4);
-	//   QuicConfigPeer::SetReceivedConnectionOptions(&config, options);
-	//   sender.SetFromConfig(config, Perspective::IS_SERVER);
-	//   sender.OnRetransmissionTimeout(true);
-	//   Expect( sender.congestion_window()).To(Equal(1u))
-	//   EXPECT_TRUE(
-	//       sender.TimeUntilSend(QuicTime::Zero(), protocol.DefaultTCPMSS).IsZero());
-	//   EXPECT_TRUE(
-	//       sender.TimeUntilSend(QuicTime::Zero(), 2 * protocol.DefaultTCPMSS).IsZero());
-	//   EXPECT_TRUE(
-	//       sender.TimeUntilSend(QuicTime::Zero(), 3 * protocol.DefaultTCPMSS).IsZero());
-	//   EXPECT_FALSE(
-	//       sender.TimeUntilSend(QuicTime::Zero(), 4 * protocol.DefaultTCPMSS).IsZero());
-	// }
-
 	It("reset after connection migration", func() {
 		Expect(sender.GetCongestionWindow()).To(Equal(defaultWindowTCP))
 		Expect(sender.SlowstartThreshold()).To(Equal(MaxCongestionWindow))
@@ -782,7 +581,7 @@ var _ = Describe("Cubic Sender", func() {
 		// start threshold is also updated.
 		expectedSendWindow = protocol.ByteCount(float32(expectedSendWindow) * renoBeta)
 		Expect(sender.GetCongestionWindow()).To(Equal(expectedSendWindow))
-		Expect(sender.SlowstartThreshold()).To(Equal(protocol.PacketNumber(expectedSendWindow / protocol.DefaultTCPMSS)))
+		Expect(sender.SlowstartThreshold()).To(Equal(expectedSendWindow))
 
 		// Resets cwnd and slow start threshold on connection migrations.
 		sender.OnConnectionMigration()
@@ -790,4 +589,52 @@ var _ = Describe("Cubic Sender", func() {
 		Expect(sender.SlowstartThreshold()).To(Equal(MaxCongestionWindow))
 		Expect(sender.HybridSlowStart().Started()).To(BeFalse())
 	})
+
+	It("default max cwnd", func() {
+		sender = NewCubicSender(&clock, rttStats, true /*reno*/, initialCongestionWindowPackets*protocol.DefaultTCPMSS, protocol.DefaultMaxCongestionWindow)
+
+		defaultMaxCongestionWindowPackets := protocol.DefaultMaxCongestionWindow / protocol.DefaultTCPMSS
+		for i := 1; i < int(defaultMaxCongestionWindowPackets); i++ {
+			sender.MaybeExitSlowStart()
+			sender.OnPacketAcked(protocol.PacketNumber(i), 1350, sender.GetCongestionWindow(), clock.Now())
+		}
+		Expect(sender.GetCongestionWindow()).To(Equal(protocol.DefaultMaxCongestionWindow))
+	})
+
+	It("limit cwnd increase in congestion avoidance", func() {
+		// Enable Cubic.
+		sender = NewCubicSender(&clock, rttStats, false, initialCongestionWindowPackets*protocol.DefaultTCPMSS, MaxCongestionWindow)
+		numSent := SendAvailableSendWindow()
+
+		// Make sure we fall out of slow start.
+		savedCwnd := sender.GetCongestionWindow()
+		LoseNPackets(1)
+		Expect(savedCwnd).To(BeNumerically(">", sender.GetCongestionWindow()))
+
+		// Ack the rest of the outstanding packets to get out of recovery.
+		for i := 1; i < numSent; i++ {
+			AckNPackets(1)
+		}
+		Expect(bytesInFlight).To(BeZero())
+
+		savedCwnd = sender.GetCongestionWindow()
+		SendAvailableSendWindow()
+
+		// Ack packets until the CWND increases.
+		for sender.GetCongestionWindow() == savedCwnd {
+			AckNPackets(1)
+			SendAvailableSendWindow()
+		}
+		// Bytes in flight may be larger than the CWND if the CWND isn't an exact
+		// multiple of the packet sizes being sent.
+		Expect(bytesInFlight).To(BeNumerically(">=", sender.GetCongestionWindow()))
+		savedCwnd = sender.GetCongestionWindow()
+
+		// Advance time 2 seconds waiting for an ack.
+		clock.Advance(2 * time.Second)
+
+		// Ack two packets.  The CWND should increase by only one packet.
+		AckNPackets(2)
+		Expect(sender.GetCongestionWindow()).To(Equal(savedCwnd + protocol.DefaultTCPMSS))
+	})
 })
diff --git a/internal/congestion/cubic_test.go b/internal/congestion/cubic_test.go
index 8372df0e..52cae145 100644
--- a/internal/congestion/cubic_test.go
+++ b/internal/congestion/cubic_test.go
@@ -11,7 +11,9 @@ import (
 
 const numConnections uint32 = 2
 const nConnectionBeta float32 = (float32(numConnections) - 1 + beta) / float32(numConnections)
+const nConnectionBetaLastMax float32 = (float32(numConnections) - 1 + betaLastMax) / float32(numConnections)
 const nConnectionAlpha float32 = 3 * float32(numConnections) * float32(numConnections) * (1 - nConnectionBeta) / (1 + nConnectionBeta)
+const maxCubicTimeInterval = 30 * time.Millisecond
 
 var _ = Describe("Cubic", func() {
 	var (
@@ -24,88 +26,211 @@ var _ = Describe("Cubic", func() {
 		cubic = NewCubic(&clock)
 	})
 
-	It("works above origin", func() {
+	renoCwnd := func(currentCwnd protocol.ByteCount) protocol.ByteCount {
+		return currentCwnd + protocol.ByteCount(float32(protocol.DefaultTCPMSS)*nConnectionAlpha*float32(protocol.DefaultTCPMSS)/float32(currentCwnd))
+	}
+
+	cubicConvexCwnd := func(initialCwnd protocol.ByteCount, rtt, elapsedTime time.Duration) protocol.ByteCount {
+		offset := protocol.ByteCount((elapsedTime+rtt)/time.Microsecond) << 10 / 1000000
+		deltaCongestionWindow := 410 * offset * offset * offset * protocol.DefaultTCPMSS >> 40
+		return initialCwnd + deltaCongestionWindow
+	}
+
+	It("works above origin (with tighter bounds)", func() {
 		// Convex growth.
 		const rttMin = 100 * time.Millisecond
 		const rttMinS = float32(rttMin/time.Millisecond) / 1000.0
-		currentCwnd := protocol.PacketNumber(10)
-		// Without the signed-integer, cubic-convex fix, we mistakenly
-		// increment cwnd after only one_ms_ and a single ack.
-		expectedCwnd := currentCwnd
-		// Initialize the state.
+		currentCwnd := 10 * protocol.DefaultTCPMSS
+		initialCwnd := currentCwnd
+
 		clock.Advance(time.Millisecond)
 		initialTime := clock.Now()
-		currentCwnd = cubic.CongestionWindowAfterAck(currentCwnd, rttMin)
-		Expect(currentCwnd).To(Equal(expectedCwnd))
-		currentCwnd = expectedCwnd
-		initialCwnd := currentCwnd
+		expectedFirstCwnd := renoCwnd(currentCwnd)
+		currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, initialTime)
+		Expect(expectedFirstCwnd).To(Equal(currentCwnd))
+
 		// Normal TCP phase.
 		// The maximum number of expected reno RTTs can be calculated by
 		// finding the point where the cubic curve and the reno curve meet.
-		maxRenoRtts := int(math.Sqrt(float64(nConnectionAlpha/(0.4*rttMinS*rttMinS*rttMinS))) - 1)
+		maxRenoRtts := int(math.Sqrt(float64(nConnectionAlpha/(0.4*rttMinS*rttMinS*rttMinS))) - 2)
 		for i := 0; i < maxRenoRtts; i++ {
-			maxPerAckCwnd := currentCwnd
-			for n := uint64(1); n < uint64(float32(maxPerAckCwnd)/nConnectionAlpha); n++ {
+			// Alternatively, we expect it to increase by one, every time we
+			// receive current_cwnd/Alpha acks back.  (This is another way of
+			// saying we expect cwnd to increase by approximately Alpha once
+			// we receive current_cwnd number ofacks back).
+			numAcksThisEpoch := int(float32(currentCwnd/protocol.DefaultTCPMSS) / nConnectionAlpha)
+
+			initialCwndThisEpoch := currentCwnd
+			for n := 0; n < numAcksThisEpoch; n++ {
 				// Call once per ACK.
-				nextCwnd := cubic.CongestionWindowAfterAck(currentCwnd, rttMin)
-				Expect(nextCwnd).To(Equal(currentCwnd))
+				expectedNextCwnd := renoCwnd(currentCwnd)
+				currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
+				Expect(currentCwnd).To(Equal(expectedNextCwnd))
 			}
+			// Our byte-wise Reno implementation is an estimate.  We expect
+			// the cwnd to increase by approximately one MSS every
+			// cwnd/kDefaultTCPMSS/Alpha acks, but it may be off by as much as
+			// half a packet for smaller values of current_cwnd.
+			cwndChangeThisEpoch := currentCwnd - initialCwndThisEpoch
+			Expect(cwndChangeThisEpoch).To(BeNumerically("~", protocol.DefaultTCPMSS, protocol.DefaultTCPMSS/2))
 			clock.Advance(100 * time.Millisecond)
-			currentCwnd = cubic.CongestionWindowAfterAck(currentCwnd, rttMin)
-			// When we fix convex mode and the uint64 arithmetic, we
-			// increase the expected_cwnd only after after the first 100ms,
-			// rather than after the initial 1ms.
-			expectedCwnd++
-			Expect(currentCwnd).To(Equal(expectedCwnd))
 		}
-		// Cubic phase.
-		for i := 0; i < 52; i++ {
-			for n := protocol.PacketNumber(1); n < currentCwnd; n++ {
-				// Call once per ACK.
-				Expect(cubic.CongestionWindowAfterAck(currentCwnd, rttMin)).To(Equal(currentCwnd))
+
+		for i := 0; i < 54; i++ {
+			maxAcksThisEpoch := currentCwnd / protocol.DefaultTCPMSS
+			interval := time.Duration(100*1000/maxAcksThisEpoch) * time.Microsecond
+			for n := 0; n < int(maxAcksThisEpoch); n++ {
+				clock.Advance(interval)
+				currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
+				expectedCwnd := cubicConvexCwnd(initialCwnd, rttMin, clock.Now().Sub(initialTime))
+				// If we allow per-ack updates, every update is a small cubic update.
+				Expect(currentCwnd).To(Equal(expectedCwnd))
 			}
-			clock.Advance(100 * time.Millisecond)
-			currentCwnd = cubic.CongestionWindowAfterAck(currentCwnd, rttMin)
 		}
-		// Total time elapsed so far; add min_rtt (0.1s) here as well.
-		elapsedTimeS := float32(clock.Now().Sub(initialTime)+rttMin) / float32(time.Second)
-		// |expected_cwnd| is initial value of cwnd + K * t^3, where K = 0.4.
-		expectedCwnd = initialCwnd + protocol.PacketNumber((elapsedTimeS*elapsedTimeS*elapsedTimeS*410)/1024)
+		expectedCwnd := cubicConvexCwnd(initialCwnd, rttMin, clock.Now().Sub(initialTime))
+		currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
 		Expect(currentCwnd).To(Equal(expectedCwnd))
 	})
 
-	It("manages loss events", func() {
+	It("works above the origin with fine grained cubing", func() {
+		// Start the test with an artificially large cwnd to prevent Reno
+		// from over-taking cubic.
+		currentCwnd := 1000 * protocol.DefaultTCPMSS
+		initialCwnd := currentCwnd
 		rttMin := 100 * time.Millisecond
-		currentCwnd := protocol.PacketNumber(422)
-		expectedCwnd := currentCwnd
+		clock.Advance(time.Millisecond)
+		initialTime := clock.Now()
+
+		currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
+		clock.Advance(600 * time.Millisecond)
+		currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
+
+		// We expect the algorithm to perform only non-zero, fine-grained cubic
+		// increases on every ack in this case.
+		for i := 0; i < 100; i++ {
+			clock.Advance(10 * time.Millisecond)
+			expectedCwnd := cubicConvexCwnd(initialCwnd, rttMin, clock.Now().Sub(initialTime))
+			nextCwnd := cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
+			// Make sure we are performing cubic increases.
+			Expect(nextCwnd).To(Equal(expectedCwnd))
+			// Make sure that these are non-zero, less-than-packet sized increases.
+			Expect(nextCwnd).To(BeNumerically(">", currentCwnd))
+			cwndDelta := nextCwnd - currentCwnd
+			Expect(protocol.DefaultTCPMSS / 10).To(BeNumerically(">", cwndDelta))
+			currentCwnd = nextCwnd
+		}
+	})
+
+	It("handles per ack updates", func() {
+		// Start the test with a large cwnd and RTT, to force the first
+		// increase to be a cubic increase.
+		initialCwndPackets := 150
+		currentCwnd := protocol.ByteCount(initialCwndPackets) * protocol.DefaultTCPMSS
+		rttMin := 350 * time.Millisecond
+
+		// Initialize the epoch
+		clock.Advance(time.Millisecond)
+		// Keep track of the growth of the reno-equivalent cwnd.
+		rCwnd := renoCwnd(currentCwnd)
+		currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
+		initialCwnd := currentCwnd
+
+		// Simulate the return of cwnd packets in less than
+		// MaxCubicInterval() time.
+		maxAcks := int(float32(initialCwndPackets) / nConnectionAlpha)
+		interval := maxCubicTimeInterval / time.Duration(maxAcks+1)
+
+		// In this scenario, the first increase is dictated by the cubic
+		// equation, but it is less than one byte, so the cwnd doesn't
+		// change.  Normally, without per-ack increases, any cwnd plateau
+		// will cause the cwnd to be pinned for MaxCubicTimeInterval().  If
+		// we enable per-ack updates, the cwnd will continue to grow,
+		// regardless of the temporary plateau.
+		clock.Advance(interval)
+		rCwnd = renoCwnd(rCwnd)
+		Expect(cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())).To(Equal(currentCwnd))
+		for i := 1; i < maxAcks; i++ {
+			clock.Advance(interval)
+			nextCwnd := cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
+			rCwnd = renoCwnd(rCwnd)
+			// The window shoud increase on every ack.
+			Expect(nextCwnd).To(BeNumerically(">", currentCwnd))
+			Expect(nextCwnd).To(Equal(rCwnd))
+			currentCwnd = nextCwnd
+		}
+
+		// After all the acks are returned from the epoch, we expect the
+		// cwnd to have increased by nearly one packet.  (Not exactly one
+		// packet, because our byte-wise Reno algorithm is always a slight
+		// under-estimation).  Without per-ack updates, the current_cwnd
+		// would otherwise be unchanged.
+		minimumExpectedIncrease := protocol.DefaultTCPMSS * 9 / 10
+		Expect(currentCwnd).To(BeNumerically(">", initialCwnd+minimumExpectedIncrease))
+	})
+
+	It("handles loss events", func() {
+		rttMin := 100 * time.Millisecond
+		currentCwnd := 422 * protocol.DefaultTCPMSS
+		expectedCwnd := renoCwnd(currentCwnd)
 		// Initialize the state.
 		clock.Advance(time.Millisecond)
-		Expect(cubic.CongestionWindowAfterAck(currentCwnd, rttMin)).To(Equal(expectedCwnd))
-		expectedCwnd = protocol.PacketNumber(float32(currentCwnd) * nConnectionBeta)
+		Expect(cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())).To(Equal(expectedCwnd))
+
+		// On the first loss, the last max congestion window is set to the
+		// congestion window before the loss.
+		preLossCwnd := currentCwnd
+		Expect(cubic.lastMaxCongestionWindow).To(BeZero())
+		expectedCwnd = protocol.ByteCount(float32(currentCwnd) * nConnectionBeta)
 		Expect(cubic.CongestionWindowAfterPacketLoss(currentCwnd)).To(Equal(expectedCwnd))
-		expectedCwnd = protocol.PacketNumber(float32(currentCwnd) * nConnectionBeta)
+		Expect(cubic.lastMaxCongestionWindow).To(Equal(preLossCwnd))
+		currentCwnd = expectedCwnd
+
+		// On the second loss, the current congestion window has not yet
+		// reached the last max congestion window.  The last max congestion
+		// window will be reduced by an additional backoff factor to allow
+		// for competition.
+		preLossCwnd = currentCwnd
+		expectedCwnd = protocol.ByteCount(float32(currentCwnd) * nConnectionBeta)
 		Expect(cubic.CongestionWindowAfterPacketLoss(currentCwnd)).To(Equal(expectedCwnd))
+		currentCwnd = expectedCwnd
+		Expect(preLossCwnd).To(BeNumerically(">", cubic.lastMaxCongestionWindow))
+		expectedLastMax := protocol.ByteCount(float32(preLossCwnd) * nConnectionBetaLastMax)
+		Expect(cubic.lastMaxCongestionWindow).To(Equal(expectedLastMax))
+		Expect(expectedCwnd).To(BeNumerically("<", cubic.lastMaxCongestionWindow))
+		// Simulate an increase, and check that we are below the origin.
+		currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
+		Expect(cubic.lastMaxCongestionWindow).To(BeNumerically(">", currentCwnd))
+
+		// On the final loss, simulate the condition where the congestion
+		// window had a chance to grow nearly to the last congestion window.
+		currentCwnd = cubic.lastMaxCongestionWindow - 1
+		preLossCwnd = currentCwnd
+		expectedCwnd = protocol.ByteCount(float32(currentCwnd) * nConnectionBeta)
+		Expect(cubic.CongestionWindowAfterPacketLoss(currentCwnd)).To(Equal(expectedCwnd))
+		expectedLastMax = preLossCwnd
+		Expect(cubic.lastMaxCongestionWindow).To(Equal(expectedLastMax))
 	})
 
 	It("works below origin", func() {
 		// Concave growth.
 		rttMin := 100 * time.Millisecond
-		currentCwnd := protocol.PacketNumber(422)
-		expectedCwnd := currentCwnd
+		currentCwnd := 422 * protocol.DefaultTCPMSS
+		expectedCwnd := renoCwnd(currentCwnd)
 		// Initialize the state.
 		clock.Advance(time.Millisecond)
-		Expect(cubic.CongestionWindowAfterAck(currentCwnd, rttMin)).To(Equal(expectedCwnd))
-		expectedCwnd = protocol.PacketNumber(float32(currentCwnd) * nConnectionBeta)
+		Expect(cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())).To(Equal(expectedCwnd))
+
+		expectedCwnd = protocol.ByteCount(float32(currentCwnd) * nConnectionBeta)
 		Expect(cubic.CongestionWindowAfterPacketLoss(currentCwnd)).To(Equal(expectedCwnd))
 		currentCwnd = expectedCwnd
 		// First update after loss to initialize the epoch.
-		currentCwnd = cubic.CongestionWindowAfterAck(currentCwnd, rttMin)
+		currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
 		// Cubic phase.
 		for i := 0; i < 40; i++ {
 			clock.Advance(100 * time.Millisecond)
-			currentCwnd = cubic.CongestionWindowAfterAck(currentCwnd, rttMin)
+			currentCwnd = cubic.CongestionWindowAfterAck(protocol.DefaultTCPMSS, currentCwnd, rttMin, clock.Now())
 		}
-		expectedCwnd = 422
+		expectedCwnd = 553632
 		Expect(currentCwnd).To(Equal(expectedCwnd))
 	})
 })
diff --git a/internal/congestion/interface.go b/internal/congestion/interface.go
index 28950dd0..7c27da64 100644
--- a/internal/congestion/interface.go
+++ b/internal/congestion/interface.go
@@ -9,11 +9,11 @@ import (
 // A SendAlgorithm performs congestion control and calculates the congestion window
 type SendAlgorithm interface {
 	TimeUntilSend(bytesInFlight protocol.ByteCount) time.Duration
-	OnPacketSent(sentTime time.Time, bytesInFlight protocol.ByteCount, packetNumber protocol.PacketNumber, bytes protocol.ByteCount, isRetransmittable bool) bool
+	OnPacketSent(sentTime time.Time, bytesInFlight protocol.ByteCount, packetNumber protocol.PacketNumber, bytes protocol.ByteCount, isRetransmittable bool)
 	GetCongestionWindow() protocol.ByteCount
 	MaybeExitSlowStart()
-	OnPacketAcked(number protocol.PacketNumber, ackedBytes protocol.ByteCount, bytesInFlight protocol.ByteCount)
-	OnPacketLost(number protocol.PacketNumber, lostBytes protocol.ByteCount, bytesInFlight protocol.ByteCount)
+	OnPacketAcked(number protocol.PacketNumber, ackedBytes protocol.ByteCount, priorInFlight protocol.ByteCount, eventTime time.Time)
+	OnPacketLost(number protocol.PacketNumber, lostBytes protocol.ByteCount, priorInFlight protocol.ByteCount)
 	SetNumEmulatedConnections(n int)
 	OnRetransmissionTimeout(packetsRetransmitted bool)
 	OnConnectionMigration()
@@ -30,7 +30,7 @@ type SendAlgorithmWithDebugInfo interface {
 	// Stuff only used in testing
 
 	HybridSlowStart() *HybridSlowStart
-	SlowstartThreshold() protocol.PacketNumber
+	SlowstartThreshold() protocol.ByteCount
 	RenoBeta() float32
 	InRecovery() bool
 }
diff --git a/internal/congestion/prr_sender.go b/internal/congestion/prr_sender.go
index 18a3736a..5c807d19 100644
--- a/internal/congestion/prr_sender.go
+++ b/internal/congestion/prr_sender.go
@@ -1,10 +1,7 @@
 package congestion
 
 import (
-	"time"
-
 	"github.com/lucas-clemente/quic-go/internal/protocol"
-	"github.com/lucas-clemente/quic-go/internal/utils"
 )
 
 // PrrSender implements the Proportional Rate Reduction (PRR) per RFC 6937
@@ -23,9 +20,9 @@ func (p *PrrSender) OnPacketSent(sentBytes protocol.ByteCount) {
 // OnPacketLost should be called on the first loss that triggers a recovery
 // period and all other methods in this class should only be called when in
 // recovery.
-func (p *PrrSender) OnPacketLost(bytesInFlight protocol.ByteCount) {
+func (p *PrrSender) OnPacketLost(priorInFlight protocol.ByteCount) {
 	p.bytesSentSinceLoss = 0
-	p.bytesInFlightBeforeLoss = bytesInFlight
+	p.bytesInFlightBeforeLoss = priorInFlight
 	p.bytesDeliveredSinceLoss = 0
 	p.ackCountSinceLoss = 0
 }
@@ -36,28 +33,22 @@ func (p *PrrSender) OnPacketAcked(ackedBytes protocol.ByteCount) {
 	p.ackCountSinceLoss++
 }
 
-// TimeUntilSend calculates the time until a packet can be sent
-func (p *PrrSender) TimeUntilSend(congestionWindow, bytesInFlight, slowstartThreshold protocol.ByteCount) time.Duration {
+// CanSend returns if packets can be sent
+func (p *PrrSender) CanSend(congestionWindow, bytesInFlight, slowstartThreshold protocol.ByteCount) bool {
 	// Return QuicTime::Zero In order to ensure limited transmit always works.
 	if p.bytesSentSinceLoss == 0 || bytesInFlight < protocol.DefaultTCPMSS {
-		return 0
+		return true
 	}
 	if congestionWindow > bytesInFlight {
 		// During PRR-SSRB, limit outgoing packets to 1 extra MSS per ack, instead
 		// of sending the entire available window. This prevents burst retransmits
 		// when more packets are lost than the CWND reduction.
 		//   limit = MAX(prr_delivered - prr_out, DeliveredData) + MSS
-		if p.bytesDeliveredSinceLoss+p.ackCountSinceLoss*protocol.DefaultTCPMSS <= p.bytesSentSinceLoss {
-			return utils.InfDuration
-		}
-		return 0
+		return p.bytesDeliveredSinceLoss+p.ackCountSinceLoss*protocol.DefaultTCPMSS > p.bytesSentSinceLoss
 	}
 	// Implement Proportional Rate Reduction (RFC6937).
 	// Checks a simplified version of the PRR formula that doesn't use division:
 	// AvailableSendWindow =
 	//   CEIL(prr_delivered * ssthresh / BytesInFlightAtLoss) - prr_sent
-	if p.bytesDeliveredSinceLoss*slowstartThreshold > p.bytesSentSinceLoss*p.bytesInFlightBeforeLoss {
-		return 0
-	}
-	return utils.InfDuration
+	return p.bytesDeliveredSinceLoss*slowstartThreshold > p.bytesSentSinceLoss*p.bytesInFlightBeforeLoss
 }
diff --git a/internal/congestion/prr_sender_test.go b/internal/congestion/prr_sender_test.go
index 52f41714..0786b9bd 100644
--- a/internal/congestion/prr_sender_test.go
+++ b/internal/congestion/prr_sender_test.go
@@ -5,7 +5,6 @@ import (
 	. "github.com/onsi/gomega"
 
 	"github.com/lucas-clemente/quic-go/internal/protocol"
-	"github.com/lucas-clemente/quic-go/internal/utils"
 )
 
 var _ = Describe("PRR sender", func() {
@@ -27,11 +26,11 @@ var _ = Describe("PRR sender", func() {
 		// Ack a packet. PRR allows one packet to leave immediately.
 		prr.OnPacketAcked(protocol.DefaultTCPMSS)
 		bytesInFlight -= protocol.DefaultTCPMSS
-		Expect(prr.TimeUntilSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeZero())
+		Expect(prr.CanSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeTrue())
 		// Send retransmission.
 		prr.OnPacketSent(protocol.DefaultTCPMSS)
 		// PRR shouldn't allow sending any more packets.
-		Expect(prr.TimeUntilSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(Equal(utils.InfDuration))
+		Expect(prr.CanSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeFalse())
 
 		// One packet is lost, and one ack was consumed above. PRR now paces
 		// transmissions through the remaining 48 acks. PRR will alternatively
@@ -40,11 +39,11 @@ var _ = Describe("PRR sender", func() {
 			// Ack a packet. PRR shouldn't allow sending a packet in response.
 			prr.OnPacketAcked(protocol.DefaultTCPMSS)
 			bytesInFlight -= protocol.DefaultTCPMSS
-			Expect(prr.TimeUntilSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(Equal(utils.InfDuration))
+			Expect(prr.CanSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeFalse())
 			// Ack another packet. PRR should now allow sending a packet in response.
 			prr.OnPacketAcked(protocol.DefaultTCPMSS)
 			bytesInFlight -= protocol.DefaultTCPMSS
-			Expect(prr.TimeUntilSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeZero())
+			Expect(prr.CanSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeTrue())
 			// Send a packet in response.
 			prr.OnPacketSent(protocol.DefaultTCPMSS)
 			bytesInFlight += protocol.DefaultTCPMSS
@@ -57,7 +56,7 @@ var _ = Describe("PRR sender", func() {
 			// Ack a packet.
 			prr.OnPacketAcked(protocol.DefaultTCPMSS)
 			bytesInFlight -= protocol.DefaultTCPMSS
-			Expect(prr.TimeUntilSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeZero())
+			Expect(prr.CanSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeTrue())
 			// Send a packet in response, since PRR allows it.
 			prr.OnPacketSent(protocol.DefaultTCPMSS)
 			bytesInFlight += protocol.DefaultTCPMSS
@@ -65,7 +64,7 @@ var _ = Describe("PRR sender", func() {
 			// Since bytes_in_flight is equal to the congestion_window,
 			// PRR disallows sending.
 			Expect(bytesInFlight).To(Equal(congestionWindow))
-			Expect(prr.TimeUntilSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(Equal(utils.InfDuration))
+			Expect(prr.CanSend(congestionWindow, bytesInFlight, sshthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeFalse())
 		}
 
 	})
@@ -86,20 +85,20 @@ var _ = Describe("PRR sender", func() {
 			bytesInFlight -= protocol.DefaultTCPMSS
 			// PRR-SSRB should allow two packets to be sent.
 			for j := 0; j < 2; j++ {
-				Expect(prr.TimeUntilSend(congestionWindow, bytesInFlight, ssthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeZero())
+				Expect(prr.CanSend(congestionWindow, bytesInFlight, ssthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeTrue())
 				// Send a packet in response.
 				prr.OnPacketSent(protocol.DefaultTCPMSS)
 				bytesInFlight += protocol.DefaultTCPMSS
 			}
 			// PRR should allow no more than 2 packets in response to an ack.
-			Expect(prr.TimeUntilSend(congestionWindow, bytesInFlight, ssthreshAfterLoss*protocol.DefaultTCPMSS)).To(Equal(utils.InfDuration))
+			Expect(prr.CanSend(congestionWindow, bytesInFlight, ssthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeFalse())
 		}
 
 		// Out of SSRB mode, PRR allows one send in response to each ack.
 		for i := 0; i < 10; i++ {
 			prr.OnPacketAcked(protocol.DefaultTCPMSS)
 			bytesInFlight -= protocol.DefaultTCPMSS
-			Expect(prr.TimeUntilSend(congestionWindow, bytesInFlight, ssthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeZero())
+			Expect(prr.CanSend(congestionWindow, bytesInFlight, ssthreshAfterLoss*protocol.DefaultTCPMSS)).To(BeTrue())
 			// Send a packet in response.
 			prr.OnPacketSent(protocol.DefaultTCPMSS)
 			bytesInFlight += protocol.DefaultTCPMSS
diff --git a/internal/mocks/congestion.go b/internal/mocks/congestion.go
index b86ccfea..d0749252 100644
--- a/internal/mocks/congestion.go
+++ b/internal/mocks/congestion.go
@@ -68,13 +68,13 @@ func (mr *MockSendAlgorithmMockRecorder) OnConnectionMigration() *gomock.Call {
 }
 
 // OnPacketAcked mocks base method
-func (m *MockSendAlgorithm) OnPacketAcked(arg0 protocol.PacketNumber, arg1, arg2 protocol.ByteCount) {
-	m.ctrl.Call(m, "OnPacketAcked", arg0, arg1, arg2)
+func (m *MockSendAlgorithm) OnPacketAcked(arg0 protocol.PacketNumber, arg1, arg2 protocol.ByteCount, arg3 time.Time) {
+	m.ctrl.Call(m, "OnPacketAcked", arg0, arg1, arg2, arg3)
 }
 
 // OnPacketAcked indicates an expected call of OnPacketAcked
-func (mr *MockSendAlgorithmMockRecorder) OnPacketAcked(arg0, arg1, arg2 interface{}) *gomock.Call {
-	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "OnPacketAcked", reflect.TypeOf((*MockSendAlgorithm)(nil).OnPacketAcked), arg0, arg1, arg2)
+func (mr *MockSendAlgorithmMockRecorder) OnPacketAcked(arg0, arg1, arg2, arg3 interface{}) *gomock.Call {
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "OnPacketAcked", reflect.TypeOf((*MockSendAlgorithm)(nil).OnPacketAcked), arg0, arg1, arg2, arg3)
 }
 
 // OnPacketLost mocks base method
@@ -88,10 +88,8 @@ func (mr *MockSendAlgorithmMockRecorder) OnPacketLost(arg0, arg1, arg2 interface
 }
 
 // OnPacketSent mocks base method
-func (m *MockSendAlgorithm) OnPacketSent(arg0 time.Time, arg1 protocol.ByteCount, arg2 protocol.PacketNumber, arg3 protocol.ByteCount, arg4 bool) bool {
-	ret := m.ctrl.Call(m, "OnPacketSent", arg0, arg1, arg2, arg3, arg4)
-	ret0, _ := ret[0].(bool)
-	return ret0
+func (m *MockSendAlgorithm) OnPacketSent(arg0 time.Time, arg1 protocol.ByteCount, arg2 protocol.PacketNumber, arg3 protocol.ByteCount, arg4 bool) {
+	m.ctrl.Call(m, "OnPacketSent", arg0, arg1, arg2, arg3, arg4)
 }
 
 // OnPacketSent indicates an expected call of OnPacketSent
diff --git a/internal/protocol/server_parameters.go b/internal/protocol/server_parameters.go
index 46b69ead..961986a2 100644
--- a/internal/protocol/server_parameters.go
+++ b/internal/protocol/server_parameters.go
@@ -12,11 +12,13 @@ const MaxPacketSizeIPv6 = 1232
 // This makes sure that those packets can always be retransmitted without splitting the contained StreamFrames
 const NonForwardSecurePacketSizeReduction = 50
 
+const defaultMaxCongestionWindowPackets = 1000
+
 // DefaultMaxCongestionWindow is the default for the max congestion window
-const DefaultMaxCongestionWindow = 1000
+const DefaultMaxCongestionWindow ByteCount = defaultMaxCongestionWindowPackets * DefaultTCPMSS
 
 // InitialCongestionWindow is the initial congestion window in QUIC packets
-const InitialCongestionWindow = 32
+const InitialCongestionWindow ByteCount = 32 * DefaultTCPMSS
 
 // MaxUndecryptablePackets limits the number of undecryptable packets that a
 // session queues for later until it sends a public reset.
@@ -70,7 +72,7 @@ const MaxStreamsMultiplier = 1.1
 const MaxStreamsMinimumIncrement = 10
 
 // MaxSessionUnprocessedPackets is the max number of packets stored in each session that are not yet processed.
-const MaxSessionUnprocessedPackets = DefaultMaxCongestionWindow
+const MaxSessionUnprocessedPackets = defaultMaxCongestionWindowPackets
 
 // SkipPacketAveragePeriodLength is the average period length in which one packet number is skipped to prevent an Optimistic ACK attack
 const SkipPacketAveragePeriodLength PacketNumber = 500
@@ -84,7 +86,7 @@ const CookieExpiryTime = 24 * time.Hour
 // MaxOutstandingSentPackets is maximum number of packets saved for retransmission.
 // When reached, it imposes a soft limit on sending new packets:
 // Sending ACKs and retransmission is still allowed, but now new regular packets can be sent.
-const MaxOutstandingSentPackets = 2 * DefaultMaxCongestionWindow
+const MaxOutstandingSentPackets = 2 * defaultMaxCongestionWindowPackets
 
 // MaxTrackedSentPackets is maximum number of sent packets saved for retransmission.
 // When reached, no more packets will be sent.
@@ -92,7 +94,7 @@ const MaxOutstandingSentPackets = 2 * DefaultMaxCongestionWindow
 const MaxTrackedSentPackets = MaxOutstandingSentPackets * 5 / 4
 
 // MaxTrackedReceivedAckRanges is the maximum number of ACK ranges tracked
-const MaxTrackedReceivedAckRanges = DefaultMaxCongestionWindow
+const MaxTrackedReceivedAckRanges = defaultMaxCongestionWindowPackets
 
 // MaxNonRetransmittableAcks is the maximum number of packets containing an ACK, but no retransmittable frames, that we send in a row
 const MaxNonRetransmittableAcks = 19
diff --git a/internal/utils/minmax.go b/internal/utils/minmax.go
index ef71c7fa..4394ab04 100644
--- a/internal/utils/minmax.go
+++ b/internal/utils/minmax.go
@@ -82,6 +82,14 @@ func MinByteCount(a, b protocol.ByteCount) protocol.ByteCount {
 	return b
 }
 
+// MaxByteCount returns the maximum of two ByteCounts
+func MaxByteCount(a, b protocol.ByteCount) protocol.ByteCount {
+	if a < b {
+		return b
+	}
+	return a
+}
+
 // MaxDuration returns the max duration
 func MaxDuration(a, b time.Duration) time.Duration {
 	if a > b {
diff --git a/internal/utils/minmax_test.go b/internal/utils/minmax_test.go
index 0a5dbb08..95816372 100644
--- a/internal/utils/minmax_test.go
+++ b/internal/utils/minmax_test.go
@@ -35,6 +35,11 @@ var _ = Describe("Min / Max", func() {
 			Expect(MaxInt64(7, 5)).To(Equal(int64(7)))
 		})
 
+		It("returns the maximum ByteCount", func() {
+			Expect(MaxByteCount(7, 5)).To(Equal(protocol.ByteCount(7)))
+			Expect(MaxByteCount(5, 7)).To(Equal(protocol.ByteCount(7)))
+		})
+
 		It("returns the maximum duration", func() {
 			Expect(MaxDuration(time.Microsecond, time.Nanosecond)).To(Equal(time.Microsecond))
 			Expect(MaxDuration(time.Nanosecond, time.Microsecond)).To(Equal(time.Microsecond))