[TCP]: Add H-TCP congestion control module.
Baruch Even [Thu, 23 Jun 2005 19:28:11 +0000 (12:28 -0700)]
H-TCP is a congestion control algorithm developed at the Hamilton Institute, by
Douglas Leith and Robert Shorten. It is extending the standard Reno algorithm
with mode switching is thus a relatively simple modification.

H-TCP is defined in a layered manner as it is still a research platform. The
basic form includes the modification of beta according to the ratio of maxRTT
to min RTT and the alpha=2*factor*(1-beta) relation, where factor is dependant
on the time since last congestion.

The other layers improve convergence by adding appropriate factors to alpha.

The following patch implements the H-TCP algorithm in it's basic form.

Signed-Off-By: Baruch Even <baruch@ev-en.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/ipv4/Kconfig
net/ipv4/Makefile
net/ipv4/tcp_htcp.c [new file with mode: 0644]

index 6c105b6..73a25b5 100644 (file)
@@ -467,6 +467,18 @@ config TCP_CONG_WESTWOOD
        TCP Westwood+ significantly increases fairness wrt TCP Reno in
        wired networks and throughput over wireless links.
 
+config TCP_CONG_HTCP
+        tristate "H-TCP"
+       depends on INET
+        default m
+       ---help---
+       H-TCP is a send-side only modifications of the TCP Reno
+       protocol stack that optimizes the performance of TCP
+       congestion control for high speed network links. It uses a
+       modeswitch to change the alpha and beta parameters of TCP Reno
+       based on network conditions and in a way so as to be fair with
+       other Reno and H-TCP flows.
+
 config TCP_CONG_HSTCP
        tristate "High Speed TCP"
        depends on INET && EXPERIMENTAL
@@ -499,6 +511,7 @@ config TCP_CONG_VEGAS
        window. TCP Vegas should provide less packet loss, but it is
        not as aggressive as TCP Reno.
 
+
 endmenu
 
 source "net/ipv4/ipvs/Kconfig"
index a801a97..e96ed17 100644 (file)
@@ -35,6 +35,7 @@ obj-$(CONFIG_TCP_CONG_BIC) += tcp_bic.o
 obj-$(CONFIG_TCP_CONG_WESTWOOD) += tcp_westwood.o
 obj-$(CONFIG_TCP_CONG_HSTCP) += tcp_highspeed.o
 obj-$(CONFIG_TCP_CONG_HYBLA) += tcp_hybla.o
+obj-$(CONFIG_TCP_CONG_HTCP) += tcp_htcp.o
 obj-$(CONFIG_TCP_CONG_VEGAS) += tcp_vegas.o
 
 obj-$(CONFIG_XFRM) += xfrm4_policy.o xfrm4_state.o xfrm4_input.o \
diff --git a/net/ipv4/tcp_htcp.c b/net/ipv4/tcp_htcp.c
new file mode 100644 (file)
index 0000000..4016827
--- /dev/null
@@ -0,0 +1,289 @@
+/*
+ * H-TCP congestion control. The algorithm is detailed in:
+ * R.N.Shorten, D.J.Leith:
+ *   "H-TCP: TCP for high-speed and long-distance networks"
+ *   Proc. PFLDnet, Argonne, 2004.
+ * http://www.hamilton.ie/net/htcp3.pdf
+ */
+
+#include <linux/config.h>
+#include <linux/mm.h>
+#include <linux/module.h>
+#include <net/tcp.h>
+
+#define ALPHA_BASE     (1<<7)  /* 1.0 with shift << 7 */
+#define BETA_MIN       (1<<6)  /* 0.5 with shift << 7 */
+#define BETA_MAX       102     /* 0.8 with shift << 7 */
+
+static int use_rtt_scaling = 1;
+module_param(use_rtt_scaling, int, 0644);
+MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling");
+
+static int use_bandwidth_switch = 1;
+module_param(use_bandwidth_switch, int, 0644);
+MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher");
+
+struct htcp {
+       u16     alpha;          /* Fixed point arith, << 7 */
+       u8      beta;           /* Fixed point arith, << 7 */
+       u8      modeswitch;     /* Delay modeswitch until we had at least one congestion event */
+       u8      ccount;         /* Number of RTTs since last congestion event */
+       u8      undo_ccount;
+       u16     packetcount;
+       u32     minRTT;
+       u32     maxRTT;
+       u32     snd_cwnd_cnt2;
+
+       u32     undo_maxRTT;
+       u32     undo_old_maxB;
+
+       /* Bandwidth estimation */
+       u32     minB;
+       u32     maxB;
+       u32     old_maxB;
+       u32     Bi;
+       u32     lasttime;
+};
+
+static inline void htcp_reset(struct htcp *ca)
+{
+       ca->undo_ccount = ca->ccount;
+       ca->undo_maxRTT = ca->maxRTT;
+       ca->undo_old_maxB = ca->old_maxB;
+
+       ca->ccount = 0;
+       ca->snd_cwnd_cnt2 = 0;
+}
+
+static u32 htcp_cwnd_undo(struct tcp_sock *tp)
+{
+       struct htcp *ca = tcp_ca(tp);
+       ca->ccount = ca->undo_ccount;
+       ca->maxRTT = ca->undo_maxRTT;
+       ca->old_maxB = ca->undo_old_maxB;
+       return max(tp->snd_cwnd, (tp->snd_ssthresh<<7)/ca->beta);
+}
+
+static inline void measure_rtt(struct tcp_sock *tp)
+{
+       struct htcp *ca = tcp_ca(tp);
+       u32 srtt = tp->srtt>>3;
+
+       /* keep track of minimum RTT seen so far, minRTT is zero at first */
+       if (ca->minRTT > srtt || !ca->minRTT)
+               ca->minRTT = srtt;
+
+       /* max RTT */
+       if (tp->ca_state == TCP_CA_Open && tp->snd_ssthresh < 0xFFFF && ca->ccount > 3) {
+               if (ca->maxRTT < ca->minRTT)
+                       ca->maxRTT = ca->minRTT;
+               if (ca->maxRTT < srtt && srtt <= ca->maxRTT+HZ/50)
+                       ca->maxRTT = srtt;
+       }
+}
+
+static void measure_achieved_throughput(struct tcp_sock *tp, u32 pkts_acked)
+{
+       struct htcp *ca = tcp_ca(tp);
+       u32 now = tcp_time_stamp;
+
+       /* achieved throughput calculations */
+       if (tp->ca_state != TCP_CA_Open && tp->ca_state != TCP_CA_Disorder) {
+               ca->packetcount = 0;
+               ca->lasttime = now;
+               return;
+       }
+
+       ca->packetcount += pkts_acked;
+
+       if (ca->packetcount >= tp->snd_cwnd - (ca->alpha>>7? : 1)
+                       && now - ca->lasttime >= ca->minRTT
+                       && ca->minRTT > 0) {
+               __u32 cur_Bi = ca->packetcount*HZ/(now - ca->lasttime);
+               if (ca->ccount <= 3) {
+                       /* just after backoff */
+                       ca->minB = ca->maxB = ca->Bi = cur_Bi;
+               } else {
+                       ca->Bi = (3*ca->Bi + cur_Bi)/4;
+                       if (ca->Bi > ca->maxB)
+                               ca->maxB = ca->Bi;
+                       if (ca->minB > ca->maxB)
+                               ca->minB = ca->maxB;
+               }
+               ca->packetcount = 0;
+               ca->lasttime = now;
+       }
+}
+
+static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
+{
+       if (use_bandwidth_switch) {
+               u32 maxB = ca->maxB;
+               u32 old_maxB = ca->old_maxB;
+               ca->old_maxB = ca->maxB;
+
+               if (!between(5*maxB, 4*old_maxB, 6*old_maxB)) {
+                       ca->beta = BETA_MIN;
+                       ca->modeswitch = 0;
+                       return;
+               }
+       }
+
+       if (ca->modeswitch && minRTT > max(HZ/100, 1) && maxRTT) {
+               ca->beta = (minRTT<<7)/maxRTT;
+               if (ca->beta < BETA_MIN)
+                       ca->beta = BETA_MIN;
+               else if (ca->beta > BETA_MAX)
+                       ca->beta = BETA_MAX;
+       } else {
+               ca->beta = BETA_MIN;
+               ca->modeswitch = 1;
+       }
+}
+
+static inline void htcp_alpha_update(struct htcp *ca)
+{
+       u32 minRTT = ca->minRTT;
+       u32 factor = 1;
+       u32 diff = ca->ccount * minRTT; /* time since last backoff */
+
+       if (diff > HZ) {
+               diff -= HZ;
+               factor = 1+ ( 10*diff + ((diff/2)*(diff/2)/HZ) )/HZ;
+       }
+
+       if (use_rtt_scaling && minRTT) {
+               u32 scale = (HZ<<3)/(10*minRTT);
+               scale = min(max(scale, 1U<<2), 10U<<3); /* clamping ratio to interval [0.5,10]<<3 */
+               factor = (factor<<3)/scale;
+               if (!factor)
+                       factor = 1;
+       }
+
+       ca->alpha = 2*factor*((1<<7)-ca->beta);
+       if (!ca->alpha)
+               ca->alpha = ALPHA_BASE;
+}
+
+/* After we have the rtt data to calculate beta, we'd still prefer to wait one
+ * rtt before we adjust our beta to ensure we are working from a consistent
+ * data.
+ *
+ * This function should be called when we hit a congestion event since only at
+ * that point do we really have a real sense of maxRTT (the queues en route
+ * were getting just too full now).
+ */
+static void htcp_param_update(struct tcp_sock *tp)
+{
+       struct htcp *ca = tcp_ca(tp);
+       u32 minRTT = ca->minRTT;
+       u32 maxRTT = ca->maxRTT;
+
+       htcp_beta_update(ca, minRTT, maxRTT);
+       htcp_alpha_update(ca);
+
+       /* add slowly fading memory for maxRTT to accommodate routing changes etc */
+       if (minRTT > 0 && maxRTT > minRTT)
+               ca->maxRTT = minRTT + ((maxRTT-minRTT)*95)/100;
+}
+
+static u32 htcp_recalc_ssthresh(struct tcp_sock *tp)
+{
+       struct htcp *ca = tcp_ca(tp);
+       htcp_param_update(tp);
+       return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
+}
+
+static void htcp_cong_avoid(struct tcp_sock *tp, u32 ack, u32 rtt,
+                           u32 in_flight, int data_acked)
+{
+       struct htcp *ca = tcp_ca(tp);
+
+       if (in_flight < tp->snd_cwnd)
+               return;
+
+        if (tp->snd_cwnd <= tp->snd_ssthresh) {
+                /* In "safe" area, increase. */
+               if (tp->snd_cwnd < tp->snd_cwnd_clamp)
+                       tp->snd_cwnd++;
+       } else {
+               measure_rtt(tp);
+
+               /* keep track of number of round-trip times since last backoff event */
+               if (ca->snd_cwnd_cnt2++ > tp->snd_cwnd) {
+                       ca->ccount++;
+                       ca->snd_cwnd_cnt2 = 0;
+                       htcp_alpha_update(ca);
+               }
+
+                /* In dangerous area, increase slowly.
+                * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
+                */
+               if ((tp->snd_cwnd_cnt++ * ca->alpha)>>7 >= tp->snd_cwnd) {
+                       if (tp->snd_cwnd < tp->snd_cwnd_clamp)
+                               tp->snd_cwnd++;
+                       tp->snd_cwnd_cnt = 0;
+                       ca->ccount++;
+               }
+       }
+}
+
+/* Lower bound on congestion window. */
+static u32 htcp_min_cwnd(struct tcp_sock *tp)
+{
+       return tp->snd_ssthresh;
+}
+
+
+static void htcp_init(struct tcp_sock *tp)
+{
+       struct htcp *ca = tcp_ca(tp);
+
+       memset(ca, 0, sizeof(struct htcp));
+       ca->alpha = ALPHA_BASE;
+       ca->beta = BETA_MIN;
+}
+
+static void htcp_state(struct tcp_sock *tp, u8 new_state)
+{
+       switch (new_state) {
+       case TCP_CA_CWR:
+       case TCP_CA_Recovery:
+       case TCP_CA_Loss:
+               htcp_reset(tcp_ca(tp));
+               break;
+       }
+}
+
+static struct tcp_congestion_ops htcp = {
+       .init           = htcp_init,
+       .ssthresh       = htcp_recalc_ssthresh,
+       .min_cwnd       = htcp_min_cwnd,
+       .cong_avoid     = htcp_cong_avoid,
+       .set_state      = htcp_state,
+       .undo_cwnd      = htcp_cwnd_undo,
+       .pkts_acked     = measure_achieved_throughput,
+       .owner          = THIS_MODULE,
+       .name           = "htcp",
+};
+
+static int __init htcp_register(void)
+{
+       BUG_ON(sizeof(struct htcp) > TCP_CA_PRIV_SIZE);
+       BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
+       if (!use_bandwidth_switch)
+               htcp.pkts_acked = NULL;
+       return tcp_register_congestion_control(&htcp);
+}
+
+static void __exit htcp_unregister(void)
+{
+       tcp_unregister_congestion_control(&htcp);
+}
+
+module_init(htcp_register);
+module_exit(htcp_unregister);
+
+MODULE_AUTHOR("Baruch Even");
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("H-TCP");