[TCP] BIC: CUBIC window growth (2.0)
Stephen Hemminger [Wed, 14 Dec 2005 07:13:28 +0000 (23:13 -0800)]
Replace existing BIC version 1.1 with new version 2.0.
The main change is to replace the window growth function
with a cubic function as described in:
  http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/cubic-paper.pdf

Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

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

index e55136a..011cca7 100644 (file)
@@ -456,6 +456,14 @@ config TCP_CONG_BIC
        increase provides TCP friendliness.
        See http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/
 
+config TCP_CONG_CUBIC
+       tristate "CUBIC TCP"
+       default m
+       ---help---
+       This is version 2.0 of BIC-TCP which uses a cubic growth function
+       among other techniques.
+       See http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/cubic-paper.pdf
+
 config TCP_CONG_WESTWOOD
        tristate "TCP Westwood+"
        default m
index f0435d0..c54edd7 100644 (file)
@@ -34,6 +34,7 @@ obj-$(CONFIG_INET_DIAG) += inet_diag.o
 obj-$(CONFIG_IP_ROUTE_MULTIPATH_CACHED) += multipath.o
 obj-$(CONFIG_INET_TCP_DIAG) += tcp_diag.o
 obj-$(CONFIG_TCP_CONG_BIC) += tcp_bic.o
+obj-$(CONFIG_TCP_CONG_CUBIC) += tcp_cubic.o
 obj-$(CONFIG_TCP_CONG_WESTWOOD) += tcp_westwood.o
 obj-$(CONFIG_TCP_CONG_HSTCP) += tcp_highspeed.o
 obj-$(CONFIG_TCP_CONG_HYBLA) += tcp_hybla.o
diff --git a/net/ipv4/tcp_cubic.c b/net/ipv4/tcp_cubic.c
new file mode 100644 (file)
index 0000000..bb5dc4b
--- /dev/null
@@ -0,0 +1,445 @@
+/*
+ * TCP CUBIC: Binary Increase Congestion control for TCP v2.0
+ *
+ * This is from the implementation of CUBIC TCP in
+ * Injong Rhee, Lisong Xu.
+ *  "CUBIC: A New TCP-Friendly High-Speed TCP Variant
+ *  in PFLDnet 2005
+ * Available from:
+ *  http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/cubic-paper.pdf
+ *
+ * Unless CUBIC is enabled and congestion window is large
+ * this behaves the same as the original Reno.
+ */
+
+#include <linux/config.h>
+#include <linux/mm.h>
+#include <linux/module.h>
+#include <net/tcp.h>
+
+
+#define BICTCP_BETA_SCALE    1024      /* Scale factor beta calculation
+                                        * max_cwnd = snd_cwnd * beta
+                                        */
+#define BICTCP_B               4        /*
+                                         * In binary search,
+                                         * go to point (max+min)/N
+                                         */
+#define        BICTCP_HZ               10      /* BIC HZ 2^10 = 1024 */
+
+static int fast_convergence = 1;
+static int max_increment = 16;
+static int beta = 819;         /* = 819/1024 (BICTCP_BETA_SCALE) */
+static int initial_ssthresh = 100;
+static int bic_scale = 41;
+static int tcp_friendliness = 1;
+
+module_param(fast_convergence, int, 0644);
+MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
+module_param(max_increment, int, 0644);
+MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
+module_param(beta, int, 0644);
+MODULE_PARM_DESC(beta, "beta for multiplicative increase");
+module_param(initial_ssthresh, int, 0644);
+MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
+module_param(bic_scale, int, 0644);
+MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
+module_param(tcp_friendliness, int, 0644);
+MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
+
+
+/* BIC TCP Parameters */
+struct bictcp {
+       u32     cnt;            /* increase cwnd by 1 after ACKs */
+       u32     last_max_cwnd;  /* last maximum snd_cwnd */
+       u32     loss_cwnd;      /* congestion window at last loss */
+       u32     last_cwnd;      /* the last snd_cwnd */
+       u32     last_time;      /* time when updated last_cwnd */
+       u32     bic_origin_point;/* origin point of bic function */
+       u32     bic_K;          /* time to origin point from the beginning of the current epoch */
+       u32     delay_min;      /* min delay */
+       u32     epoch_start;    /* beginning of an epoch */
+       u32     ack_cnt;        /* number of acks */
+       u32     tcp_cwnd;       /* estimated tcp cwnd */
+#define ACK_RATIO_SHIFT        4
+       u32     delayed_ack;    /* estimate the ratio of Packets/ACKs << 4 */
+};
+
+static inline void bictcp_reset(struct bictcp *ca)
+{
+       ca->cnt = 0;
+       ca->last_max_cwnd = 0;
+       ca->loss_cwnd = 0;
+       ca->last_cwnd = 0;
+       ca->last_time = 0;
+       ca->bic_origin_point = 0;
+       ca->bic_K = 0;
+       ca->delay_min = 0;
+       ca->epoch_start = 0;
+       ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
+       ca->ack_cnt = 0;
+       ca->tcp_cwnd = 0;
+}
+
+static void bictcp_init(struct sock *sk)
+{
+       bictcp_reset(inet_csk_ca(sk));
+       if (initial_ssthresh)
+               tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
+}
+
+/* 65536 times the cubic root */
+static const u64 cubic_table[8]
+       = {0, 65536, 82570, 94519, 104030, 112063, 119087, 125367};
+
+/*
+ * calculate the cubic root of x
+ * the basic idea is that x can be expressed as i*8^j
+ * so cubic_root(x) = cubic_root(i)*2^j
+ *  in the following code, x is i, and y is 2^j
+ *  because of integer calculation, there are errors in calculation
+ *  so finally use binary search to find out the exact solution
+ */
+static u32 cubic_root(u64 x)
+{
+        u64 y, app, target, start, end, mid, start_diff, end_diff;
+
+        if (x == 0)
+                return 0;
+
+        target = x;
+
+        /* first estimate lower and upper bound */
+        y = 1;
+        while (x >= 8){
+                x = (x >> 3);
+                y = (y << 1);
+        }
+        start = (y*cubic_table[x])>>16;
+        if (x==7)
+                end = (y<<1);
+        else
+                end = (y*cubic_table[x+1]+65535)>>16;
+
+        /* binary search for more accurate one */
+        while (start < end-1) {
+                mid = (start+end) >> 1;
+                app = mid*mid*mid;
+                if (app < target)
+                        start = mid;
+                else if (app > target)
+                        end = mid;
+                else
+                        return mid;
+        }
+
+        /* find the most accurate one from start and end */
+        app = start*start*start;
+        if (app < target)
+                start_diff = target - app;
+        else
+                start_diff = app - target;
+        app = end*end*end;
+        if (app < target)
+                end_diff = target - app;
+        else
+                end_diff = app - target;
+
+        if (start_diff < end_diff)
+                return (u32)start;
+        else
+                return (u32)end;
+}
+
+static inline u32 bictcp_K(u32 dist, u32 srtt)
+{
+        u64 d64;
+        u32 d32;
+        u32 count;
+        u32 result;
+
+        /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
+           so K = cubic_root( (wmax-cwnd)*rtt/c )
+           the unit of K is bictcp_HZ=2^10, not HZ
+
+           c = bic_scale >> 10
+           rtt = (tp->srtt >> 3 ) / HZ
+
+           the following code has been designed and tested for
+           cwnd < 1 million packets
+           RTT < 100 seconds
+           HZ < 1,000,00  (corresponding to 10 nano-second)
+
+        */
+
+        /* 1/c * 2^2*bictcp_HZ */
+        d32 = (1 << (10+2*BICTCP_HZ)) / bic_scale;
+        d64 = (__u64)d32;
+
+        /* srtt * 2^count / HZ
+           1) to get a better accuracy of the following d32,
+           the larger the "count", the better the accuracy
+           2) and avoid overflow of the following d64
+           the larger the "count", the high possibility of overflow
+           3) so find a "count" between bictcp_hz-3 and bictcp_hz
+           "count" may be less than bictcp_HZ,
+           then d64 becomes 0. that is OK
+        */
+        d32 = srtt;
+        count = 0;
+        while (((d32 & 0x80000000)==0) && (count < BICTCP_HZ)){
+                d32 = d32 << 1;
+                count++;
+        }
+        d32 = d32 / HZ;
+
+        /* (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)  */
+        d64 = (d64 * dist * d32) >> (count+3-BICTCP_HZ);
+
+        /* cubic root */
+        d64 = cubic_root(d64);
+
+        result = (u32)d64;
+        return result;
+}
+
+/*
+ * Compute congestion window to use.
+ */
+static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
+{
+       u64 d64;
+       u32 d32, t, srtt, bic_target, min_cnt, max_cnt;
+
+       ca->ack_cnt++;  /* count the number of ACKs */
+
+       if (ca->last_cwnd == cwnd &&
+           (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
+               return;
+
+       ca->last_cwnd = cwnd;
+       ca->last_time = tcp_time_stamp;
+
+       srtt = (HZ << 3)/10;    /* use real time-based growth function */
+
+       if (ca->epoch_start == 0) {
+               ca->epoch_start = tcp_time_stamp;       /* record the beginning of an epoch */
+               ca->ack_cnt = 1;                        /* start counting */
+               ca->tcp_cwnd = cwnd;                    /* syn with cubic */
+
+               if (ca->last_max_cwnd <= cwnd) {
+                       ca->bic_K = 0;
+                       ca->bic_origin_point = cwnd;
+               } else {
+                       ca->bic_K = bictcp_K(ca->last_max_cwnd-cwnd, srtt);
+                       ca->bic_origin_point = ca->last_max_cwnd;
+               }
+       }
+
+        /* cubic function - calc*/
+        /* calculate c * time^3 / rtt,
+         *  while considering overflow in calculation of time^3
+        * (so time^3 is done by using d64)
+        * and without the support of division of 64bit numbers
+        * (so all divisions are done by using d32)
+         *  also NOTE the unit of those veriables
+         *       time  = (t - K) / 2^bictcp_HZ
+         *       c = bic_scale >> 10
+        * rtt  = (srtt >> 3) / HZ
+        * !!! The following code does not have overflow problems,
+        * if the cwnd < 1 million packets !!!
+         */
+
+       /* change the unit from HZ to bictcp_HZ */
+        t = ((tcp_time_stamp + ca->delay_min - ca->epoch_start)
+            << BICTCP_HZ) / HZ;
+
+        if (t < ca->bic_K)             /* t - K */
+                d32 = ca->bic_K - t;
+        else
+                d32 = t - ca->bic_K;
+
+        d64 = (u64)d32;
+        d32 = (bic_scale << 3) * HZ / srtt;                    /* 1024*c/rtt */
+        d64 = (d32 * d64 * d64 * d64) >> (10+3*BICTCP_HZ);     /* c/rtt * (t-K)^3 */
+        d32 = (u32)d64;
+        if (t < ca->bic_K)                                     /* below origin*/
+                bic_target = ca->bic_origin_point - d32;
+        else                                                   /* above origin*/
+                bic_target = ca->bic_origin_point + d32;
+
+        /* cubic function - calc bictcp_cnt*/
+        if (bic_target > cwnd) {
+               ca->cnt = cwnd / (bic_target - cwnd);
+        } else {
+                ca->cnt = 100 * cwnd;              /* very small increment*/
+        }
+
+       if (ca->delay_min > 0) {
+               /* max increment = Smax * rtt / 0.1  */
+               min_cnt = (cwnd * HZ * 8)/(10 * max_increment * ca->delay_min);
+               if (ca->cnt < min_cnt)
+                       ca->cnt = min_cnt;
+       }
+
+        /* slow start and low utilization  */
+       if (ca->loss_cwnd == 0)         /* could be aggressive in slow start */
+               ca->cnt = 50;
+
+       /* TCP Friendly */
+       if (tcp_friendliness) {
+               u32 scale = 8*(BICTCP_BETA_SCALE+beta)/3/(BICTCP_BETA_SCALE-beta);
+               d32 = (cwnd * scale) >> 3;
+               while (ca->ack_cnt > d32) {             /* update tcp cwnd */
+                       ca->ack_cnt -= d32;
+                       ca->tcp_cwnd++;
+               }
+
+               if (ca->tcp_cwnd > cwnd){       /* if bic is slower than tcp */
+                       d32 = ca->tcp_cwnd - cwnd;
+                       max_cnt = cwnd / d32;
+                       if (ca->cnt > max_cnt)
+                               ca->cnt = max_cnt;
+               }
+        }
+
+       ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
+       if (ca->cnt == 0)                       /* cannot be zero */
+               ca->cnt = 1;
+}
+
+
+/* Keep track of minimum rtt */
+static inline void measure_delay(struct sock *sk)
+{
+       const struct tcp_sock *tp = tcp_sk(sk);
+       struct bictcp *ca = inet_csk_ca(sk);
+       u32 delay;
+
+       /* No time stamp */
+       if (!(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) ||
+            /* Discard delay samples right after fast recovery */
+           (s32)(tcp_time_stamp - ca->epoch_start) < HZ)
+               return;
+
+       delay = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
+       if (delay == 0)
+               delay = 1;
+
+       /* first time call or link delay decreases */
+       if (ca->delay_min == 0 || ca->delay_min > delay)
+               ca->delay_min = delay;
+}
+
+static void bictcp_cong_avoid(struct sock *sk, u32 ack,
+                             u32 seq_rtt, u32 in_flight, int data_acked)
+{
+       struct tcp_sock *tp = tcp_sk(sk);
+       struct bictcp *ca = inet_csk_ca(sk);
+
+       if (data_acked)
+               measure_delay(sk);
+
+       if (!tcp_is_cwnd_limited(sk, in_flight))
+               return;
+
+       if (tp->snd_cwnd <= tp->snd_ssthresh)
+               tcp_slow_start(tp);
+       else {
+               bictcp_update(ca, tp->snd_cwnd);
+
+               /* In dangerous area, increase slowly.
+                * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
+                */
+               if (tp->snd_cwnd_cnt >= ca->cnt) {
+                       if (tp->snd_cwnd < tp->snd_cwnd_clamp)
+                               tp->snd_cwnd++;
+                       tp->snd_cwnd_cnt = 0;
+               } else
+                       tp->snd_cwnd_cnt++;
+       }
+
+}
+
+static u32 bictcp_recalc_ssthresh(struct sock *sk)
+{
+       const struct tcp_sock *tp = tcp_sk(sk);
+       struct bictcp *ca = inet_csk_ca(sk);
+
+       ca->epoch_start = 0;    /* end of epoch */
+
+       /* Wmax and fast convergence */
+       if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
+               ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
+                       / (2 * BICTCP_BETA_SCALE);
+       else
+               ca->last_max_cwnd = tp->snd_cwnd;
+
+       ca->loss_cwnd = tp->snd_cwnd;
+
+       return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
+}
+
+static u32 bictcp_undo_cwnd(struct sock *sk)
+{
+       struct bictcp *ca = inet_csk_ca(sk);
+
+       return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
+}
+
+static u32 bictcp_min_cwnd(struct sock *sk)
+{
+       return tcp_sk(sk)->snd_ssthresh;
+}
+
+static void bictcp_state(struct sock *sk, u8 new_state)
+{
+       if (new_state == TCP_CA_Loss)
+               bictcp_reset(inet_csk_ca(sk));
+}
+
+/* Track delayed acknowledgment ratio using sliding window
+ * ratio = (15*ratio + sample) / 16
+ */
+static void bictcp_acked(struct sock *sk, u32 cnt)
+{
+       const struct inet_connection_sock *icsk = inet_csk(sk);
+
+       if (cnt > 0 && icsk->icsk_ca_state == TCP_CA_Open) {
+               struct bictcp *ca = inet_csk_ca(sk);
+               cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
+               ca->delayed_ack += cnt;
+       }
+}
+
+
+static struct tcp_congestion_ops cubictcp = {
+       .init           = bictcp_init,
+       .ssthresh       = bictcp_recalc_ssthresh,
+       .cong_avoid     = bictcp_cong_avoid,
+       .set_state      = bictcp_state,
+       .undo_cwnd      = bictcp_undo_cwnd,
+       .min_cwnd       = bictcp_min_cwnd,
+       .pkts_acked     = bictcp_acked,
+       .owner          = THIS_MODULE,
+       .name           = "cubic",
+};
+
+static int __init cubictcp_register(void)
+{
+       BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
+       return tcp_register_congestion_control(&cubictcp);
+}
+
+static void __exit cubictcp_unregister(void)
+{
+       tcp_unregister_congestion_control(&cubictcp);
+}
+
+module_init(cubictcp_register);
+module_exit(cubictcp_unregister);
+
+MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("CUBIC TCP");
+MODULE_VERSION("2.0");