[TCP]: TCP Compound congestion control
authorAngelo P. Castellani <angelo.castellani@gmail.com>
Tue, 6 Jun 2006 00:29:09 +0000 (17:29 -0700)
committerDavid S. Miller <davem@sunset.davemloft.net>
Sun, 18 Jun 2006 04:29:25 +0000 (21:29 -0700)
TCP Compound is a sender-side only change to TCP that uses
a mixed Reno/Vegas approach to calculate the cwnd.

For further details look here:
  ftp://ftp.research.microsoft.com/pub/tr/TR-2005-86.pdf

Signed-off-by: Angelo P. Castellani <angelo.castellani@gmail.com>
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_compound.c [new file with mode: 0644]

index 8514106761b0ce6637b3fde1ead4af203d43e9d1..da33393be45f3d2e4d964dcdb9dbde58e59a1d84 100644 (file)
@@ -572,6 +572,16 @@ config TCP_CONG_VENO
        loss packets.
        See http://www.ntu.edu.sg/home5/ZHOU0022/papers/CPFu03a.pdf
 
+config TCP_CONG_COMPOUND
+       tristate "TCP Compound"
+       depends on EXPERIMENTAL
+       default n
+       ---help---
+       TCP Compound is a sender-side only change to TCP that uses
+       a mixed Reno/Vegas approach to calculate the cwnd.
+       For further details look here:
+         ftp://ftp.research.microsoft.com/pub/tr/TR-2005-86.pdf
+
 endmenu
 
 config TCP_CONG_BIC
index 00fcd2c1ba788e2ccee24c7f3027e137792c7f9f..f30faefc26727a3e54cf912582625d2bf4a9b0d6 100644 (file)
@@ -46,6 +46,7 @@ obj-$(CONFIG_TCP_CONG_VEGAS) += tcp_vegas.o
 obj-$(CONFIG_TCP_CONG_VENO) += tcp_veno.o
 obj-$(CONFIG_TCP_CONG_SCALABLE) += tcp_scalable.o
 obj-$(CONFIG_TCP_CONG_LP) += tcp_lp.o
+obj-$(CONFIG_TCP_CONG_COMPOUND) += tcp_compound.o
 
 obj-$(CONFIG_XFRM) += xfrm4_policy.o xfrm4_state.o xfrm4_input.o \
                      xfrm4_output.o
diff --git a/net/ipv4/tcp_compound.c b/net/ipv4/tcp_compound.c
new file mode 100644 (file)
index 0000000..8c1ebfb
--- /dev/null
@@ -0,0 +1,407 @@
+/*
+ * TCP Vegas congestion control
+ *
+ * This is based on the congestion detection/avoidance scheme described in
+ *    Lawrence S. Brakmo and Larry L. Peterson.
+ *    "TCP Vegas: End to end congestion avoidance on a global internet."
+ *    IEEE Journal on Selected Areas in Communication, 13(8):1465--1480,
+ *    October 1995. Available from:
+ *     ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps
+ *
+ * See http://www.cs.arizona.edu/xkernel/ for their implementation.
+ * The main aspects that distinguish this implementation from the
+ * Arizona Vegas implementation are:
+ *   o We do not change the loss detection or recovery mechanisms of
+ *     Linux in any way. Linux already recovers from losses quite well,
+ *     using fine-grained timers, NewReno, and FACK.
+ *   o To avoid the performance penalty imposed by increasing cwnd
+ *     only every-other RTT during slow start, we increase during
+ *     every RTT during slow start, just like Reno.
+ *   o Largely to allow continuous cwnd growth during slow start,
+ *     we use the rate at which ACKs come back as the "actual"
+ *     rate, rather than the rate at which data is sent.
+ *   o To speed convergence to the right rate, we set the cwnd
+ *     to achieve the right ("actual") rate when we exit slow start.
+ *   o To filter out the noise caused by delayed ACKs, we use the
+ *     minimum RTT sample observed during the last RTT to calculate
+ *     the actual rate.
+ *   o When the sender re-starts from idle, it waits until it has
+ *     received ACKs for an entire flight of new data before making
+ *     a cwnd adjustment decision. The original Vegas implementation
+ *     assumed senders never went idle.
+ *
+ *
+ *   TCP Compound based on TCP Vegas
+ *
+ *   further details can be found here:
+ *      ftp://ftp.research.microsoft.com/pub/tr/TR-2005-86.pdf
+ */
+
+#include <linux/config.h>
+#include <linux/mm.h>
+#include <linux/module.h>
+#include <linux/skbuff.h>
+#include <linux/inet_diag.h>
+
+#include <net/tcp.h>
+
+/* Default values of the Vegas variables, in fixed-point representation
+ * with V_PARAM_SHIFT bits to the right of the binary point.
+ */
+#define V_PARAM_SHIFT 1
+
+#define TCP_COMPOUND_ALPHA          3U
+#define TCP_COMPOUND_BETA           1U
+#define TCP_COMPOUND_KAPPA_POW      3
+#define TCP_COMPOUND_KAPPA_NSQRT    2
+#define TCP_COMPOUND_GAMMA         30
+#define TCP_COMPOUND_ZETA           1
+
+/* TCP compound variables */
+struct compound {
+       u32 beg_snd_nxt;        /* right edge during last RTT */
+       u32 beg_snd_una;        /* left edge  during last RTT */
+       u32 beg_snd_cwnd;       /* saves the size of the cwnd */
+       u8 doing_vegas_now;     /* if true, do vegas for this RTT */
+       u16 cntRTT;             /* # of RTTs measured within last RTT */
+       u32 minRTT;             /* min of RTTs measured within last RTT (in usec) */
+       u32 baseRTT;            /* the min of all Vegas RTT measurements seen (in usec) */
+
+       u32 cwnd;
+       u32 dwnd;
+};
+
+/* There are several situations when we must "re-start" Vegas:
+ *
+ *  o when a connection is established
+ *  o after an RTO
+ *  o after fast recovery
+ *  o when we send a packet and there is no outstanding
+ *    unacknowledged data (restarting an idle connection)
+ *
+ * In these circumstances we cannot do a Vegas calculation at the
+ * end of the first RTT, because any calculation we do is using
+ * stale info -- both the saved cwnd and congestion feedback are
+ * stale.
+ *
+ * Instead we must wait until the completion of an RTT during
+ * which we actually receive ACKs.
+ */
+static inline void vegas_enable(struct sock *sk)
+{
+       const struct tcp_sock *tp = tcp_sk(sk);
+       struct compound *vegas = inet_csk_ca(sk);
+
+       /* Begin taking Vegas samples next time we send something. */
+       vegas->doing_vegas_now = 1;
+
+       /* Set the beginning of the next send window. */
+       vegas->beg_snd_nxt = tp->snd_nxt;
+
+       vegas->cntRTT = 0;
+       vegas->minRTT = 0x7fffffff;
+}
+
+/* Stop taking Vegas samples for now. */
+static inline void vegas_disable(struct sock *sk)
+{
+       struct compound *vegas = inet_csk_ca(sk);
+
+       vegas->doing_vegas_now = 0;
+}
+
+static void tcp_compound_init(struct sock *sk)
+{
+       struct compound *vegas = inet_csk_ca(sk);
+       const struct tcp_sock *tp = tcp_sk(sk);
+
+       vegas->baseRTT = 0x7fffffff;
+       vegas_enable(sk);
+
+       vegas->dwnd = 0;
+       vegas->cwnd = tp->snd_cwnd;
+}
+
+/* Do RTT sampling needed for Vegas.
+ * Basically we:
+ *   o min-filter RTT samples from within an RTT to get the current
+ *     propagation delay + queuing delay (we are min-filtering to try to
+ *     avoid the effects of delayed ACKs)
+ *   o min-filter RTT samples from a much longer window (forever for now)
+ *     to find the propagation delay (baseRTT)
+ */
+static void tcp_compound_rtt_calc(struct sock *sk, u32 usrtt)
+{
+       struct compound *vegas = inet_csk_ca(sk);
+       u32 vrtt = usrtt + 1;   /* Never allow zero rtt or baseRTT */
+
+       /* Filter to find propagation delay: */
+       if (vrtt < vegas->baseRTT)
+               vegas->baseRTT = vrtt;
+
+       /* Find the min RTT during the last RTT to find
+        * the current prop. delay + queuing delay:
+        */
+
+       vegas->minRTT = min(vegas->minRTT, vrtt);
+       vegas->cntRTT++;
+}
+
+static void tcp_compound_state(struct sock *sk, u8 ca_state)
+{
+
+       if (ca_state == TCP_CA_Open)
+               vegas_enable(sk);
+       else
+               vegas_disable(sk);
+}
+
+/*
+ * If the connection is idle and we are restarting,
+ * then we don't want to do any Vegas calculations
+ * until we get fresh RTT samples.  So when we
+ * restart, we reset our Vegas state to a clean
+ * slate. After we get acks for this flight of
+ * packets, _then_ we can make Vegas calculations
+ * again.
+ */
+static void tcp_compound_cwnd_event(struct sock *sk, enum tcp_ca_event event)
+{
+       if (event == CA_EVENT_CWND_RESTART || event == CA_EVENT_TX_START)
+               tcp_compound_init(sk);
+}
+
+static void tcp_compound_cong_avoid(struct sock *sk, u32 ack,
+                                   u32 seq_rtt, u32 in_flight, int flag)
+{
+       struct tcp_sock *tp = tcp_sk(sk);
+       struct compound *vegas = inet_csk_ca(sk);
+       u8 inc = 0;
+
+       if (vegas->cwnd + vegas->dwnd > tp->snd_cwnd) {
+               if (vegas->cwnd > tp->snd_cwnd || vegas->dwnd > tp->snd_cwnd) {
+                       vegas->cwnd = tp->snd_cwnd;
+                       vegas->dwnd = 0;
+               } else
+                       vegas->cwnd = tp->snd_cwnd - vegas->dwnd;
+
+       }
+
+       if (!tcp_is_cwnd_limited(sk, in_flight))
+               return;
+
+       if (vegas->cwnd <= tp->snd_ssthresh)
+               inc = 1;
+       else if (tp->snd_cwnd_cnt < tp->snd_cwnd)
+               tp->snd_cwnd_cnt++;
+
+       if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
+               inc = 1;
+               tp->snd_cwnd_cnt = 0;
+       }
+
+       if (inc && tp->snd_cwnd < tp->snd_cwnd_clamp)
+               vegas->cwnd++;
+
+       /* The key players are v_beg_snd_una and v_beg_snd_nxt.
+        *
+        * These are so named because they represent the approximate values
+        * of snd_una and snd_nxt at the beginning of the current RTT. More
+        * precisely, they represent the amount of data sent during the RTT.
+        * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
+        * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
+        * bytes of data have been ACKed during the course of the RTT, giving
+        * an "actual" rate of:
+        *
+        *     (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
+        *
+        * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
+        * because delayed ACKs can cover more than one segment, so they
+        * don't line up nicely with the boundaries of RTTs.
+        *
+        * Another unfortunate fact of life is that delayed ACKs delay the
+        * advance of the left edge of our send window, so that the number
+        * of bytes we send in an RTT is often less than our cwnd will allow.
+        * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
+        */
+
+       if (after(ack, vegas->beg_snd_nxt)) {
+               /* Do the Vegas once-per-RTT cwnd adjustment. */
+               u32 old_wnd, old_snd_cwnd;
+
+               /* Here old_wnd is essentially the window of data that was
+                * sent during the previous RTT, and has all
+                * been acknowledged in the course of the RTT that ended
+                * with the ACK we just received. Likewise, old_snd_cwnd
+                * is the cwnd during the previous RTT.
+                */
+               if (!tp->mss_cache)
+                       return;
+
+               old_wnd = (vegas->beg_snd_nxt - vegas->beg_snd_una) /
+                   tp->mss_cache;
+               old_snd_cwnd = vegas->beg_snd_cwnd;
+
+               /* Save the extent of the current window so we can use this
+                * at the end of the next RTT.
+                */
+               vegas->beg_snd_una = vegas->beg_snd_nxt;
+               vegas->beg_snd_nxt = tp->snd_nxt;
+               vegas->beg_snd_cwnd = tp->snd_cwnd;
+
+               /* We do the Vegas calculations only if we got enough RTT
+                * samples that we can be reasonably sure that we got
+                * at least one RTT sample that wasn't from a delayed ACK.
+                * If we only had 2 samples total,
+                * then that means we're getting only 1 ACK per RTT, which
+                * means they're almost certainly delayed ACKs.
+                * If  we have 3 samples, we should be OK.
+                */
+
+               if (vegas->cntRTT > 2) {
+                       u32 rtt, target_cwnd, diff;
+                       u32 brtt, dwnd;
+
+                       /* We have enough RTT samples, so, using the Vegas
+                        * algorithm, we determine if we should increase or
+                        * decrease cwnd, and by how much.
+                        */
+
+                       /* Pluck out the RTT we are using for the Vegas
+                        * calculations. This is the min RTT seen during the
+                        * last RTT. Taking the min filters out the effects
+                        * of delayed ACKs, at the cost of noticing congestion
+                        * a bit later.
+                        */
+                       rtt = vegas->minRTT;
+
+                       /* Calculate the cwnd we should have, if we weren't
+                        * going too fast.
+                        *
+                        * This is:
+                        *     (actual rate in segments) * baseRTT
+                        * We keep it as a fixed point number with
+                        * V_PARAM_SHIFT bits to the right of the binary point.
+                        */
+                       if (!rtt)
+                               return;
+
+                       brtt = vegas->baseRTT;
+                       target_cwnd = ((old_wnd * brtt)
+                                      << V_PARAM_SHIFT) / rtt;
+
+                       /* Calculate the difference between the window we had,
+                        * and the window we would like to have. This quantity
+                        * is the "Diff" from the Arizona Vegas papers.
+                        *
+                        * Again, this is a fixed point number with
+                        * V_PARAM_SHIFT bits to the right of the binary
+                        * point.
+                        */
+
+                       diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
+
+                       dwnd = vegas->dwnd;
+
+                       if (diff < (TCP_COMPOUND_GAMMA << V_PARAM_SHIFT)) {
+                               u32 i, j, x, x2;
+                               u64 v;
+
+                               v = 1;
+
+                               for (i = 0; i < TCP_COMPOUND_KAPPA_POW; i++)
+                                       v *= old_wnd;
+
+                               for (i = 0; i < TCP_COMPOUND_KAPPA_NSQRT; i++) {
+                                       x = 1;
+                                       for (j = 0; j < 200; j++) {
+                                               x2 = (x + v / x) / 2;
+
+                                               if (x2 == x || !x2)
+                                                       break;
+
+                                               x = x2;
+                                       }
+                                       v = x;
+                               }
+
+                               x = (u32) v >> TCP_COMPOUND_ALPHA;
+
+                               if (x > 1)
+                                       dwnd = x - 1;
+                               else
+                                       dwnd = 0;
+
+                               dwnd += vegas->dwnd;
+
+                       } else if ((dwnd << V_PARAM_SHIFT) <
+                                  (diff * TCP_COMPOUND_BETA))
+                               dwnd = 0;
+                       else
+                               dwnd =
+                                   ((dwnd << V_PARAM_SHIFT) -
+                                    (diff *
+                                     TCP_COMPOUND_BETA)) >> V_PARAM_SHIFT;
+
+                       vegas->dwnd = dwnd;
+
+               }
+
+               /* Wipe the slate clean for the next RTT. */
+               vegas->cntRTT = 0;
+               vegas->minRTT = 0x7fffffff;
+       }
+
+       tp->snd_cwnd = vegas->cwnd + vegas->dwnd;
+}
+
+/* Extract info for Tcp socket info provided via netlink. */
+static void tcp_compound_get_info(struct sock *sk, u32 ext, struct sk_buff *skb)
+{
+       const struct compound *ca = inet_csk_ca(sk);
+       if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
+               struct tcpvegas_info *info;
+
+               info = RTA_DATA(__RTA_PUT(skb, INET_DIAG_VEGASINFO,
+                                         sizeof(*info)));
+
+               info->tcpv_enabled = ca->doing_vegas_now;
+               info->tcpv_rttcnt = ca->cntRTT;
+               info->tcpv_rtt = ca->baseRTT;
+               info->tcpv_minrtt = ca->minRTT;
+       rtattr_failure:;
+       }
+}
+
+static struct tcp_congestion_ops tcp_compound = {
+       .init           = tcp_compound_init,
+       .ssthresh       = tcp_reno_ssthresh,
+       .cong_avoid     = tcp_compound_cong_avoid,
+       .min_cwnd       = tcp_reno_min_cwnd,
+       .rtt_sample     = tcp_compound_rtt_calc,
+       .set_state      = tcp_compound_state,
+       .cwnd_event     = tcp_compound_cwnd_event,
+       .get_info       = tcp_compound_get_info,
+
+       .owner          = THIS_MODULE,
+       .name           = "compound",
+};
+
+static int __init tcp_compound_register(void)
+{
+       BUG_ON(sizeof(struct compound) > ICSK_CA_PRIV_SIZE);
+       tcp_register_congestion_control(&tcp_compound);
+       return 0;
+}
+
+static void __exit tcp_compound_unregister(void)
+{
+       tcp_unregister_congestion_control(&tcp_compound);
+}
+
+module_init(tcp_compound_register);
+module_exit(tcp_compound_unregister);
+
+MODULE_AUTHOR("Angelo P. Castellani, Stephen Hemminger");
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("TCP Compound");