bpf: Add BPF_SOCKET_OPS_BASE_RTT support to tcp_nv

TCP_NV will try to get the base RTT from a socket_ops BPF program if one
is loaded. NV will then use the base RTT to bound its min RTT (its
notion of the base RTT). It uses the base RTT as an upper bound and 80%
of the base RTT as its lower bound.

In other words, NV will consider filtered RTTs larger than base RTT as a
sign of congestion. As a result, there is no minRTT inflation when there
is a lot of congestion. For example, in a DC where the RTTs are less
than 40us when there is no congestion, a base RTT value of 80us improves
the performance of NV. The difference between the uncongested RTT and
the base RTT provided represents how much queueing we are willing to
have (in practice it can be higher).

NV has been tunned to reduce congestion when there are many flows at the
cost of one flow not achieving full bandwith utilization. When a
reasonable base RTT is provided, one NV flow can now fully utilize the
full bandwidth. In addition, the performance is also improved when there
are many flows.

In the following examples the NV results are using a kernel with this
patch set (i.e. both NV results are using the new nv_loss_dec_factor).

With one host sending to another host and only one flow the
goodputs are:
  Cubic: 9.3 Gbps, NV: 5.5 Gbps, NV (baseRTT=80us): 9.2 Gbps

With 2 hosts sending to one host (1 flow per host, the goodput per flow
is:
  Cubic: 4.6 Gbps, NV: 4.5 Gbps, NV (baseRTT=80us)L 4.6 Gbps

But the RTTs seen by a ping process in the sender is:
  Cubic: 3.3ms  NV: 97us,  NV (baseRTT=80us): 146us

With a lot of flows things look even better for NV with baseRTT. Here we
have 3 hosts sending to one host. Each sending host has 6 flows: 1
stream, 4x1MB RPC, 1x10KB RPC. Cubic, NV and NV with baseRTT all fully
utilize the full available bandwidth. However, the distribution of
bandwidth among the flows is very different. For the 10KB RPC flow:
  Cubic: 27Mbps, NV: 111Mbps, NV (baseRTT=80us): 222Mbps

The 99% latencies for the 10KB flows are:
  Cubic: 26ms,  NV: 1ms,  NV (baseRTT=80us): 500us

The RTT seen by a ping process at the senders:
  Cubic: 3.2ms  NV: 720us,  NV (baseRTT=80us): 330us

Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 38 insertions, 2 deletions

diff --git a/net/ipv4/tcp_nv.c b/net/ipv4/tcp_nv.c
index 1ff73982e28c..a978e3fa71ec 100644
--- a/net/ipv4/tcp_nv.c
+++ b/net/ipv4/tcp_nv.c
@@ -39,7 +39,7 @@
  * nv_cong_dec_mult	Decrease cwnd by X% (30%) of congestion when detected
  * nv_ssthresh_factor	On congestion set ssthresh to this * <desired cwnd> / 8
  * nv_rtt_factor	RTT averaging factor
- * nv_loss_dec_factor	Decrease cwnd by this (50%) when losses occur
+ * nv_loss_dec_factor	Decrease cwnd to this (80%) when losses occur
  * nv_dec_eval_min_calls	Wait this many RTT measurements before dec cwnd
  * nv_inc_eval_min_calls	Wait this many RTT measurements before inc cwnd
  * nv_ssthresh_eval_min_calls	Wait this many RTT measurements before stopping
@@ -61,7 +61,7 @@ static int nv_min_cwnd __read_mostly = 2;
 static int nv_cong_dec_mult __read_mostly = 30 * 128 / 100; /* = 30% */
 static int nv_ssthresh_factor __read_mostly = 8; /* = 1 */
 static int nv_rtt_factor __read_mostly = 128; /* = 1/2*old + 1/2*new */
-static int nv_loss_dec_factor __read_mostly = 512; /* => 50% */
+static int nv_loss_dec_factor __read_mostly = 819; /* => 80% */
 static int nv_cwnd_growth_rate_neg __read_mostly = 8;
 static int nv_cwnd_growth_rate_pos __read_mostly; /* 0 => fixed like Reno */
 static int nv_dec_eval_min_calls __read_mostly = 60;
@@ -101,6 +101,11 @@ struct tcpnv {
 	u32 nv_last_rtt;	/* last rtt */
 	u32 nv_min_rtt;		/* active min rtt. Used to determine slope */
 	u32 nv_min_rtt_new;	/* min rtt for future use */
+	u32 nv_base_rtt;        /* If non-zero it represents the threshold for
+				 * congestion */
+	u32 nv_lower_bound_rtt; /* Used in conjunction with nv_base_rtt. It is
+				 * set to 80% of nv_base_rtt. It helps reduce
+				 * unfairness between flows */
 	u32 nv_rtt_max_rate;	/* max rate seen during current RTT */
 	u32 nv_rtt_start_seq;	/* current RTT ends when packet arrives
 				 * acking beyond nv_rtt_start_seq */
@@ -132,9 +137,24 @@ static inline void tcpnv_reset(struct tcpnv *ca, struct sock *sk)
 static void tcpnv_init(struct sock *sk)
 {
 	struct tcpnv *ca = inet_csk_ca(sk);
+	int base_rtt;
 
 	tcpnv_reset(ca, sk);
 
+	/* See if base_rtt is available from socket_ops bpf program.
+	 * It is meant to be used in environments, such as communication
+	 * within a datacenter, where we have reasonable estimates of
+	 * RTTs
+	 */
+	base_rtt = tcp_call_bpf(sk, BPF_SOCK_OPS_BASE_RTT);
+	if (base_rtt > 0) {
+		ca->nv_base_rtt = base_rtt;
+		ca->nv_lower_bound_rtt = (base_rtt * 205) >> 8; /* 80% */
+	} else {
+		ca->nv_base_rtt = 0;
+		ca->nv_lower_bound_rtt = 0;
+	}
+
 	ca->nv_allow_cwnd_growth = 1;
 	ca->nv_min_rtt_reset_jiffies = jiffies + 2 * HZ;
 	ca->nv_min_rtt = NV_INIT_RTT;
@@ -144,6 +164,19 @@ static void tcpnv_init(struct sock *sk)
 	ca->cwnd_growth_factor = 0;
 }
 
+/* If provided, apply upper (base_rtt) and lower (lower_bound_rtt)
+ * bounds to RTT.
+ */
+inline u32 nv_get_bounded_rtt(struct tcpnv *ca, u32 val)
+{
+	if (ca->nv_lower_bound_rtt > 0 && val < ca->nv_lower_bound_rtt)
+		return ca->nv_lower_bound_rtt;
+	else if (ca->nv_base_rtt > 0 && val > ca->nv_base_rtt)
+		return ca->nv_base_rtt;
+	else
+		return val;
+}
+
 static void tcpnv_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
@@ -265,6 +298,9 @@ static void tcpnv_acked(struct sock *sk, const struct ack_sample *sample)
 	if (ca->nv_eval_call_cnt < 255)
 		ca->nv_eval_call_cnt++;
 
+	/* Apply bounds to rtt. Only used to update min_rtt */
+	avg_rtt = nv_get_bounded_rtt(ca, avg_rtt);
+
 	/* update min rtt if necessary */
 	if (avg_rtt < ca->nv_min_rtt)
 		ca->nv_min_rtt = avg_rtt;