This is a request for comments. TCP-NV (New Vegas) is a major update to TCP-Vegas. An earlier version of NV was presented at 2010's LPC (slides). It is a delayed based congestion avoidance for the data center. This version has been tested within a 10G rack where the HW RTTs are 20-50us.
A description of TCP-NV, including implementation and experimental results, can be found at: http://www.brakmo.org/networking/tcp-nv/TCPNV.html The current version includes many module parameters to support experimentation with the parameters. Signed-off-by: Lawrence Brakmo <bra...@fb.com> --- net/ipv4/Kconfig | 16 ++ net/ipv4/Makefile | 1 + net/ipv4/tcp_nv.c | 479 ++++++++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 496 insertions(+) create mode 100644 net/ipv4/tcp_nv.c diff --git a/net/ipv4/Kconfig b/net/ipv4/Kconfig index 6fb3c90..f11f2f8 100644 --- a/net/ipv4/Kconfig +++ b/net/ipv4/Kconfig @@ -539,6 +539,22 @@ config TCP_CONG_VEGAS window. TCP Vegas should provide less packet loss, but it is not as aggressive as TCP Reno. +config TCP_CONG_NV + tristate "TCP NV" + default n + ---help--- + TCP NV is a follow up to TCP Vegas. It has been modified to deal with + 10G networks, measurement noise introduced by LRO, GRO and interrupt + coalescence. In addition, it will decrease its cwnd multiplicatively + instead of linearly. + + Note that in general congestion avoidance (cwnd decreased when # packets + queued grows) cannot coexist with congestion control (cwnd decreased only + when there is packet loss) due to fairness issues. One scenario when they + can coexist safely is when the CA flows have RTTs << CC flows RTTs. + + For further details see http://www.brakmo.org/networking/tcp-nv/ + config TCP_CONG_SCALABLE tristate "Scalable TCP" default n diff --git a/net/ipv4/Makefile b/net/ipv4/Makefile index efc43f3..06f335f 100644 --- a/net/ipv4/Makefile +++ b/net/ipv4/Makefile @@ -50,6 +50,7 @@ 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-$(CONFIG_TCP_CONG_NV) += tcp_nv.o obj-$(CONFIG_TCP_CONG_VENO) += tcp_veno.o obj-$(CONFIG_TCP_CONG_SCALABLE) += tcp_scalable.o obj-$(CONFIG_TCP_CONG_LP) += tcp_lp.o diff --git a/net/ipv4/tcp_nv.c b/net/ipv4/tcp_nv.c new file mode 100644 index 0000000..c4379b8 --- /dev/null +++ b/net/ipv4/tcp_nv.c @@ -0,0 +1,479 @@ +/* + * TCP NV: TCP with Congestion Avoidance + * + * TCP-NV is a successor of TCP-Vegas that has been developed to + * deal with the issues that occur in modern networks. + * Like TCP-Vegas, TCP-NV supports true congestion avoidance, + * the ability to detect congestion before packet losses occur. + * When congestion (queue buildup) starts to occur, TCP-NV + * predicts what the cwnd size should be for the current + * throughput and it reduces the cwnd proportionally to + * the difference between the current cwnd and the predicted cwnd. + * TCP-NV behaves like Reno when no congestion is detected, or when + * recovering from packet losses. + * + * TODO: + * 1) Add option to not decrease cwnd on losses below certain level + * 2) Add mechanism to deal with reverse congestion. + */ + +#include <linux/mm.h> +#include <linux/module.h> +#include <linux/math64.h> +#include <net/tcp.h> +#include <linux/inet_diag.h> + +/* TCP NV parameters */ +static int nv_enable __read_mostly = 1; +static int nv_pad __read_mostly = 10; +static int nv_pad_buffer __read_mostly = 2; +static int nv_reset_period __read_mostly = 5; +static int nv_min_cwnd = 10; +static int nv_dec_eval_min_calls = 100; +static int nv_ssthresh_eval_min_calls = 30; +static int nv_rtt_min_cnt = 2; +static int nv_cong_decrease_mult = 30*128/100; +static int nv_ssthresh_factor = 8; +static int nv_rtt_factor = 128; +static int nv_rtt_cnt_dec_delta = 20; /* dec cwnd by this many RTTs */ +static int nv_dec_factor = 5; /* actual value is factor/8 */ +static int nv_loss_dec_factor = 820; /* on loss reduce cwnd by 20% */ +static int nv_cwnd_growth_factor = 2; /* larger => cwnd grows slower */ + +module_param(nv_pad, int, 0644); +MODULE_PARM_DESC(nv_pad, "extra packets above congestion level"); +module_param(nv_pad_buffer, int, 0644); +MODULE_PARM_DESC(nv_pad_buffer, "no growth buffer zone"); +module_param(nv_reset_period, int, 0644); +MODULE_PARM_DESC(nv_reset_period, "nv_min_rtt reset period (secs)"); +module_param(nv_min_cwnd, int, 0644); +MODULE_PARM_DESC(nv_min_cwnd, "NV will not decrease cwnd below this value" + " without losses"); +module_param(nv_dec_eval_min_calls, int, 0644); +MODULE_PARM_DESC(nv_dec_eval_min_calls, "Wait for this many data points " + "before declaring congestion (< 256)"); +module_param(nv_ssthresh_eval_min_calls, int, 0644); +MODULE_PARM_DESC(nv_ssthresh_eval_min_calls, "Wait for this many data points " + "before declaring congestion during initial slow-start"); +module_param(nv_rtt_min_cnt, int, 0644); +MODULE_PARM_DESC(nv_rtt_min_cnt, "Wait for this many RTTs before declaring" + " congestion (<64)"); +module_param(nv_cong_decrease_mult, int, 0644); +MODULE_PARM_DESC(nv_cong_decrease_mult, "Congestion decrease factor"); +module_param(nv_ssthresh_factor, int, 0644); +MODULE_PARM_DESC(nv_ssthresh_factor, "ssthresh factor"); +module_param(nv_rtt_factor, int, 0644); +MODULE_PARM_DESC(nv_rtt_factor, "rtt averaging factor (0-256)"); +module_param(nv_rtt_cnt_dec_delta, int, 0644); +MODULE_PARM_DESC(nv_rtt_cnt_dec_delta, "decrease cwnd for this many RTTs " + "every 100 RTTs"); +module_param(nv_dec_factor, int, 0644); +MODULE_PARM_DESC(nv_dec_factor, "decrease cwnd every ~192 RTTS by factor/8"); +module_param(nv_loss_dec_factor, int, 0644); +MODULE_PARM_DESC(nv_loss_dec_factor, "on loss new cwnd = cwnd * this / 1024"); +module_param(nv_cwnd_growth_factor, int, 0644); +MODULE_PARM_DESC(nv_cwnd_growth_factor, "larger => cwnd grows slower"); + +/* TCP NV Parameters */ +struct tcpnv { + unsigned long nv_min_rtt_reset_jiffies; /* when to switch to + * nv_min_rtt_new */ + u32 cnt; /* increase cwnd by 1 after ACKs */ + u32 loss_cwnd; /* cwnd at last loss */ + u8 nv_enable:1, + nv_allow_cwnd_growth:1, /* whether cwnd can grow */ + nv_rtt_cnt:6; /* RTTs without making ca decision */ + u8 nv_eval_call_cnt;/* call count since last eval */ + u8 nv_min_cwnd; /* nv won't make a ca decision if cwnd is + * smaller than this. It may grow to handle + * TSO, LRO and interrupt coalescence because + * with these a small cwnd cannot saturate + * the link. Note that this is different from + * sysctl_tcp_nv_min_cwnd */ + u8 available; + 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_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 */ + u32 nv_last_snd_una; /* Previous value of tp->snd_una. It is + * used to determine bytes acked since last + * call to bictcp_acked */ + u32 nv_no_cong_cnt; /* Consecutive no congestion decisions */ + u32 nv_rtt_cnt_dec; /* RTTs since last temporary cwnd decrease */ +}; + +#define NV_INIT_RTT 0xffffffff +#define NV_MIN_CWND 4 +#define NV_MIN_CWND_GROW 2 +#define NV_TSO_CWND_BOUND 80 + +static inline void tcpnv_reset(struct tcpnv *ca, struct sock *sk) +{ + struct tcp_sock *tp = tcp_sk(sk); + + ca->loss_cwnd = 0; + ca->nv_no_cong_cnt = 0; + ca->cnt = 0; + ca->nv_rtt_cnt = 0; + ca->nv_rtt_cnt_dec = 0; + ca->nv_allow_cwnd_growth = 1; + ca->nv_last_rtt = 0; + ca->nv_rtt_max_rate = 0; + ca->nv_rtt_start_seq = tp->snd_una; + ca->nv_eval_call_cnt = 0; + ca->nv_last_snd_una = tp->snd_una; +} + +static void tcpnv_init(struct sock *sk) +{ + struct tcpnv *ca = inet_csk_ca(sk); + + tcpnv_reset(ca, sk); + + ca->nv_min_rtt_reset_jiffies = jiffies + 2*HZ; + ca->nv_min_rtt = NV_INIT_RTT; + ca->nv_min_rtt_new = NV_INIT_RTT; + ca->nv_enable = nv_enable; + ca->nv_min_cwnd = NV_MIN_CWND; + if (nv_dec_eval_min_calls > 255) + nv_dec_eval_min_calls = 255; + if (nv_rtt_min_cnt > 63) + nv_rtt_min_cnt = 63; +} + +static void tcpnv_cong_avoid(struct sock *sk, u32 ack, u32 acked) +{ + struct tcp_sock *tp = tcp_sk(sk); + struct tcpnv *ca = inet_csk_ca(sk); + + if (!tcp_is_cwnd_limited(sk)) + return; + + /* Only grow cwnd if NV has not detected congestion */ + if (nv_enable && ca->nv_enable && !ca->nv_allow_cwnd_growth) + return; + + if (tcp_in_slow_start(tp)) { + acked = tcp_slow_start(tp, acked); + if (!acked) + return; + } + if (ca->cnt == 0 || !(nv_enable || ca->nv_enable)) + ca->cnt = tp->snd_cwnd; + + tcp_cong_avoid_ai(tp, ca->cnt, acked); +} + +static u32 tcpnv_recalc_ssthresh(struct sock *sk) +{ + const struct tcp_sock *tp = tcp_sk(sk); + struct tcpnv *ca = inet_csk_ca(sk); + + ca->loss_cwnd = tp->snd_cwnd; + return max((tp->snd_cwnd * nv_loss_dec_factor) >> 10, 2U); +} + +static u32 tcpnv_undo_cwnd(struct sock *sk) +{ + struct tcpnv *ca = inet_csk_ca(sk); + + return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd); +} + +static void tcpnv_state(struct sock *sk, u8 new_state) +{ + struct tcpnv *ca = inet_csk_ca(sk); + + if (new_state == TCP_CA_Open) { + ca->nv_enable = 1; + tcpnv_reset(ca, sk); + } else if (new_state == TCP_CA_Loss) { + ca->nv_enable = 0; + } +} + +/* Do congestion avoidance calculaitons for TCP-NV + */ +static void tcpnv_acked(struct sock *sk, struct ack_sample *sample) +{ + const struct inet_connection_sock *icsk = inet_csk(sk); + struct tcp_sock *tp = tcp_sk(sk); + struct tcpnv *ca = inet_csk_ca(sk); + unsigned long now = jiffies; + s64 rate64 = 0; + u32 rate, max_win, cwnd_by_slope; + u32 avg_rtt; + u32 bytes_acked = 0; + + /* Some calls are for duplicates without timetamps */ + if (sample->rtt_us < 0) + return; + + /* If not in TCP_CA_Open state, skip. */ + if (icsk->icsk_ca_state != TCP_CA_Open) + return; + + /* If NV mode is not enabled, behave like Reno */ + if (!nv_enable || !ca->nv_enable) { + ca->nv_allow_cwnd_growth = 1; + return; + } + + bytes_acked = tp->snd_una - ca->nv_last_snd_una; + ca->nv_last_snd_una = tp->snd_una; + + if (sample->in_flight == 0) + return; + + /* Calculate moving average of RTT */ + if (nv_rtt_factor > 0) { + if (ca->nv_last_rtt > 0) { + avg_rtt = (((u64)sample->rtt_us) * nv_rtt_factor + + ((u64)ca->nv_last_rtt) + * (256 - nv_rtt_factor)) >> 8; + } else { + avg_rtt = sample->rtt_us; + ca->nv_min_rtt = avg_rtt << 1; + } + ca->nv_last_rtt = avg_rtt; + } else { + avg_rtt = sample->rtt_us; + } + + /* rate in 100's bits per second */ + rate64 = ((u64)sample->in_flight) * 8000000; + rate = (u32)div64_u64(rate64, (u64)(avg_rtt*100)); + + /* Remember the maximum rate seen during this RTT + * Note: It may be more than one RTT. This function should be + * called at least nv_dec_eval_min_calls times. + */ + if (ca->nv_rtt_max_rate < rate) + ca->nv_rtt_max_rate = rate; + + /* We have valid information, increment counter */ + if (ca->nv_eval_call_cnt < 255) + ca->nv_eval_call_cnt++; + + /* update min rtt if necessary */ + if (avg_rtt < ca->nv_min_rtt) + ca->nv_min_rtt = avg_rtt; + + /* update future min_rtt if necessary */ + if (avg_rtt < ca->nv_min_rtt_new) + ca->nv_min_rtt_new = avg_rtt; + + /* nv_min_rtt is updated with the minimum (possibley averaged) rtt + * seen in the last sysctl_tcp_nv_reset_period seconds (i.e. a + * warm reset). This new nv_min_rtt will be continued to be updated + * and be used for another sysctl_tcp_nv_reset_period seconds, + * when it will be updated again. + * In practice we introduce some randomness, so the actual period used + * is chosen randomly from the range: + * [sysctl_tcp_nv_reset_period*3/4, sysctl_tcp_nv_reset_period*5/4) + */ + if (time_after_eq(now, ca->nv_min_rtt_reset_jiffies)) { + unsigned char rand; + ca->nv_min_rtt = ca->nv_min_rtt_new; + ca->nv_min_rtt_new = NV_INIT_RTT; + get_random_bytes(&rand, 1); + ca->nv_min_rtt_reset_jiffies = + now + ((nv_reset_period*(384 + rand)*HZ)>>9); + /* Every so often we decrease nv_min_cwnd in case previous + * value is no longer accurate. + */ + ca->nv_min_cwnd = max(ca->nv_min_cwnd/2, NV_MIN_CWND); + } + + /* Once per RTT check if we need to do congestion avoidance */ + if (before(ca->nv_rtt_start_seq, tp->snd_una)) { + ca->nv_rtt_start_seq = tp->snd_nxt; + if (ca->nv_rtt_cnt < 63) + /* Increase counter for RTTs without CA decision */ + ca->nv_rtt_cnt++; + if (ca->nv_rtt_cnt_dec < 255) + /* Increase counter for temporary cwnd decrease */ + ca->nv_rtt_cnt_dec++; + + /* If this function is only called once within an RTT + * the cwnd is probably too small (in some cases due to + * tso, lro or interrupt coalescence), so we increase + * nv_min_cwnd. + */ + if (ca->nv_eval_call_cnt == 1 + && bytes_acked >= (ca->nv_min_cwnd - 1) * tp->mss_cache + && ca->nv_min_cwnd < (NV_TSO_CWND_BOUND + 1) + && ca->nv_rtt_cnt_dec < 192) { + ca->nv_min_cwnd = min(ca->nv_min_cwnd + + NV_MIN_CWND_GROW, + NV_TSO_CWND_BOUND + 1); + ca->nv_rtt_start_seq = tp->snd_nxt + + ca->nv_min_cwnd*tp->mss_cache; + ca->nv_eval_call_cnt = 0; + ca->nv_allow_cwnd_growth = 1; + return; + } + + /* Every 192 to 320 RTTs decrease cwnd to get better min RTT + * measurement. In practice we accomplish this by initializing + * nv_rtt_cnd_dec randomly form the range [0, 128) and + * stopping at 320. + * We keep the value low for nv_rtt_cnt_dec_delta RTTs and then + * we restore cwnd to its previous value (by setting + * ssthresh to the previous value). + */ + if (ca->nv_rtt_cnt_dec == 320) { + /* decrease cwnd and ssthresh */ + tp->snd_cwnd = + max((unsigned int)nv_min_cwnd, + ((tp->snd_cwnd * nv_dec_factor) >> 3)); + tp->snd_ssthresh = + max(tp->snd_cwnd, + ((tp->snd_ssthresh * nv_dec_factor) >> 3)); + ca->nv_allow_cwnd_growth = 0; + return; + } else if (ca->nv_rtt_cnt_dec > 320) { + if (ca->nv_rtt_cnt_dec - 320 >= nv_rtt_cnt_dec_delta) { + /* Restore ssthresh to restore cwnd */ + unsigned char rand; + get_random_bytes(&rand, 1); + ca->nv_rtt_cnt_dec = rand >> 1; + tp->snd_ssthresh = (tp->snd_ssthresh << 3) + / nv_dec_factor; + ca->nv_allow_cwnd_growth = 1; + ca->nv_no_cong_cnt = 0; + } + return; + } + + /* Find the ideal cwnd for current rate from slope + * slope = 80000.0 * mss / nv_min_rtt + * cwnd_by_slope = nv_rtt_max_rate / slope + */ + cwnd_by_slope = (u32) + div64_u64(((u64)ca->nv_rtt_max_rate) * ca->nv_min_rtt, + (u64)(80000 * tp->mss_cache)); + max_win = cwnd_by_slope + nv_pad; + + /* If cwnd > max_win, decrease cwnd + * if cwnd < max_win, grow cwnd + * else leave the same + */ + if (tp->snd_cwnd > max_win) { + /* there is congestion, check that it is ok + * to make a CA decision + * 1. We should have at least nv_dec_eval_min_calls + * data points before making a CA decision + * 2. We only make a congesion decision after + * nv_rtt_min_cnt RTTs + */ + if (ca->nv_rtt_cnt < nv_rtt_min_cnt) + return; + else if (tp->snd_ssthresh == TCP_INFINITE_SSTHRESH) { + if (ca->nv_eval_call_cnt < + nv_ssthresh_eval_min_calls) + return; + } else if (ca->nv_eval_call_cnt < + nv_dec_eval_min_calls) { + return; + } + + /* We have enough data to determine we are congested */ + ca->nv_allow_cwnd_growth = 0; + tp->snd_ssthresh = + (nv_ssthresh_factor * max_win) >> 3; + if (tp->snd_cwnd - max_win > 2) { + /* gap > 2, we do exponential cwnd decrease */ + int dec; + dec = max(2U, ((tp->snd_cwnd - max_win) * + nv_cong_decrease_mult) >> 7); + tp->snd_cwnd -= dec; + } else if (nv_cong_decrease_mult > 0) { + tp->snd_cwnd = max_win; + } + ca->cnt = tp->snd_cwnd; + ca->nv_no_cong_cnt = 0; + } else if (tp->snd_cwnd <= max_win - nv_pad_buffer) { + /* We allow growth of cwnd every RTT since we would + * have grown even if we waited (just slower) + */ + ca->nv_allow_cwnd_growth = 1; + ca->nv_no_cong_cnt++; + if (nv_cwnd_growth_factor > 0 && + ca->nv_no_cong_cnt > nv_cwnd_growth_factor) { + ca->cnt = max(ca->cnt >> 1, (u32) 4); + ca->nv_no_cong_cnt = 0; + } + } else { + ca->nv_allow_cwnd_growth = 0; + } + + /* update state */ + ca->nv_eval_call_cnt = 0; + ca->nv_rtt_cnt = 0; + ca->nv_rtt_max_rate = 0; + + /* Don't want to make cwnd < nv_min_cwnd + * (it wasn't before, if it is now is because nv + * decreased it). + */ + if (tp->snd_cwnd < nv_min_cwnd) + tp->snd_cwnd = nv_min_cwnd; + + } +} + +/* Extract info for Tcp socket info provided via netlink */ +size_t tcpnv_get_info(struct sock *sk, u32 ext, int *attr, + union tcp_cc_info *info) +{ + const struct tcpnv *ca = inet_csk_ca(sk); + + if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) { + info->vegas.tcpv_enabled = ca->nv_enable + && nv_enable; + info->vegas.tcpv_rttcnt = ca->nv_rtt_cnt; + info->vegas.tcpv_rtt = ca->nv_last_rtt; + info->vegas.tcpv_minrtt = ca->nv_min_rtt; + + *attr = INET_DIAG_VEGASINFO; + return sizeof(struct tcpvegas_info); + } + return 0; +} +EXPORT_SYMBOL_GPL(tcpnv_get_info); + +static struct tcp_congestion_ops tcpnv __read_mostly = { + .init = tcpnv_init, + .ssthresh = tcpnv_recalc_ssthresh, + .cong_avoid = tcpnv_cong_avoid, + .set_state = tcpnv_state, + .undo_cwnd = tcpnv_undo_cwnd, + .pkts_acked = tcpnv_acked, + .get_info = tcpnv_get_info, + + .owner = THIS_MODULE, + .name = "nv", +}; + +static int __init tcpnv_register(void) +{ + BUILD_BUG_ON(sizeof(struct tcpnv) > ICSK_CA_PRIV_SIZE); + + return tcp_register_congestion_control(&tcpnv); +} + +static void __exit tcpnv_unregister(void) +{ + tcp_unregister_congestion_control(&tcpnv); +} + +module_init(tcpnv_register); +module_exit(tcpnv_unregister); + +MODULE_AUTHOR("Lawrence Brakmo"); +MODULE_LICENSE("GPL"); +MODULE_DESCRIPTION("TCP NV"); +MODULE_VERSION("1.0"); -- 1.8.1 -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
