Yonhhong, thank you for the feedbac. Repies inlined. I will submit a new
revision later today.
On 7/1/19, 2:04 PM, "Yonghong Song" <y...@fb.com> wrote:
On 6/28/19 12:41 PM, brakmo wrote:
> Adds support for fq's Earliest Departure Time to HBM (Host Bandwidth
> Manager). Includes a new BPF program supporting EDT, and also updates
> corresponding programs.
>
> It will drop packets with an EDT of more than 500us in the future
> unless the packet belongs to a flow with less than 2 packets in flight.
> This is done so each flow has at least 2 packets in flight, so they
> will not starve, and also to help prevent delayed ACK timeouts.
>
> It will also work with ECN enabled traffic, where the packets will be
> CE marked if their EDT is more than 50us in the future.
>
> The table below shows some performance numbers. The flows are back to
> back RPCS. One server sending to another, either 2 or 4 flows.
> One flow is a 10KB RPC, the rest are 1MB RPCs. When there are more
> than one flow of a given RPC size, the numbers represent averages.
>
> The rate limit applies to all flows (they are in the same cgroup).
> Tests ending with "-edt" ran with the new BPF program supporting EDT.
> Tests ending with "-hbt" ran on top HBT qdisc with the specified rate
> (i.e. no HBM). The other tests ran with the HBM BPF program included
> in the HBM patch-set.
>
> EDT has limited value when using DCTCP, but it helps in many cases when
> using Cubic. It usually achieves larger link utilization and lower
> 99% latencies for the 1MB RPCs.
> HBM ends up queueing a lot of packets with its default parameter values,
> reducing the goodput of the 10KB RPCs and increasing their latency. Also,
> the RTTs seen by the flows are quite large.
>
> Aggr 10K 10K 10K 1MB 1MB 1MB
> Limit rate drops RTT rate P90 P99 rate P90 P99
> Test rate Flows Mbps % us Mbps us us Mbps ms ms
> -------- ---- ----- ---- ----- --- ---- ---- ---- ---- ---- ----
> cubic 1G 2 904 0.02 108 257 511 539 647 13.4 24.5
> cubic-edt 1G 2 982 0.01 156 239 656 967 743 14.0 17.2
> dctcp 1G 2 977 0.00 105 324 408 744 653 14.5 15.9
> dctcp-edt 1G 2 981 0.01 142 321 417 811 660 15.7 17.0
> cubic-htb 1G 2 919 0.00 1825 40 2822 4140 879 9.7 9.9
>
> cubic 200M 2 155 0.30 220 81 532 655 74 283 450
> cubic-edt 200M 2 188 0.02 222 87 1035 1095 101 84 85
> dctcp 200M 2 188 0.03 111 77 912 939 111 76 325
> dctcp-edt 200M 2 188 0.03 217 74 1416 1738 114 76 79
> cubic-htb 200M 2 188 0.00 5015 8 14ms 15ms 180 48 50
>
> cubic 1G 4 952 0.03 110 165 516 546 262 38 154
> cubic-edt 1G 4 973 0.01 190 111 1034 1314 287 65 79
> dctcp 1G 4 951 0.00 103 180 617 905 257 37 38
> dctcp-edt 1G 4 967 0.00 163 151 732 1126 272 43 55
> cubic-htb 1G 4 914 0.00 3249 13 7ms 8ms 300 29 34
>
> cubic 5G 4 4236 0.00 134 305 490 624 1310 10 17
> cubic-edt 5G 4 4865 0.00 156 306 425 759 1520 10 16
> dctcp 5G 4 4936 0.00 128 485 221 409 1484 7 9
> dctcp-edt 5G 4 4924 0.00 148 390 392 623 1508 11 26
>
> v1 -> v2: Incorporated Andrii's suggestions
>
> Signed-off-by: Lawrence Brakmo <bra...@fb.com>
> ---
> samples/bpf/Makefile | 2 +
> samples/bpf/do_hbm_test.sh | 22 ++---
> samples/bpf/hbm.c | 18 +++-
> samples/bpf/hbm_edt_kern.c | 171 +++++++++++++++++++++++++++++++++++++
> samples/bpf/hbm_kern.h | 40 +++++++--
> 5 files changed, 234 insertions(+), 19 deletions(-)
> create mode 100644 samples/bpf/hbm_edt_kern.c
>
> diff --git a/samples/bpf/Makefile b/samples/bpf/Makefile
> index 0917f8cf4fab..35640414ebb3 100644
> --- a/samples/bpf/Makefile
> +++ b/samples/bpf/Makefile
> @@ -168,6 +168,7 @@ always += task_fd_query_kern.o
> always += xdp_sample_pkts_kern.o
> always += ibumad_kern.o
> always += hbm_out_kern.o
> +always += hbm_edt_kern.o
>
> KBUILD_HOSTCFLAGS += -I$(objtree)/usr/include
> KBUILD_HOSTCFLAGS += -I$(srctree)/tools/lib/bpf/
> @@ -272,6 +273,7 @@ $(src)/*.c: verify_target_bpf $(LIBBPF)
> $(obj)/tracex5_kern.o: $(obj)/syscall_nrs.h
> $(obj)/hbm_out_kern.o: $(src)/hbm.h $(src)/hbm_kern.h
> $(obj)/hbm.o: $(src)/hbm.h
> +$(obj)/hbm_edt_kern.o: $(src)/hbm.h $(src)/hbm_kern.h
>
> # asm/sysreg.h - inline assembly used by it is incompatible with llvm.
> # But, there is no easy way to fix it, so just exclude it since it is
> diff --git a/samples/bpf/do_hbm_test.sh b/samples/bpf/do_hbm_test.sh
> index e48b047d4646..ffe4c0607341 100755
> --- a/samples/bpf/do_hbm_test.sh
> +++ b/samples/bpf/do_hbm_test.sh
> @@ -14,7 +14,7 @@ Usage() {
> echo "loads. The output is the goodput in Mbps (unless -D was used)."
> echo ""
> echo "USAGE: $name [out] [-b=<prog>|--bpf=<prog>] [-c=<cc>|--cc=<cc>]"
> - echo " [-D] [-d=<delay>|--delay=<delay>] [--debug] [-E]"
> + echo " [-D] [-d=<delay>|--delay=<delay>] [--debug] [-E]
[--edt]"
> echo " [-f=<#flows>|--flows=<#flows>] [-h]
[-i=<id>|--id=<id >]"
> echo " [-l] [-N] [--no_cn] [-p=<port>|--port=<port>] [-P]"
> echo " [-q=<qdisc>] [-R] [-s=<server>|--server=<server]"
> @@ -30,6 +30,7 @@ Usage() {
> echo " other detailed information. This
information is"
> echo " test dependent (i.e. iperf3 or netperf)."
> echo " -E enable ECN (not required for dctcp)"
> + echo " --edt use fq's Earliest Departure Time (requires
fq)"
> echo " -f or --flows number of concurrent flows (default=1)"
> echo " -i or --id cgroup id (an integer, default is 1)"
> echo " -N use netperf instead of iperf3"
> @@ -130,13 +131,12 @@ processArgs () {
> details=1
> ;;
> -E)
> - ecn=1
> + ecn=1
> + ;;
> + --edt)
> + flags="$flags --edt"
> + qdisc="fq"
> ;;
> - # Support for upcomming fq Early Departure Time egress rate limiting
> - #--edt)
> - # prog="hbm_out_edt_kern.o"
> - # qdisc="fq"
> - # ;;
> -f=*|--flows=*)
> flows="${i#*=}"
> ;;
> @@ -228,8 +228,8 @@ if [ "$netem" -ne "0" ] ; then
> tc qdisc del dev lo root > /dev/null 2>&1
> tc qdisc add dev lo root netem delay $netem\ms > /dev/null 2>&1
> elif [ "$qdisc" != "" ] ; then
> - tc qdisc del dev lo root > /dev/null 2>&1
> - tc qdisc add dev lo root $qdisc > /dev/null 2>&1
> + tc qdisc del dev eth0 root > /dev/null 2>&1
> + tc qdisc add dev eth0 root $qdisc > /dev/null 2>&1
> fi
>
> n=0
> @@ -399,7 +399,9 @@ fi
> if [ "$netem" -ne "0" ] ; then
> tc qdisc del dev lo root > /dev/null 2>&1
> fi
> -
> +if [ "$qdisc" != "" ] ; then
> + tc qdisc del dev eth0 root > /dev/null 2>&1
> +fi
> sleep 2
>
> hbmPid=`ps ax | grep "hbm " | grep --invert-match "grep" | awk '{ print
$1 }'`
> diff --git a/samples/bpf/hbm.c b/samples/bpf/hbm.c
> index b905b32ff185..e0fbab9bec83 100644
> --- a/samples/bpf/hbm.c
> +++ b/samples/bpf/hbm.c
> @@ -62,6 +62,7 @@ bool loopback_flag;
> bool debugFlag;
> bool work_conserving_flag;
> bool no_cn_flag;
> +bool edt_flag;
>
> static void Usage(void);
> static void read_trace_pipe2(void);
> @@ -372,9 +373,14 @@ static int run_bpf_prog(char *prog, int cg_id)
> fprintf(fout, "avg rtt:%d\n",
> (int)(qstats.sum_rtt / (qstats.pkts_total +
1)));
> // Average credit
> - fprintf(fout, "avg credit:%d\n",
> - (int)(qstats.sum_credit /
> - (1500 * ((int)qstats.pkts_total) + 1)));
> + if (edt_flag)
> + fprintf(fout, "avg credit_ms:%.03f\n",
> + (qstats.sum_credit /
> + (qstats.pkts_total + 1.0)) / 1000000.0);
> + else
> + fprintf(fout, "avg credit:%d\n",
> + (int)(qstats.sum_credit /
> + (1500 * ((int)qstats.pkts_total ) + 1)));
>
> // Return values stats
> for (k = 0; k < RET_VAL_COUNT; k++) {
> @@ -408,6 +414,7 @@ static void Usage(void)
> " Where:\n"
> " -o indicates egress direction (default)\n"
> " -d print BPF trace debug buffer\n"
> + " --edt use fq's Earliest Departure Time\n"
> " -l also limit flows using loopback\n"
> " -n <#> to create cgroup \"/hbm#\" and attach
prog\n"
> " Default is /hbm1\n"
> @@ -433,6 +440,7 @@ int main(int argc, char **argv)
> char *optstring = "iodln:r:st:wh";
> struct option loptions[] = {
> {"no_cn", 0, NULL, 1},
> + {"edt", 0, NULL, 2},
> {NULL, 0, NULL, 0}
> };
>
> @@ -441,6 +449,10 @@ int main(int argc, char **argv)
> case 1:
> no_cn_flag = true;
> break;
> + case 2:
> + prog = "hbm_edt_kern.o";
> + edt_flag = true;
> + break;
> case'o':
> break;
> case 'd':
> diff --git a/samples/bpf/hbm_edt_kern.c b/samples/bpf/hbm_edt_kern.c
> new file mode 100644
> index 000000000000..004a44a83e1e
> --- /dev/null
> +++ b/samples/bpf/hbm_edt_kern.c
> @@ -0,0 +1,171 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/* Copyright (c) 2019 Facebook
> + *
> + * This program is free software; you can redistribute it and/or
> + * modify it under the terms of version 2 of the GNU General Public
> + * License as published by the Free Software Foundation.
> + *
> + * Sample Host Bandwidth Manager (HBM) BPF program.
> + *
> + * A cgroup skb BPF egress program to limit cgroup output bandwidth.
> + * It uses a modified virtual token bucket queue to limit average
> + * egress bandwidth. The implementation uses credits instead of tokens.
> + * Negative credits imply that queueing would have happened (this is
> + * a virtual queue, so no queueing is done by it. However, queueing may
> + * occur at the actual qdisc (which is not used for rate limiting).
> + *
> + * This implementation uses 3 thresholds, one to start marking packets
and
> + * the other two to drop packets:
> + * CREDIT
> + * - <--------------------------|------------------------> +
> + * | | | 0
> + * | Large pkt |
> + * | drop thresh |
> + * Small pkt drop Mark threshold
> + * thresh
> + *
> + * The effect of marking depends on the type of packet:
> + * a) If the packet is ECN enabled and it is a TCP packet, then the
packet
> + * is ECN marked.
> + * b) If the packet is a TCP packet, then we probabilistically call
tcp_cwr
> + * to reduce the congestion window. The current implementation uses a
linear
> + * distribution (0% probability at marking threshold, 100% probability
> + * at drop threshold).
> + * c) If the packet is not a TCP packet, then it is dropped.
> + *
> + * If the credit is below the drop threshold, the packet is dropped. If
it
> + * is a TCP packet, then it also calls tcp_cwr since packets dropped by
> + * by a cgroup skb BPF program do not automatically trigger a call to
> + * tcp_cwr in the current kernel code.
> + *
> + * This BPF program actually uses 2 drop thresholds, one threshold
> + * for larger packets (>= 120 bytes) and another for smaller packets.
This
> + * protects smaller packets such as SYNs, ACKs, etc.
> + *
> + * The default bandwidth limit is set at 1Gbps but this can be changed by
> + * a user program through a shared BPF map. In addition, by default this
BPF
> + * program does not limit connections using loopback. This behavior can
be
> + * overwritten by the user program. There is also an option to calculate
> + * some statistics, such as percent of packets marked or dropped, which
> + * a user program, such as hbm, can access.
> + */
> +
> +#include "hbm_kern.h"
> +
> +SEC("cgroup_skb/egress")
> +int _hbm_out_cg(struct __sk_buff *skb)
> +{
> + signed long long delta = 0, delta_send;
"signed" is not needed here.
Removed
> + unsigned long long curtime, sendtime;
> + struct hbm_queue_stats *qsp = NULL;
> + unsigned int queue_index = 0;
> + bool congestion_flag = false;
> + bool ecn_ce_flag = false;
> + struct hbm_pkt_info pkti;
> + struct hbm_vqueue *qdp;
> + bool drop_flag = false;
> + bool cwr_flag = false;
> + int len = skb->len;
> + int rv = ALLOW_PKT;
> +
> + qsp = bpf_map_lookup_elem(&queue_stats, &queue_index);
> + if (qsp != NULL && !qsp->loopback && (skb->ifindex == 1))
The skb->ifindex = 1 is skipped. Why? Some comments will
be helpful to understand.
Added comment
> + return ALLOW_PKT;
> +
> + hbm_get_pkt_info(skb, &pkti);
> +
> + // We may want to account for the length of headers in len
> + // calculation, like ETH header + overhead, specially if it
> + // is a gso packet. But I am not doing it right now.
> +
> + qdp = bpf_get_local_storage(&queue_state, 0);
> + if (!qdp)
> + return ALLOW_PKT;
> + if (qdp->lasttime == 0)
> + hbm_init_edt_vqueue(qdp, 1024);
> +
> + curtime = bpf_ktime_get_ns();
> +
> + // Begin critical section
> + bpf_spin_lock(&qdp->lock);
> + delta = qdp->lasttime - curtime;
> + // bound bursts to 100us
> + if (delta < -BURST_SIZE_NS) {
> + // negative delta is a credit that allows bursts
> + qdp->lasttime = curtime - BURST_SIZE_NS;
> + delta = -BURST_SIZE_NS;
> + }
> + sendtime = qdp->lasttime;
> + delta_send = BYTES_TO_NS(len, qdp->rate);
> + qdp->lasttime += delta_send;
> + bpf_spin_unlock(&qdp->lock);
> + // End critical section
> +
> + // Set EDT of packet
> + skb->tstamp = sendtime;
> +
> + // Check if we should update rate
> + if (qsp != NULL && (qsp->rate * 128) != qdp->rate) {
> + qdp->rate = qsp->rate * 128;
> + bpf_printk("Updating rate: %d (1sec:%llu bits)\n",
> + (int)qdp->rate,
> + CREDIT_PER_NS(1000000000, qdp->rate) * 8);
How often this bpf_printk will fire at runtime?
If it fires relatively frequently, maybe worth to put it
under debug?
Removed
> + }
> +
> + // Set flags (drop, congestion, cwr)
> + // last packet will be sent in the future, bound latency
> + if (delta > DROP_THRESH_NS || (delta > LARGE_PKT_DROP_THRESH_NS &&
> + len > LARGE_PKT_THRESH)) {
> + drop_flag = true;
> + if (pkti.is_tcp && pkti.ecn == 0)
> + cwr_flag = true;
> + } else if (delta > MARK_THRESH_NS) {
> + if (pkti.is_tcp)
> + congestion_flag = true;
> + else
> + drop_flag = true;
> + }
> +
> + if (congestion_flag) {
> + if (bpf_skb_ecn_set_ce(skb)) {
> + ecn_ce_flag = true;
> + } else {
> + if (pkti.is_tcp) {
> + unsigned int rand = bpf_get_prandom_u32();
> +
> + if (delta >= MARK_THRESH_NS +
> + (rand % MARK_REGION_SIZE_NS)) {
> + // Do congestion control
> + cwr_flag = true;
> + }
> + } else if (len > LARGE_PKT_THRESH) {
> + // Problem if too many small packets?
> + drop_flag = true;
> + congestion_flag = false;
> + }
> + }
> + }
> +
> + if (pkti.is_tcp && drop_flag && pkti.packets_out <= 1) {
> + drop_flag = false;
> + cwr_flag = true;
> + congestion_flag = false;
> + }
> +
> + if (qsp != NULL && qsp->no_cn)
> + cwr_flag = false;
> +
> + hbm_update_stats(qsp, len, curtime, congestion_flag, drop_flag,
> + cwr_flag, ecn_ce_flag, &pkti, (int) delta);
> +
> + if (drop_flag) {
> + __sync_add_and_fetch(&(qdp->credit), len);
> + __sync_add_and_fetch(&(qdp->lasttime), -delta_send);
We have race here. Updating qdp->lasttime may happen concurrently with
the update in bpf_spin_lock region. Do we have any issue here?
Good catch. Should not be an issue as long as both updates are concurrent. I
changed the update within the spin_lock_region to also be atomic.
> + rv = DROP_PKT;
> + }
> +
> + if (cwr_flag)
> + rv |= CWR;
> + return rv;
> +}
> +char _license[] SEC("license") = "GPL";
> diff --git a/samples/bpf/hbm_kern.h b/samples/bpf/hbm_kern.h
> index be19cf1d5cd5..aa207a2eebbd 100644
> --- a/samples/bpf/hbm_kern.h
> +++ b/samples/bpf/hbm_kern.h
> @@ -29,6 +29,7 @@
> #define DROP_PKT 0
> #define ALLOW_PKT 1
> #define TCP_ECN_OK 1
> +#define CWR 2
>
> #ifndef HBM_DEBUG // Define HBM_DEBUG to enable debugging
> #undef bpf_printk
> @@ -45,8 +46,18 @@
> #define MAX_CREDIT (100 * MAX_BYTES_PER_PACKET)
> #define INIT_CREDIT (INITIAL_CREDIT_PACKETS *
MAX_BYTES_PER_PACKET)
>
> +// Time base accounting for fq's EDT
> +#define BURST_SIZE_NS 100000 // 100us
> +#define MARK_THRESH_NS 50000 // 50us
> +#define DROP_THRESH_NS 500000 // 500us
> +// Reserve 20us of queuing for small packets (less than 120 bytes)
> +#define LARGE_PKT_DROP_THRESH_NS (DROP_THRESH_NS - 20000)
> +#define MARK_REGION_SIZE_NS (LARGE_PKT_DROP_THRESH_NS -
MARK_THRESH_NS)
> +
> // rate in bytes per ns << 20
> #define CREDIT_PER_NS(delta, rate) ((((u64)(delta)) * (rate)) >> 20)
> +#define BYTES_PER_NS(delta, rate) ((((u64)(delta)) * (rate)) >> 20)
> +#define BYTES_TO_NS(bytes, rate) div64_u64(((u64)(bytes)) << 20,
(u64)(rate))
>
> struct bpf_map_def SEC("maps") queue_state = {
> .type = BPF_MAP_TYPE_CGROUP_STORAGE,
> @@ -67,6 +78,7 @@ BPF_ANNOTATE_KV_PAIR(queue_stats, int, struct
hbm_queue_stats);
> struct hbm_pkt_info {
> int cwnd;
> int rtt;
> + int packets_out;
> bool is_ip;
> bool is_tcp;
> short ecn;
> @@ -86,16 +98,20 @@ static int get_tcp_info(struct __sk_buff *skb, struct
hbm_pkt_info *pkti)
> if (tp) {
> pkti->cwnd = tp->snd_cwnd;
> pkti->rtt = tp->srtt_us >> 3;
> + pkti->packets_out = tp->packets_out;
> return 0;
> }
> }
> }
> }
> + pkti->cwnd = 0;
> + pkti->rtt = 0;
> + pkti->packets_out = 0;
> return 1;
> }
>
> -static __always_inline void hbm_get_pkt_info(struct __sk_buff *skb,
> - struct hbm_pkt_info *pkti)
> +static void hbm_get_pkt_info(struct __sk_buff *skb,
> + struct hbm_pkt_info *pkti)
> {
> struct iphdr iph;
> struct ipv6hdr *ip6h;
> @@ -123,10 +139,22 @@ static __always_inline void hbm_get_pkt_info(struct
__sk_buff *skb,
>
> static __always_inline void hbm_init_vqueue(struct hbm_vqueue *qdp, int
rate)
> {
> - bpf_printk("Initializing queue_state, rate:%d\n", rate * 128);
> - qdp->lasttime = bpf_ktime_get_ns();
> - qdp->credit = INIT_CREDIT;
> - qdp->rate = rate * 128;
> + bpf_printk("Initializing queue_state, rate:%d\n", rate * 128);
> + qdp->lasttime = bpf_ktime_get_ns();
> + qdp->credit = INIT_CREDIT;
> + qdp->rate = rate * 128;
> +}
> +
> +static __always_inline void hbm_init_edt_vqueue(struct hbm_vqueue *qdp,
> + int rate)
> +{
> + unsigned long long curtime;
> +
> + curtime = bpf_ktime_get_ns();
> + bpf_printk("Initializing queue_state, rate:%d\n", rate * 128);
> + qdp->lasttime = curtime - BURST_SIZE_NS; // support initial burst
> + qdp->credit = 0; // not used
> + qdp->rate = rate * 128;
> }
>
> static __always_inline void hbm_update_stats(struct hbm_queue_stats
*qsp,
>
