Now that we have the __bpf_kfunc tag, we should use add it to all
existing kfuncs to ensure that they'll never be elided in LTO builds.
Signed-off-by: David Vernet <void@manifault.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Stanislav Fomichev <sdf@google.com>
Link: https://lore.kernel.org/bpf/20230201173016.342758-4-void@manifault.com
PLB support is added to TCP DCTCP code. As DCTCP uses ECN as the
congestion signal, PLB also uses ECN to make decisions whether to change
the path or not upon sustained congestion.
Signed-off-by: Mubashir Adnan Qureshi <mubashirq@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of populating multiple sets to indicate some attribute and then
researching the same BTF ID in them, prepare a single unified BTF set
which indicates whether a kfunc is allowed to be called, and also its
attributes if any at the same time. Now, only one call is needed to
perform the lookup for both kfunc availability and its attributes.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20220721134245.2450-4-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
We had various bugs over the years with code
breaking the assumption that tp->snd_cwnd is greater
than zero.
Lately, syzbot reported the WARN_ON_ONCE(!tp->prior_cwnd) added
in commit 8b8a321ff7 ("tcp: fix zero cwnd in tcp_cwnd_reduction")
can trigger, and without a repro we would have to spend
considerable time finding the bug.
Instead of complaining too late, we want to catch where
and when tp->snd_cwnd is set to an illegal value.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Suggested-by: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Link: https://lore.kernel.org/r/20220405233538.947344-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Completely remove the old code for check_kfunc_call to help it work
with modules, and also the callback itself.
The previous commit adds infrastructure to register all sets and put
them in vmlinux or module BTF, and concatenates all related sets
organized by the hook and the type. Once populated, these sets remain
immutable for the lifetime of the struct btf.
Also, since we don't need the 'owner' module anywhere when doing
check_kfunc_call, drop the 'btf_modp' module parameter from
find_kfunc_desc_btf.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20220114163953.1455836-4-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
This commit moves BTF ID lookup into the newly added registration
helper, in a way that the bbr, cubic, and dctcp implementation set up
their sets in the bpf_tcp_ca kfunc_btf_set list, while the ones not
dependent on modules are looked up from the wrapper function.
This lifts the restriction for them to be compiled as built in objects,
and can be loaded as modules if required. Also modify Makefile.modfinal
to call resolve_btfids for each module.
Note that since kernel kfunc_ids never overlap with module kfunc_ids, we
only match the owner for module btf id sets.
See following commits for background on use of:
CONFIG_X86 ifdef:
569c484f99 (bpf: Limit static tcp-cc functions in the .BTF_ids list to x86)
CONFIG_DYNAMIC_FTRACE ifdef:
7aae231ac9 (bpf: tcp: Limit calling some tcp cc functions to CONFIG_DYNAMIC_FTRACE)
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20211002011757.311265-6-memxor@gmail.com
Based on 1 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license as published by
the free software foundation either version 2 of the license or at
your option any later version
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-or-later
has been chosen to replace the boilerplate/reference in 3029 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190527070032.746973796@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
After commit e21db6f69a ("tcp: track total bytes delivered with ECN CE marks")
core TCP stack does a very good job tracking ECN signals.
The "sender's best estimate of CE information" Yuchung mentioned in his
patch is indeed the best we can do.
DCTCP can use tp->delivered_ce and tp->delivered to not duplicate the logic,
and use the existing best estimate.
This solves some problems, since current DCTCP logic does not deal with losses
and/or GRO or ack aggregation very well.
This also removes a dubious use of inet_csk(sk)->icsk_ack.rcv_mss
(this should have been tp->mss_cache), and a 64 bit divide.
Finally, we can see that the DCTCP logic, calling dctcp_update_alpha() for
every ACK could be done differently, calling it only once per RTT.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Soheil Hassas Yeganeh <soheil@google.com>
Cc: Florian Westphal <fw@strlen.de>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Lawrence Brakmo <brakmo@fb.com>
Cc: Abdul Kabbani <akabbani@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
RFC8257 §3.5 explicitly states that "A DCTCP sender MUST react to
loss episodes in the same way as conventional TCP".
Currently, Linux DCTCP performs no cwnd reduction when losses
are encountered. Optionally, the dctcp_clamp_alpha_on_loss resets
alpha to its maximal value if a RTO happens. This behavior
is sub-optimal for at least two reasons: i) it ignores losses
triggering fast retransmissions; and ii) it causes unnecessary large
cwnd reduction in the future if the loss was isolated as it resets
the historical term of DCTCP's alpha EWMA to its maximal value (i.e.,
denoting a total congestion). The second reason has an especially
noticeable effect when using DCTCP in high BDP environments, where
alpha normally stays at low values.
This patch replace the clamping of alpha by setting ssthresh to
half of cwnd for both fast retransmissions and RTOs, at most once
per RTT. Consequently, the dctcp_clamp_alpha_on_loss module parameter
has been removed.
The table below shows experimental results where we measured the
drop probability of a PIE AQM (not applying ECN marks) at a
bottleneck in the presence of a single TCP flow with either the
alpha-clamping option enabled or the cwnd halving proposed by this
patch. Results using reno or cubic are given for comparison.
| Link | RTT | Drop
TCP CC | speed | base+AQM | probability
==================|=========|==========|============
CUBIC | 40Mbps | 7+20ms | 0.21%
RENO | | | 0.19%
DCTCP-CLAMP-ALPHA | | | 25.80%
DCTCP-HALVE-CWND | | | 0.22%
------------------|---------|----------|------------
CUBIC | 100Mbps | 7+20ms | 0.03%
RENO | | | 0.02%
DCTCP-CLAMP-ALPHA | | | 23.30%
DCTCP-HALVE-CWND | | | 0.04%
------------------|---------|----------|------------
CUBIC | 800Mbps | 1+1ms | 0.04%
RENO | | | 0.05%
DCTCP-CLAMP-ALPHA | | | 18.70%
DCTCP-HALVE-CWND | | | 0.06%
We see that, without halving its cwnd for all source of losses,
DCTCP drives the AQM to large drop probabilities in order to keep
the queue length under control (i.e., it repeatedly faces RTOs).
Instead, if DCTCP reacts to all source of losses, it can then be
controlled by the AQM using similar drop levels than cubic or reno.
Signed-off-by: Koen De Schepper <koen.de_schepper@nokia-bell-labs.com>
Signed-off-by: Olivier Tilmans <olivier.tilmans@nokia-bell-labs.com>
Cc: Bob Briscoe <research@bobbriscoe.net>
Cc: Lawrence Brakmo <brakmo@fb.com>
Cc: Florian Westphal <fw@strlen.de>
Cc: Daniel Borkmann <borkmann@iogearbox.net>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Andrew Shewmaker <agshew@gmail.com>
Cc: Glenn Judd <glenn.judd@morganstanley.com>
Acked-by: Florian Westphal <fw@strlen.de>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
DCTCP has two parts - a new ECN signalling mechanism and the response
function to it. The first part can be used by other congestion
control for DCTCP-ECN deployed networks. This patch moves that part
into a separate tcp_dctcp.h to be used by other congestion control
module (like how Yeah uses Vegas algorithmas). For example, BBR is
experimenting such ECN signal currently
https://tinyurl.com/ietf-102-iccrg-bbr2
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Yousuk Seung <ysseung@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The recent fix of acking immediately in DCTCP on CE status change
has an undesirable side-effect: it also resets TCP ack timer and
disables pingpong mode (interactive session). But the CE status
change has nothing to do with them. This patch addresses that by
using the new one-time immediate ACK flag instead of calling
tcp_enter_quickack_mode().
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Wei Wang <weiwan@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Per DCTCP RFC8257 (Section 3.2) the ACK reflecting the CE status change
has to be sent immediately so the sender can respond quickly:
""" When receiving packets, the CE codepoint MUST be processed as follows:
1. If the CE codepoint is set and DCTCP.CE is false, set DCTCP.CE to
true and send an immediate ACK.
2. If the CE codepoint is not set and DCTCP.CE is true, set DCTCP.CE
to false and send an immediate ACK.
"""
Previously DCTCP implementation may continue to delay the ACK. This
patch fixes that to implement the RFC by forcing an immediate ACK.
Tested with this packetdrill script provided by Larry Brakmo
0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
0.000 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
0.000 setsockopt(3, SOL_TCP, TCP_CONGESTION, "dctcp", 5) = 0
0.000 bind(3, ..., ...) = 0
0.000 listen(3, 1) = 0
0.100 < [ect0] SEW 0:0(0) win 32792 <mss 1000,sackOK,nop,nop,nop,wscale 7>
0.100 > SE. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8>
0.110 < [ect0] . 1:1(0) ack 1 win 257
0.200 accept(3, ..., ...) = 4
+0 setsockopt(4, SOL_SOCKET, SO_DEBUG, [1], 4) = 0
0.200 < [ect0] . 1:1001(1000) ack 1 win 257
0.200 > [ect01] . 1:1(0) ack 1001
0.200 write(4, ..., 1) = 1
0.200 > [ect01] P. 1:2(1) ack 1001
0.200 < [ect0] . 1001:2001(1000) ack 2 win 257
+0.005 < [ce] . 2001:3001(1000) ack 2 win 257
+0.000 > [ect01] . 2:2(0) ack 2001
// Previously the ACK below would be delayed by 40ms
+0.000 > [ect01] E. 2:2(0) ack 3001
+0.500 < F. 9501:9501(0) ack 4 win 257
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently when a DCTCP receiver delays an ACK and receive a
data packet with a different CE mark from the previous one's, it
sends two immediate ACKs acking previous and latest sequences
respectly (for ECN accounting).
Previously sending the first ACK may mark off the delayed ACK timer
(tcp_event_ack_sent). This may subsequently prevent sending the
second ACK to acknowledge the latest sequence (tcp_ack_snd_check).
The culprit is that tcp_send_ack() assumes it always acknowleges
the latest sequence, which is not true for the first special ACK.
The fix is to not make the assumption in tcp_send_ack and check the
actual ack sequence before cancelling the delayed ACK. Further it's
safer to pass the ack sequence number as a local variable into
tcp_send_ack routine, instead of intercepting tp->rcv_nxt to avoid
future bugs like this.
Reported-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After fixing the way DCTCP tracking delayed ACKs, the delayed-ACK
related callbacks are no longer needed
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously, when a data segment was sent an ACK was piggybacked
on the data segment without generating a CA_EVENT_NON_DELAYED_ACK
event to notify congestion control modules. So the DCTCP
ca->delayed_ack_reserved flag could incorrectly stay set when
in fact there were no delayed ACKs being reserved. This could result
in sending a special ECN notification ACK that carries an older
ACK sequence, when in fact there was no need for such an ACK.
DCTCP keeps track of the delayed ACK status with its own separate
state ca->delayed_ack_reserved. Previously it may accidentally cancel
the delayed ACK without updating this field upon sending a special
ACK that carries a older ACK sequence. This inconsistency would
lead to DCTCP receiver never acknowledging the latest data until the
sender times out and retry in some cases.
Packetdrill script (provided by Larry Brakmo)
0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
0.000 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
0.000 setsockopt(3, SOL_TCP, TCP_CONGESTION, "dctcp", 5) = 0
0.000 bind(3, ..., ...) = 0
0.000 listen(3, 1) = 0
0.100 < [ect0] SEW 0:0(0) win 32792 <mss 1000,sackOK,nop,nop,nop,wscale 7>
0.100 > SE. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 8>
0.110 < [ect0] . 1:1(0) ack 1 win 257
0.200 accept(3, ..., ...) = 4
0.200 < [ect0] . 1:1001(1000) ack 1 win 257
0.200 > [ect01] . 1:1(0) ack 1001
0.200 write(4, ..., 1) = 1
0.200 > [ect01] P. 1:2(1) ack 1001
0.200 < [ect0] . 1001:2001(1000) ack 2 win 257
0.200 write(4, ..., 1) = 1
0.200 > [ect01] P. 2:3(1) ack 2001
0.200 < [ect0] . 2001:3001(1000) ack 3 win 257
0.200 < [ect0] . 3001:4001(1000) ack 3 win 257
0.200 > [ect01] . 3:3(0) ack 4001
0.210 < [ce] P. 4001:4501(500) ack 3 win 257
+0.001 read(4, ..., 4500) = 4500
+0 write(4, ..., 1) = 1
+0 > [ect01] PE. 3:4(1) ack 4501
+0.010 < [ect0] W. 4501:5501(1000) ack 4 win 257
// Previously the ACK sequence below would be 4501, causing a long RTO
+0.040~+0.045 > [ect01] . 4:4(0) ack 5501 // delayed ack
+0.311 < [ect0] . 5501:6501(1000) ack 4 win 257 // More data
+0 > [ect01] . 4:4(0) ack 6501 // now acks everything
+0.500 < F. 9501:9501(0) ack 4 win 257
Reported-by: Larry Brakmo <brakmo@fb.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Neal Cardwell says:
If I am reading the code correctly, then I would have two concerns:
1) Has that been tested? That seems like an extremely dramatic
decrease in cwnd. For example, if the cwnd is 80, and there are 40
ACKs, and half the ACKs are ECE marked, then my back-of-the-envelope
calculations seem to suggest that after just 11 ACKs the cwnd would be
down to a minimal value of 2 [..]
2) That seems to contradict another passage in the draft [..] where it
sazs:
Just as specified in [RFC3168], DCTCP does not react to congestion
indications more than once for every window of data.
Neal is right. Fortunately we don't have to complicate this by testing
vs. current rtt estimate, we can just revert the patch.
Normal stack already handles this for us: receiving ACKs with ECE
set causes a call to tcp_enter_cwr(), from there on the ssthresh gets
adjusted and prr will take care of cwnd adjustment.
Fixes: 4780566784 ("dctcp: update cwnd on congestion event")
Cc: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The undo_cwnd fallback in the stack doubles cwnd based on ssthresh,
which un-does reno halving behaviour.
It seems more appropriate to let congctl algorithms pair .ssthresh
and .undo_cwnd properly. Add a 'tcp_reno_undo_cwnd' function and wire it
up for all congestion algorithms that used to rely on the fallback.
Cc: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
draft-ietf-tcpm-dctcp-02 says:
... when the sender receives an indication of congestion
(ECE), the sender SHOULD update cwnd as follows:
cwnd = cwnd * (1 - DCTCP.Alpha / 2)
So, lets do this and reduce cwnd more smoothly (and faster), as per
current congestion estimate.
Cc: Lawrence Brakmo <brakmo@fb.com>
Cc: Andrew Shewmaker <agshew@gmail.com>
Cc: Glenn Judd <glenn.judd@morganstanley.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
If a congestion control module doesn't provide .undo_cwnd function,
tcp_undo_cwnd_reduction() will set cwnd to
tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
... which makes sense for reno (it sets ssthresh to half the current cwnd),
but it makes no sense for dctcp, which sets ssthresh based on the current
congestion estimate.
This can cause severe growth of cwnd (eventually overflowing u32).
Fix this by saving last cwnd on loss and restore cwnd based on that,
similar to cubic and other algorithms.
Fixes: e3118e8359 ("net: tcp: add DCTCP congestion control algorithm")
Cc: Lawrence Brakmo <brakmo@fb.com>
Cc: Andrew Shewmaker <agshew@gmail.com>
Cc: Glenn Judd <glenn.judd@morganstanley.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Make sure that dctcp_get_info() returns only the size of the
info->dctcp struct that it zeroes out and fills in. Previously it had
been returning the size of the enclosing tcp_cc_info union,
sizeof(*info). There is no problem yet, but that union that may one
day be larger than struct tcp_dctcp_info, in which case the
TCP_CC_INFO code might accidentally copy uninitialized bytes from the
stack.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
If alpha is strictly reduced by alpha >> dctcp_shift_g and if alpha is less
than 1 << dctcp_shift_g, then alpha may never reach zero. For example,
given shift_g=4 and alpha=15, alpha >> dctcp_shift_g yields 0 and alpha
remains 15. The effect isn't noticeable in this case below cwnd=137, but
could gradually drive uncongested flows with leftover alpha down to
cwnd=137. A larger dctcp_shift_g would have a greater effect.
This change causes alpha=15 to drop to 0 instead of being decrementing by 1
as it would when alpha=16. However, it requires one less conditional to
implement since it doesn't have to guard against subtracting 1 from 0U. A
decay of 15 is not unreasonable since an equal or greater amount occurs at
alpha >= 240.
Signed-off-by: Andrew G. Shewmaker <agshew@gmail.com>
Acked-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
dctcp_alpha can be read by from dctcp_get_info() without
synchro, so use WRITE_ONCE() to prevent compiler from using
dctcp_alpha as a temporary variable.
Also, playing with small dctcp_shift_g (like 1), can expose
an overflow with 32bit values shifted 9 times before divide.
Use an u64 field to avoid this problem, and perform the divide
only if acked_bytes_ecn is not zero.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We would like that optional info provided by Congestion Control
modules using netlink can also be read using getsockopt()
This patch changes get_info() to put this information in a buffer,
instead of skb, like tcp_get_info(), so that following patch
can reuse this common infrastructure.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Two different problems are fixed here :
1) inet_sk_diag_fill() might be called without socket lock held.
icsk->icsk_ca_ops can change under us and module be unloaded.
-> Access to freed memory.
Fix this using rcu_read_lock() to prevent module unload.
2) Some TCP Congestion Control modules provide information
but again this is not safe against icsk->icsk_ca_ops
change and nla_put() errors were ignored. Some sockets
could not get the additional info if skb was almost full.
Fix this by returning a status from get_info() handlers and
using rcu protection as well.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
This work adds the DataCenter TCP (DCTCP) congestion control
algorithm [1], which has been first published at SIGCOMM 2010 [2],
resp. follow-up analysis at SIGMETRICS 2011 [3] (and also, more
recently as an informational IETF draft available at [4]).
DCTCP is an enhancement to the TCP congestion control algorithm for
data center networks. Typical data center workloads are i.e.
i) partition/aggregate (queries; bursty, delay sensitive), ii) short
messages e.g. 50KB-1MB (for coordination and control state; delay
sensitive), and iii) large flows e.g. 1MB-100MB (data update;
throughput sensitive). DCTCP has therefore been designed for such
environments to provide/achieve the following three requirements:
* High burst tolerance (incast due to partition/aggregate)
* Low latency (short flows, queries)
* High throughput (continuous data updates, large file
transfers) with commodity, shallow buffered switches
The basic idea of its design consists of two fundamentals: i) on the
switch side, packets are being marked when its internal queue
length > threshold K (K is chosen so that a large enough headroom
for marked traffic is still available in the switch queue); ii) the
sender/host side maintains a moving average of the fraction of marked
packets, so each RTT, F is being updated as follows:
F := X / Y, where X is # of marked ACKs, Y is total # of ACKs
alpha := (1 - g) * alpha + g * F, where g is a smoothing constant
The resulting alpha (iow: probability that switch queue is congested)
is then being used in order to adaptively decrease the congestion
window W:
W := (1 - (alpha / 2)) * W
The means for receiving marked packets resp. marking them on switch
side in DCTCP is the use of ECN.
RFC3168 describes a mechanism for using Explicit Congestion Notification
from the switch for early detection of congestion, rather than waiting
for segment loss to occur.
However, this method only detects the presence of congestion, not
the *extent*. In the presence of mild congestion, it reduces the TCP
congestion window too aggressively and unnecessarily affects the
throughput of long flows [4].
DCTCP, as mentioned, enhances Explicit Congestion Notification (ECN)
processing to estimate the fraction of bytes that encounter congestion,
rather than simply detecting that some congestion has occurred. DCTCP
then scales the TCP congestion window based on this estimate [4],
thus it can derive multibit feedback from the information present in
the single-bit sequence of marks in its control law. And thus act in
*proportion* to the extent of congestion, not its *presence*.
Switches therefore set the Congestion Experienced (CE) codepoint in
packets when internal queue lengths exceed threshold K. Resulting,
DCTCP delivers the same or better throughput than normal TCP, while
using 90% less buffer space.
It was found in [2] that DCTCP enables the applications to handle 10x
the current background traffic, without impacting foreground traffic.
Moreover, a 10x increase in foreground traffic did not cause any
timeouts, and thus largely eliminates TCP incast collapse problems.
The algorithm itself has already seen deployments in large production
data centers since then.
We did a long-term stress-test and analysis in a data center, short
summary of our TCP incast tests with iperf compared to cubic:
This test measured DCTCP throughput and latency and compared it with
CUBIC throughput and latency for an incast scenario. In this test, 19
senders sent at maximum rate to a single receiver. The receiver simply
ran iperf -s.
The senders ran iperf -c <receiver> -t 30. All senders started
simultaneously (using local clocks synchronized by ntp).
This test was repeated multiple times. Below shows the results from a
single test. Other tests are similar. (DCTCP results were extremely
consistent, CUBIC results show some variance induced by the TCP timeouts
that CUBIC encountered.)
For this test, we report statistics on the number of TCP timeouts,
flow throughput, and traffic latency.
1) Timeouts (total over all flows, and per flow summaries):
CUBIC DCTCP
Total 3227 25
Mean 169.842 1.316
Median 183 1
Max 207 5
Min 123 0
Stddev 28.991 1.600
Timeout data is taken by measuring the net change in netstat -s
"other TCP timeouts" reported. As a result, the timeout measurements
above are not restricted to the test traffic, and we believe that it
is likely that all of the "DCTCP timeouts" are actually timeouts for
non-test traffic. We report them nevertheless. CUBIC will also include
some non-test timeouts, but they are drawfed by bona fide test traffic
timeouts for CUBIC. Clearly DCTCP does an excellent job of preventing
TCP timeouts. DCTCP reduces timeouts by at least two orders of
magnitude and may well have eliminated them in this scenario.
2) Throughput (per flow in Mbps):
CUBIC DCTCP
Mean 521.684 521.895
Median 464 523
Max 776 527
Min 403 519
Stddev 105.891 2.601
Fairness 0.962 0.999
Throughput data was simply the average throughput for each flow
reported by iperf. By avoiding TCP timeouts, DCTCP is able to
achieve much better per-flow results. In CUBIC, many flows
experience TCP timeouts which makes flow throughput unpredictable and
unfair. DCTCP, on the other hand, provides very clean predictable
throughput without incurring TCP timeouts. Thus, the standard deviation
of CUBIC throughput is dramatically higher than the standard deviation
of DCTCP throughput.
Mean throughput is nearly identical because even though cubic flows
suffer TCP timeouts, other flows will step in and fill the unused
bandwidth. Note that this test is something of a best case scenario
for incast under CUBIC: it allows other flows to fill in for flows
experiencing a timeout. Under situations where the receiver is issuing
requests and then waiting for all flows to complete, flows cannot fill
in for timed out flows and throughput will drop dramatically.
3) Latency (in ms):
CUBIC DCTCP
Mean 4.0088 0.04219
Median 4.055 0.0395
Max 4.2 0.085
Min 3.32 0.028
Stddev 0.1666 0.01064
Latency for each protocol was computed by running "ping -i 0.2
<receiver>" from a single sender to the receiver during the incast
test. For DCTCP, "ping -Q 0x6 -i 0.2 <receiver>" was used to ensure
that traffic traversed the DCTCP queue and was not dropped when the
queue size was greater than the marking threshold. The summary
statistics above are over all ping metrics measured between the single
sender, receiver pair.
The latency results for this test show a dramatic difference between
CUBIC and DCTCP. CUBIC intentionally overflows the switch buffer
which incurs the maximum queue latency (more buffer memory will lead
to high latency.) DCTCP, on the other hand, deliberately attempts to
keep queue occupancy low. The result is a two orders of magnitude
reduction of latency with DCTCP - even with a switch with relatively
little RAM. Switches with larger amounts of RAM will incur increasing
amounts of latency for CUBIC, but not for DCTCP.
4) Convergence and stability test:
This test measured the time that DCTCP took to fairly redistribute
bandwidth when a new flow commences. It also measured DCTCP's ability
to remain stable at a fair bandwidth distribution. DCTCP is compared
with CUBIC for this test.
At the commencement of this test, a single flow is sending at maximum
rate (near 10 Gbps) to a single receiver. One second after that first
flow commences, a new flow from a distinct server begins sending to
the same receiver as the first flow. After the second flow has sent
data for 10 seconds, the second flow is terminated. The first flow
sends for an additional second. Ideally, the bandwidth would be evenly
shared as soon as the second flow starts, and recover as soon as it
stops.
The results of this test are shown below. Note that the flow bandwidth
for the two flows was measured near the same time, but not
simultaneously.
DCTCP performs nearly perfectly within the measurement limitations
of this test: bandwidth is quickly distributed fairly between the two
flows, remains stable throughout the duration of the test, and
recovers quickly. CUBIC, in contrast, is slow to divide the bandwidth
fairly, and has trouble remaining stable.
CUBIC DCTCP
Seconds Flow 1 Flow 2 Seconds Flow 1 Flow 2
0 9.93 0 0 9.92 0
0.5 9.87 0 0.5 9.86 0
1 8.73 2.25 1 6.46 4.88
1.5 7.29 2.8 1.5 4.9 4.99
2 6.96 3.1 2 4.92 4.94
2.5 6.67 3.34 2.5 4.93 5
3 6.39 3.57 3 4.92 4.99
3.5 6.24 3.75 3.5 4.94 4.74
4 6 3.94 4 5.34 4.71
4.5 5.88 4.09 4.5 4.99 4.97
5 5.27 4.98 5 4.83 5.01
5.5 4.93 5.04 5.5 4.89 4.99
6 4.9 4.99 6 4.92 5.04
6.5 4.93 5.1 6.5 4.91 4.97
7 4.28 5.8 7 4.97 4.97
7.5 4.62 4.91 7.5 4.99 4.82
8 5.05 4.45 8 5.16 4.76
8.5 5.93 4.09 8.5 4.94 4.98
9 5.73 4.2 9 4.92 5.02
9.5 5.62 4.32 9.5 4.87 5.03
10 6.12 3.2 10 4.91 5.01
10.5 6.91 3.11 10.5 4.87 5.04
11 8.48 0 11 8.49 4.94
11.5 9.87 0 11.5 9.9 0
SYN/ACK ECT test:
This test demonstrates the importance of ECT on SYN and SYN-ACK packets
by measuring the connection probability in the presence of competing
flows for a DCTCP connection attempt *without* ECT in the SYN packet.
The test was repeated five times for each number of competing flows.
Competing Flows 1 | 2 | 4 | 8 | 16
------------------------------
Mean Connection Probability 1 | 0.67 | 0.45 | 0.28 | 0
Median Connection Probability 1 | 0.65 | 0.45 | 0.25 | 0
As the number of competing flows moves beyond 1, the connection
probability drops rapidly.
Enabling DCTCP with this patch requires the following steps:
DCTCP must be running both on the sender and receiver side in your
data center, i.e.:
sysctl -w net.ipv4.tcp_congestion_control=dctcp
Also, ECN functionality must be enabled on all switches in your
data center for DCTCP to work. The default ECN marking threshold (K)
heuristic on the switch for DCTCP is e.g., 20 packets (30KB) at
1Gbps, and 65 packets (~100KB) at 10Gbps (K > 1/7 * C * RTT, [4]).
In above tests, for each switch port, traffic was segregated into two
queues. For any packet with a DSCP of 0x01 - or equivalently a TOS of
0x04 - the packet was placed into the DCTCP queue. All other packets
were placed into the default drop-tail queue. For the DCTCP queue,
RED/ECN marking was enabled, here, with a marking threshold of 75 KB.
More details however, we refer you to the paper [2] under section 3).
There are no code changes required to applications running in user
space. DCTCP has been implemented in full *isolation* of the rest of
the TCP code as its own congestion control module, so that it can run
without a need to expose code to the core of the TCP stack, and thus
nothing changes for non-DCTCP users.
Changes in the CA framework code are minimal, and DCTCP algorithm
operates on mechanisms that are already available in most Silicon.
The gain (dctcp_shift_g) is currently a fixed constant (1/16) from
the paper, but we leave the option that it can be chosen carefully
to a different value by the user.
In case DCTCP is being used and ECN support on peer site is off,
DCTCP falls back after 3WHS to operate in normal TCP Reno mode.
ss {-4,-6} -t -i diag interface:
... dctcp wscale:7,7 rto:203 rtt:2.349/0.026 mss:1448 cwnd:2054
ssthresh:1102 ce_state 0 alpha 15 ab_ecn 0 ab_tot 735584
send 10129.2Mbps pacing_rate 20254.1Mbps unacked:1822 retrans:0/15
reordering:101 rcv_space:29200
... dctcp-reno wscale:7,7 rto:201 rtt:0.711/1.327 ato:40 mss:1448
cwnd:10 ssthresh:1102 fallback_mode send 162.9Mbps pacing_rate
325.5Mbps rcv_rtt:1.5 rcv_space:29200
More information about DCTCP can be found in [1-4].
[1] http://simula.stanford.edu/~alizade/Site/DCTCP.html
[2] http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf
[3] http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf
[4] http://tools.ietf.org/html/draft-bensley-tcpm-dctcp-00
Joint work with Florian Westphal and Glenn Judd.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Glenn Judd <glenn.judd@morganstanley.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
David Vernet
Mubashir Adnan Qureshi
Kumar Kartikeya Dwivedi
Eric Dumazet
Thomas Gleixner
Koen De Schepper
Yuchung Cheng
Florian Westphal
Neal Cardwell
Andrew Shewmaker
Daniel Borkmann