| TCP Maintenance Working Group | Y. Cheng |
| Internet-Draft | N. Cardwell |
| Intended status: Experimental | N. Dukkipati |
| Expires: September 6, 2018 | P. Jha |
| Google, Inc | |
| March 5, 2018 |
RACK: a time-based fast loss detection algorithm for TCP
draft-ietf-tcpm-rack-03
This document presents a new TCP loss detection algorithm called RACK ("Recent ACKnowledgment"). RACK uses the notion of time, instead of packet or sequence counts, to detect losses, for modern TCP implementations that can support per-packet timestamps and the selective acknowledgment (SACK) option. It is intended to replace the conventional DUPACK threshold approach and its variants, as well as other nonstandard approaches.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on September 6, 2018.
Copyright (c) 2018 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
This document presents a new loss detection algorithm called RACK ("Recent ACKnowledgment"). RACK uses the notion of time instead of the conventional packet or sequence counting approaches for detecting losses. RACK deems a packet lost if some packet sent sufficiently later has been delivered. It does this by recording packet transmission times and inferring losses using cumulative acknowledgments or selective acknowledgment (SACK) TCP options.
In the last couple of years we have been observing several increasingly common loss and reordering patterns in the Internet:
Despite TCP stacks (e.g. Linux) that implement many of the standard and proposed loss detection algorithms [RFC3517][RFC4653][RFC5827][RFC5681][RFC6675][RFC7765][FACK][THIN-STREAM][TLP], we've found that together they do not perform well. The main reason is that many of them are based on the classic rule of counting duplicate acknowledgments [RFC5681]. They can either detect loss quickly or accurately, but not both, especially when the sender is application-limited or under reordering that is unpredictable. And under these conditions none of them can detect lost retransmissions well.
Also, these algorithms, including RFCs, rarely address the interactions with other algorithms. For example, FACK may consider a packet is lost while RFC3517 may not. Implementing N algorithms while dealing with N^2 interactions is a daunting task and error-prone.
The goal of RACK is to solve all the problems above by replacing many of the loss detection algorithms above with one simpler, and also more effective, algorithm.
The main idea behind RACK is that if a packet has been delivered out of order, then the packets sent chronologically before that were either lost or reordered. This concept is not fundamentally different from [RFC5681][RFC3517][FACK]. But the key innovation in RACK is to use a per-packet transmission timestamp and widely deployed SACK options to conduct time-based inferences instead of inferring losses with packet or sequence counting approaches.
Using a threshold for counting duplicate acknowledgments (i.e., DupThresh) is no longer reliable because of today's prevalent reordering patterns. A common type of reordering is that the last "runt" packet of a window's worth of packet bursts gets delivered first, then the rest arrive shortly after in order. To handle this effectively, a sender would need to constantly adjust the DupThresh to the burst size; but this would risk increasing the frequency of RTOs on real losses.
Today's prevalent lost retransmissions also cause problems with packet-counting approaches [RFC5681][RFC3517][FACK], since those approaches depend on reasoning in sequence number space. Retransmissions break the direct correspondence between ordering in sequence space and ordering in time. So when retransmissions are lost, sequence-based approaches are often unable to infer and quickly repair losses that can be deduced with time-based approaches.
Instead of counting packets, RACK uses the most recently delivered packet’s transmission time to judge if some packets sent previous to that time have "expired" by passing a certain reordering settling window. On each ACK, RACK marks any already-expired packets lost, and for any packets that have not yet expired it waits until the reordering window passes and then marks those lost as well. In either case, RACK can repair the loss without waiting for a (long) RTO. RACK can be applied to both fast recovery and timeout recovery, and can detect losses on both originally transmitted and retransmitted packets, making it a great all-weather loss detection mechanism.
The reader is expected to be familiar with the definitions given in the TCP congestion control [RFC5681] and selective acknowledgment [RFC2018] RFCs. Familiarity with the conservative SACK-based recovery for TCP [RFC6675] is not expected but helps.
RACK has three requirements:
We assume that requirement 1 implies the sender keeps a SACK scoreboard, which is a data structure to store selective acknowledgment information on a per-connection basis ([RFC6675] section 3). For the ease of explaining the algorithm, we use a pseudo-scoreboard that manages the data in sequence number ranges. But the specifics of the data structure are left to the implementor.
RACK does not need any change on the receiver.
A sender needs to store these new RACK variables:
“Packet.xmit_ts” is the time of the last transmission of a data packet, including retransmissions, if any. The sender needs to record the transmission time for each packet sent and not yet acknowledged. The time MUST be stored at millisecond granularity or finer.
“RACK.packet”. Among all the packets that have been either selectively or cumulatively acknowledged, RACK.packet is the one that was sent most recently (including retransmissions).
“RACK.xmit_ts” is the latest transmission timestamp of RACK.packet.
“RACK.end_seq” is the ending TCP sequence number of RACK.packet.
“RACK.RTT” is the associated RTT measured when RACK.xmit_ts, above, was changed. It is the RTT of the most recently transmitted packet that has been delivered (either cumulatively acknowledged or selectively acknowledged) on the connection.
“RACK.reo_wnd” is a reordering window for the connection, computed in the unit of time used for recording packet transmission times. It is used to defer the moment at which RACK marks a packet lost.
“RACK.min_RTT” is the estimated minimum round-trip time (RTT) of the connection.
“RACK.ack_ts” is the time when all the sequences in RACK.packet were selectively or cumulatively acknowledged.
“RACK.reo_wnd_incr” is the multiplier applied to adjust RACK.reo_wnd
“RACK.reo_wnd_persist” is the number of loss recoveries before resetting RACK.reo_wnd “RACK.dsack” indicates if RACK.reo_wnd has been adjusted upon receiving a DSACK option
Note that the Packet.xmit_ts variable is per packet in flight. The RACK.xmit_ts, RACK.end_seq, RACK.RTT, RACK.reo_wnd, and RACK.min_RTT variables are kept in the per-connection TCP control block. RACK.packet and RACK.ack_ts are used as local variables in the algorithm.
Upon transmitting a new packet or retransmitting an old packet, record the time in Packet.xmit_ts. RACK does not care if the retransmission is triggered by an ACK, new application data, an RTO, or any other means.
Step 1: Update RACK.min_RTT.
Use the RTT measurements obtained via [RFC6298] or [RFC7323] to update the estimated minimum RTT in RACK.min_RTT. The sender can track a simple global minimum of all RTT measurements from the connection, or a windowed min-filtered value of recent RTT measurements. This document does not specify an exact approach.
Step 2: Update RACK stats
Given the information provided in an ACK, each packet cumulatively ACKed or SACKed is marked as delivered in the scoreboard. Among all the packets newly ACKed or SACKed in the connection, record the most recent Packet.xmit_ts in RACK.xmit_ts if it is ahead of RACK.xmit_ts. Sometimes the timestamps of RACK.Packet and Packet could carry the same transmit timestamps due to clock granularity or segmentation offloading (i.e. the two packets were sent as a jumbo frame into the NIC). In that case the sequence numbers of RACK.end_seq and Packet.end_seq are compared to break the tie.
Since an ACK can also acknowledge retransmitted data packets, RACK.RTT can be vastly underestimated if the retransmission was spurious. To avoid that, ignore a packet if any of its TCP sequences have been retransmitted before and either of two conditions is true:
If the ACK is not ignored as invalid, update the RACK.RTT to be the RTT sample calculated using this ACK, and continue. If this ACK or SACK was for the most recently sent packet, then record the RACK.xmit_ts timestamp and RACK.end_seq sequence implied by this ACK. Otherwise exit here and omit the following steps.
Step 2 may be summarized in pseudocode as:
RACK_sent_after(t1, seq1, t2, seq2):
If t1 > t2:
Return true
Else if t1 == t2 AND seq1 > seq2:
Return true
Else:
Return false
RACK_update():
For each Packet newly acknowledged cumulatively or selectively:
rtt = Now() - RACK.xmit_ts
If Packet has been retransmitted:
If ACK.ts_option.echo_reply < Packet.xmit_ts:
Return
If rtt < RACK.min_rtt:
Return
RACK.RTT = rtt
If RACK_sent_after(Packet.xmit_ts, Packet.end_seq
RACK.xmit_ts, RACK.end_seq):
RACK.xmit_ts = Packet.xmit_ts
RACK.end_seq = Packet.end_seq
Step 3: Update RACK reordering window
To handle the prevalent small degree of reordering, RACK.reo_wnd serves as an allowance for settling time before marking a packet lost. Use a conservative window of min_RTT / 4 if the connection is not currently in loss recovery. When in loss recovery, use a RACK.reo_wnd of zero in order to retransmit quickly.
Extension 1: Optionally size the window based on DSACK Further, the sender MAY leverage DSACK [RFC3708] to adapt the reordering window to higher degrees of reordering. Receiving an ACK with a DSACK indicates a spurious retransmission, which in turn suggests that the RACK reordering window, RACK.reo_wnd, is likely too small. The sender MAY increase the RACK.reo_wnd window linearly for every round trip in which the sender receives a DSACK, so that after N distinct round trips in which a DSACK is received, the RACK.reo_wnd is N * min_RTT / 4. The inflated RACK.reo_wnd would persist for 16 loss recoveries and then reset to its starting value, min_RTT / 4.
Extension 2: Optionally size the window if reordering has been observed
If the reordering window is too small or the connection does not support DSACK, then RACK can trigger spurious loss recoveries and reduce the congestion window unnecessarily. If the implementation supports reordering detection such as [REORDER-DETECT], then the sender MAY use the dynamically-sized reordering window based on min_RTT during loss recovery instead of a zero reordering window to compensate. Extension 3: Optionally size the window with the classic DUPACK threshold heuristic The DUPACK threshold approach in the current standards [RFC5681][RFC6675] is simple, and for decades has been effective in quickly detecting losses, despite the drawbacks discussed earlier. RACK can easily maintain the DUPACK threshold's advantages of quick detection by resetting the reordering window to zero (using RACK.reo_wnd = 0) when the DUPACK threshold is met (i.e. when at least three packets have been selectively acknowledged). The subtle differences are discussed in the section “RACK and TLP discussions”.
The following algorithm includes the basic and all the extensions mentioned above. Note that individual extensions that require additional TCP features (e.g. DSACK) would work if the feature functions simply return false.
RACK_update_reo_wnd:
RACK.min_RTT = TCP_min_RTT()
If RACK_ext_TCP_ACK_has_DSACK_option():
RACK.dsack = true
If SND.UNA < RACK.roundtrip_seq:
RACK.dsack = false /* React to DSACK once within a round trip */
If RACK.dsack:
RACK.reo_wnd_incr += 1
RACK.dsack = false
RACK.roundtrip_seq = SND.NXT
RACK.reo_wnd_persist = 16 /* Keep window for 16 loss recoveries */
Else if exiting loss recovery:
RACK.reo_wnd_persist -= 1
If RACK.reo_wnd_persist <= 0:
RACK.reo_wnd_incr = 1
If in loss recovery and not RACK_ext_TCP_seen_reordering():
RACK.reo_wnd = 0
Else if RACK_ext_TCP_dupack_threshold_hit(): /* DUPTHRESH emulation mode */
RACK.reo_wnd = 0
Else:
RACK.reo_wnd = RACK.min_RTT / 4 * RACK.reo_wnd_incr
RACK.reo_wnd = min(RACK.reo_wnd, SRTT)
Step 4: Detect losses.
For each packet that has not been SACKed, if RACK.xmit_ts is after Packet.xmit_ts + RACK.reo_wnd, then mark the packet (or its corresponding sequence range) lost in the scoreboard. The rationale is that if another packet that was sent later has been delivered, and the reordering window or "reordering settling time" has already passed, then the packet was likely lost.
If another packet that was sent later has been delivered, but the reordering window has not passed, then it is not yet safe to deem the unacked packet lost. Using the basic algorithm above, the sender would wait for the next ACK to further advance RACK.xmit_ts; but this risks a timeout (RTO) if no more ACKs come back (e.g, due to losses or application limit). For timely loss detection, the sender MAY install a "reordering settling" timer set to fire at the earliest moment at which it is safe to conclude that some packet is lost. The earliest moment is the time it takes to expire the reordering window of the earliest unacked packet in flight.
This timer expiration value can be derived as follows. As a starting point, we consider that the reordering window has passed if the RACK.packet was sent sufficiently after the packet in question, or a sufficient time has elapsed since the RACK.packet was S/ACKed, or some combination of the two. More precisely, RACK marks a packet as lost if the reordering window for a packet has elapsed through the sum of:
So we mark a packet as lost if:
RACK.xmit_ts >= Packet.xmit_ts AND (RACK.xmit_ts - Packet.xmit_ts) + (now - RACK.ack_ts) >= RACK.reo_wnd
If we solve this second condition for "now", the moment at which we can declare a packet lost, then we get:
now >= Packet.xmit_ts + RACK.reo_wnd + (RACK.ack_ts - RACK.xmit_ts)
Then (RACK.ack_ts - RACK.xmit_ts) is just the RTT of the packet we used to set RACK.xmit_ts, so this reduces to:
Packet.xmit_ts + RACK.RTT + RACK.reo_wnd - now <= 0
The following pseudocode implements the algorithm above. When an ACK is received or the RACK timer expires, call RACK_detect_loss(). The algorithm includes an additional optimization to break timestamp ties by using the TCP sequence space. The optimization is particularly useful to detect losses in a timely manner with TCP Segmentation Offload, where multiple packets in one TSO blob have identical timestamps. It is also useful when the timestamp clock granularity is close to or longer than the actual round trip time.
RACK_detect_loss():
timeout = 0
For each packet, Packet, in the scoreboard:
If Packet is already SACKed
or marked lost and not yet retransmitted:
Continue
If RACK_sent_after(RACK.xmit_ts, RACK.end_seq,
Packet.xmit_ts, Packet.end_seq):
remaining = Packet.xmit_ts + RACK.RTT + RACK.reo_wnd - Now()
If remaining <= 0:
Mark Packet lost
Else:
timeout = max(remaining, timeout)
If timeout != 0
Arm a timer to call RACK_detect_loss() after timeout
Implementation optimization: looping through packets in the SACK scoreboard above could be very costly on large BDP networks since the inflight could be very large. If the implementation can organize the scoreboard data structures to have packets sorted by the last (re)transmission time, then the loop can start on the least recently sent packet and aborts on the first packet sent after RACK.time_ts. This can be implemented by using a seperate list sorted in time order. The implementation inserts the packet to the tail of the list when it is (re)transmitted, and removes a packet from the list when it is delivered or marked lost. We RECOMMEND such an optimization for implementations for support high BDP networks. The optimization is implemented in Linux and sees orders of magnitude improvement on CPU usage on high speed WAN networks.
Tail Loss Probe: fast recovery on tail losses
This section describes a supplemental algorithm, Tail Loss Probe (TLP), which leverages RACK to further reduce RTO recoveries. TLP triggers fast recovery to quickly repair tail losses that can otherwise be recovered by RTOs only. After an original data transmission, TLP sends a probe data segment within one to two RTTs. The probe data segment can either be new, previously unsent data, or a retransmission of previously sent data just below SND.NXT. In either case the goal is to elicit more feedback from the receiver, in the form of an ACK (potentially with SACK blocks), to allow RACK to trigger fast recovery instead of an RTO.
An RTO occurs when the first unacknowledged sequence number is not acknowledged after a conservative period of time has elapsed [RFC6298]. Common causes of RTOs include:
Following is an example of TLP. All events listed are at a TCP sender.
We define the terminology used in specifying the TLP algorithm:
FlightSize: amount of outstanding data in the network, as defined in [RFC5681].
RTO: The transport's retransmission timeout (RTO) is based on measured round-trip times (RTT) between the sender and receiver, as specified in [RFC6298] for TCP. PTO: Probe timeout (PTO) is a timer event indicating that an ACK is overdue. Its value is constrained to be smaller than or equal to an RTO.
SRTT: smoothed round-trip time, computed as specified in [RFC6298].
Open state: the sender's loss recovery state machine is in its normal, default state: there are no SACKed sequence ranges in the SACK scoreboard, and neither fast recovery, timeout-based recovery, nor ECN-based cwnd reduction are underway.
The TLP algorithm has three phases, which we discuss in turn.
Step 1: Check conditions for scheduling a PTO.
A sender should check to see if it should schedule a PTO in two situations:
A sender should schedule a PTO only if all of the following conditions are met:
If a PTO cannot be scheduled according to these conditions, then the sender MUST arm the RTO timer if there is unacknowledged data in flight.
Step 2: Select the duration of the PTO.
A sender SHOULD use the following logic to select the duration of a PTO:
TLP_timeout():
If SRTT is available:
PTO = 2 * SRTT
If FlightSize = 1:
PTO += WCDelAckT
Else:
PTO += 2ms
Else:
PTO = 1 sec
If Now() + PTO > TCP_RTO_expire():
PTO = TCP_RTO_expire() - Now()
Aiming for a PTO value of 2*SRTT allows a sender to wait long enough to know that an ACK is overdue. Under normal circumstances, i.e. no losses, an ACK typically arrives in one SRTT. But choosing PTO to be exactly an SRTT is likely to generate spurious probes given that network delay variance and even end-system timings can easily push an ACK to be above an SRTT. We chose PTO to be the next integral multiple of SRTT.
Similarly, current end-system processing latencies and timer granularities can easily delay ACKs, so senders SHOULD add at least 2ms to a computed PTO value (and MAY add more if the sending host OS timer granularity is more coarse than 1ms).
WCDelAckT stands for worst case delayed ACK timer. When FlightSize is 1, PTO is inflated by WCDelAckT time to compensate for a potential long delayed ACK timer at the receiver. The RECOMMENDED value for WCDelAckT is 200ms.
Finally, if the time at which an RTO would fire (here denoted "TCP_RTO_expire") is sooner than the computed time for the PTO, then a probe is scheduled to be sent at that earlier time..
When the PTO fires, transmit a probe data segment:
TLP_send_probe():
If a previously unsent segment exists AND
the receive window allows new data to be sent:
Transmit that new segment
FlightSize += SMSS
Else:
Retransmit the last segment
The cwnd remains unchanged
On each incoming ACK, the sender should cancel any existing loss probe timer. The sender should then reschedule the loss probe timer if the conditions in Step 1 of Phase 1 allow.
If the only loss in an outstanding window of data was the last segment, then a TLP loss probe retransmission of that data segment might repair the loss. TLP recovery detection examines ACKs to detect when the probe might have repaired a loss, and thus allows congestion control to properly reduce the congestion window (cwnd) [RFC5681].
Consider a TLP retransmission episode where a sender retransmits a tail packet in a flight. The TLP retransmission episode ends when the sender receives an ACK with a SEG.ACK above the SND.NXT at the time the episode started. During the TLP retransmission episode the sender checks for a duplicate ACK or D-SACK indicating that both the original segment and TLP retransmission arrived at the receiver, meaning there was no loss that needed repairing. If the TLP sender does not receive such an indication before the end of the TLP retransmission episode, then it MUST estimate that either the original data segment or the TLP retransmission were lost, and congestion control MUST react appropriately to that loss as it would any other loss.
Since a significant fraction of the hosts that support SACK do not support duplicate selective acknowledgments (D-SACKs) [RFC2883] the TLP algorithm for detecting such lost segments relies only on basic SACK support [RFC2018].
Definitions of variables
TLPRxtOut: a boolean indicating whether there is an unacknowledged TLP retransmission.
TLPHighRxt: the value of SND.NXT at the time of sending a TLP retransmission.
When a connection is created, or suffers a retransmission timeout, or enters fast recovery, it executes the following:
TLPRxtOut = false
Senders must only send a TLP loss probe retransmission if TLPRxtOut is false. This ensures that at any given time a connection has at most one outstanding TLP retransmission. This allows the sender to use the algorithm described in this section to estimate whether any data segments were lost.
Note that this condition only restricts TLP loss probes that are retransmissions. There may be an arbitrary number of outstanding unacknowledged TLP loss probes that consist of new, previously-unsent data, since the retransmission timeout and fast recovery algorithms are sufficient to detect losses of such probe segments.
Upon sending a TLP probe that is a retransmission, the sender sets TLPRxtOut to true and TLPHighRxt to SND.NXT.
Detecting recoveries accomplished by loss probes
Step 1: Track ACKs indicating receipt of original and retransmitted segments
A sender considers both the original segment and TLP probe retransmission segment as acknowledged if either 1 or 2 are true:
If neither conditions are met, then the sender estimates that the receiver received both the original data segment and the TLP probe retransmission, and so the sender considers the TLP episode to be done, and records that fact by setting TLPRxtOut to false.
Step 2: Mark the end of a TLP retransmission episode and detect losses
If the sender receives a cumulative ACK for data beyond the TLP loss probe retransmission then, in the absence of reordering on the return path of ACKs, it should have received any ACKs for the original segment and TLP probe retransmission segment. At that time, if the TLPRxtOut flag is still true and thus indicates that the TLP probe retransmission remains unacknowledged, then the sender should presume that at least one of its data segments was lost, so it SHOULD invoke a congestion control response equivalent to fast recovery.
More precisely, on each ACK the sender executes the following:
if (TLPRxtOut and SEG.ACK >= TLPHighRxt) {
TLPRxtOut = false
EnterRecovery()
ExitRecovery()
}
The biggest advantage of RACK is that every data packet, whether it is an original data transmission or a retransmission, can be used to detect losses of the packets sent chronologically prior to it.
Example: TAIL DROP. Consider a sender that transmits a window of three data packets (P1, P2, P3), and P1 and P3 are lost. Suppose the transmission of each packet is at least RACK.reo_wnd (1 millisecond by default) after the transmission of the previous packet. RACK will mark P1 as lost when the SACK of P2 is received, and this will trigger the retransmission of P1 as R1. When R1 is cumulatively acknowledged, RACK will mark P3 as lost and the sender will retransmit P3 as R3. This example illustrates how RACK is able to repair certain drops at the tail of a transaction without any timer. Notice that neither the conventional duplicate ACK threshold [RFC5681], nor [RFC6675], nor the Forward Acknowledgment [FACK] algorithm can detect such losses, because of the required packet or sequence count.
Example: LOST RETRANSMIT. Consider a window of three data packets (P1, P2, P3) that are sent; P1 and P2 are dropped. Suppose the transmission of each packet is at least RACK.reo_wnd (1 millisecond by default) after the transmission of the previous packet. When P3 is SACKed, RACK will mark P1 and P2 lost and they will be retransmitted as R1 and R2. Suppose R1 is lost again but R2 is SACKed; RACK will mark R1 lost for retransmission again. Again, neither the conventional three duplicate ACK threshold approach, nor [RFC6675], nor the Forward Acknowledgment [FACK] algorithm can detect such losses. And such a lost retransmission is very common when TCP is being rate-limited, particularly by token bucket policers with large bucket depth and low rate limit. Retransmissions are often lost repeatedly because standard congestion control requires multiple round trips to reduce the rate below the policed rate.
Example: SMALL DEGREE OF REORDERING. Consider a common reordering event: a window of packets are sent as (P1, P2, P3). P1 and P2 carry a full payload of MSS octets, but P3 has only a 1-octet payload. Suppose the sender has detected reordering previously (e.g., by implementing the algorithm in [REORDER-DETECT]) and thus RACK.reo_wnd is min_RTT/4. Now P3 is reordered and delivered first, before P1 and P2. As long as P1 and P2 are delivered within min_RTT/4, RACK will not consider P1 and P2 lost. But if P1 and P2 are delivered outside the reordering window, then RACK will still falsely mark P1 and P2 lost. We discuss how to reduce false positives in the end of this section.
The examples above show that RACK is particularly useful when the sender is limited by the application, which is common for interactive, request/response traffic. Similarly, RACK still works when the sender is limited by the receive window, which is common for applications that use the receive window to throttle the sender.
For some implementations (e.g., Linux), RACK works quite efficiently with TCP Segmentation Offload (TSO). RACK always marks the entire TSO blob lost because the packets in the same TSO blob have the same transmission timestamp. By contrast, the counting based algorithms (e.g., [RFC3517][RFC5681]) may mark only a subset of packets in the TSO blob lost, forcing the stack to perform expensive fragmentation of the TSO blob, or to selectively tag individual packets lost in the scoreboard.
RACK requires the sender to record the transmission time of each packet sent at a clock granularity of one millisecond or finer. TCP implementations that record this already for RTT estimation do not require any new per-packet state. But implementations that are not yet recording packet transmission times will need to add per-packet internal state (commonly either 4 or 8 octets per packet or TSO blob) to track transmission times. In contrast, the conventional [RFC6675] loss detection approach does not require any per-packet state beyond the SACK scoreboard. This is particularly useful on ultra-low RTT networks where the RTT is far less than the sender TCP clock grainularity (e.g. inside data-centers).
RACK can easily and optionally support the conventional approach in [RFC6675][RFC5681] by resetting the reordering window to zero when the threshold is met. Note that this approach differs slightly from [RFC6675] which considers a packet lost when at least #DupThresh higher-sequenc packets are SACKed. RACK’s approach considers a packet lost when at least one higher sequence packet is SACKed and the total number of SACKed packets is at least DupThresh. For example, suppose a connection sends 10 packets, and packets 3, 5, 7 are SACKed. [RFC6675] considers packets 1 and 2 lost. RACK considers packets 1, 2, 4, 6 lost.
When the sender detects packet reordering, RACK uses a reordering window of min_rtt / 4. It uses the minimum RTT to accommodate reordering introduced by packets traversing slightly different paths (e.g., router-based parallelism schemes) or out-of-order deliveries in the lower link layer (e.g., wireless links using link-layer retransmission). RACK uses a quarter of minimum RTT because Linux TCP used the same factor in its implementation to delay Early Retransmit [RFC5827] to reduce spurious loss detections in the presence of reordering, and experience shows that this seems to work reasonably well. We have evaluated using the smoothed RTT (SRTT from [RFC6298] RTT estimation) or the most recently measured RTT (RACK.RTT) using an experiment similar to that in the Performance Evaluation section. They do not make any significant difference in terms of total recovery latency.
The primary motivation of RACK is to ultimately provide a simple and general replacement for some of the standard loss recovery algorithms [RFC5681][RFC6675][RFC5827][RFC4653], as well as some nonstandard ones [FACK][THIN-STREAM]. While RACK can be a supplemental loss detection mechanism on top of these algorithms, this is not necessary, because RACK implicitly subsumes most of them.
[RFC5827][RFC4653][THIN-STREAM] dynamically adjusts the duplicate ACK threshold based on the current or previous flight sizes. RACK takes a different approach, by using only one ACK event and a reordering window. RACK can be seen as an extended Early Retransmit [RFC5827] without a FlightSize limit but with an additional reordering window. [FACK] considers an original packet to be lost when its sequence range is sufficiently far below the highest SACKed sequence. In some sense RACK can be seen as a generalized form of FACK that operates in time space instead of sequence space, enabling it to better handle reordering, application-limited traffic, and lost retransmissions.
Nevertheless RACK is still an experimental algorithm. Since the oldest loss detection algorithm, the 3 duplicate ACK threshold [RFC5681], has been standardized and widely deployed. RACK can easily and optionally support the conventional approach for compatibility.
RACK is compatible with and does not interfere with the the standard RTO [RFC6298], RTO-restart [RFC7765], F-RTO [RFC5682] and Eifel algorithms [RFC3522]. This is because RACK only detects loss by using ACK events. It neither changes the RTO timer calculation nor detects spurious timeouts.
Furthermore, RACK naturally works well with Tail Loss Probe [TLP] because a tail loss probe solicits either an ACK or SACK, which can be used by RACK to detect more losses. RACK can be used to relax TLP’s requirement for using FACK and retransmitting the the highest-sequenced packet, because RACK is agnostic to packet sequence numbers, and uses transmission time instead. Thus TLP could be modified to retransmit the first unacknowledged packet, which could improve application latency.
RACK intentionally decouples loss detection from congestion control. RACK only detects losses; it does not modify the congestion control algorithm [RFC5681][RFC6937]. However, RACK may detect losses earlier or later than the conventional duplicate ACK threshold approach does. A packet marked lost by RACK SHOULD NOT be retransmitted until congestion control deems this appropriate. Specifically, Proportional Rate Reduction [RFC6937] SHOULD be used when using RACK.
RACK is applicable for both fast recovery and recovery after a retransmission timeout (RTO) in [RFC5681]. RACK applies equally to fast recovery and RTO recovery because RACK is purely based on the transmission time order of packets. When a packet retransmitted by RTO is acknowledged, RACK will mark any unacked packet sent sufficiently prior to the RTO as lost, because at least one RTT has elapsed since these packets were sent.
The following simple example compares how RACK and non-RACK loss detection interacts with congestion control: suppose a TCP sender has a congestion window (cwnd) of 20 packets on a SACK-enabled connection. It sends 10 data packets and all of them are lost.
Without RACK, the sender would time out, reset cwnd to 1, and retransmit the first packet. It would take four round trips (1 + 2 + 4 + 3 = 10) to retransmit all the 10 lost packets using slow start. The recovery latency would be RTO + 4*RTT, with an ending cwnd of 4 packets due to congestion window validation.
With RACK, a sender would send the TLP after 2*RTT and get a DUPACK. If the sender implements Proportional Rate Reduction [RFC6937] it would slow start to retransmit the remaining 9 lost packets since the number of packets in flight (0) is lower than the slow start threshold (10). The slow start would again take four round trips (1 + 2 + 4 + 3 = 10). The recovery latency would be 2*RTT + 4*RTT, with an ending cwnd set to the slow start threshold of 10 packets.
In both cases, the sender after the recovery would be in congestion avoidance. The difference in recovery latency (RTO + 4*RTT vs 6*RTT) can be significant if the RTT is much smaller than the minimum RTO (1 second in RFC6298) or if the RTT is large. The former case is common in local area networks, data-center networks, or content distribution networks with deep deployments. The latter case is more common in developing regions with highly congested and/or high-latency networks. The ending congestion window after recovery also impacts subsequent data transfer.
Delayed ACKs complicate the detection of repairs done by TLP, since with a delayed ACK the sender receives one fewer ACK than would normally be expected. To mitigate this complication, before sending a TLP loss probe retransmission, the sender should attempt to wait long enough that the receiver has sent any delayed ACKs that it is withholding. The sender algorithm described above features such a delay, in the form of WCDelAckT. Furthermore, if the receiver supports duplicate selective acknowledgments (D-SACKs) [RFC2883] then in the case of a delayed ACK the sender's TLP recovery detection algorithm (see above) can use the D-SACK information to infer that the original and TLP retransmission both arrived at the receiver.
If there is ACK loss or a delayed ACK without a D-SACK, then this algorithm is conservative, because the sender will reduce cwnd when in fact there was no packet loss. In practice this is acceptable, and potentially even desirable: if there is reverse path congestion then reducing cwnd can be prudent.
RACK can be implemented in other transport protocols. The algorithm can be simplified by skipping step 3 if the protocol can support a unique transmission or packet identifier (e.g. TCP echo options). For example, the QUIC protocol implements RACK [QUIC-LR].
RACK and TLP have been deployed at Google, for both connections to users in the Internet and internally. We conducted a performance evaluation experiment for RACK and TLP on a small set of Google Web servers in Western Europe that serve mostly European and some African countries. The experiment lasted three days in March 2017. The servers were divided evenly into four groups of roughly 5.3 million flows each:
Group 1 (control): RACK off, TLP off, RFC 3517 on
Group 2: RACK on, TLP off, RFC 3517 on
Group 3: RACK on, TLP on, RFC 3517 on
Group 4: RACK on, TLP on, RFC 3517 off
All groups used Linux with CUBIC congestion control, an initial congestion window of 10 packets, and the fq/pacing qdisc. In terms of specific recovery features, all groups enabled RFC5682 (F-RTO) but disabled FACK because it is not an IETF RFC. FACK was excluded because the goal of this setup is to compare RACK and TLP to RFC-based loss recoveries. Since TLP depends on either FACK or RACK, we could not run another group that enables TLP only (with both RACK and FACK disabled). Group 4 is to test whether RACK plus TLP can completely replace the DupThresh-based [RFC3517].
The servers sit behind a load balancer that distributes the connections evenly across the four groups.
Each group handles a similar number of connections and sends and receives similar amounts of data. We compare total time spent in loss recovery across groups. The recovery time is measured from when the recovery and retransmission starts, until the remote host has acknowledged the highest sequence (SND.NXT) at the time the recovery started. Therefore the recovery includes both fast recoveries and timeout recoveries.
Our data shows that Group 2 recovery latency is only 0.3% lower than the Group 1 recovery latency. But Group 3 recovery latency is 25% lower than Group 1 due to a 40% reduction in RTO-triggered recoveries! Therefore it is important to implement both TLP and RACK for performance. Group 4’s total recovery latency is 0.02% lower than Group 3’s, indicating that RACK plus TLP can successfully replace RFC3517 as a standalone recovery mechanism.
We want to emphasize that the current experiment is limited in terms of network coverage. The connectivity in Western Europe is fairly good, therefore loss recovery is not a major performance bottleneck. We plan to expand our experiments to regions with worse connectivity, in particular on networks with strong traffic policing.
RACK does not change the risk profile for TCP.
An interesting scenario is ACK-splitting attacks [SCWA99]: for an MSS-size packet sent, the receiver or the attacker might send MSS ACKs that SACK or acknowledge one additional byte per ACK. This would not fool RACK. RACK.xmit_ts would not advance because all the sequences of the packet are transmitted at the same time (carry the same transmission timestamp). In other words, SACKing only one byte of a packet or SACKing the packet in entirety have the same effect on RACK.
This document makes no request of IANA.
Note to RFC Editor: this section may be removed on publication as an RFC.
The authors thank Matt Mathis for his insights in FACK and Michael Welzl for his per-packet timer idea that inspired this work. Eric Dumazet, Randy Stewart, Van Jacobson, Ian Swett, Rick Jones, Jana Iyengar, and Hiren Panchasara contributed to the draft and the implementations in Linux, FreeBSD and QUIC.