| Transport Area Working Group | B. Briscoe, Ed. |
| Internet-Draft | Simula Research Lab |
| Intended status: Informational | K. De Schepper |
| Expires: September 14, 2017 | Nokia Bell Labs |
| M. Bagnulo Braun | |
| Universidad Carlos III de Madrid | |
| March 13, 2017 |
Low Latency, Low Loss, Scalable Throughput (L4S) Internet Service: Architecture
draft-briscoe-tsvwg-l4s-arch-01
This document describes the L4S architecture for the provision of a new service that the Internet could provide to eventually replace best efforts for all traffic: Low Latency, Low Loss, Scalable throughput (L4S). It is becoming common for all (or most) applications being run by a user at any one time to require low latency. However, the only solution the IETF can offer for ultra-low queuing delay is Diffserv, which only favours a minority of packets at the expense of others. In extensive testing the new L4S service keeps average queuing delay under a millisecond for all applications even under very heavy load, without sacrificing utilization; and it keeps congestion loss to zero. It is becoming widely recognized that adding more access capacity gives diminishing returns, because latency is becoming the critical problem. Even with a high capacity broadband access, the reduced latency of L4S remarkably and consistently improves performance under load for applications such as interactive video, conversational video, voice, Web, gaming, instant messaging, remote desktop and cloud-based apps (even when all being used at once over the same access link). The insight is that the root cause of queuing delay is in TCP, not in the queue. By fixing the sending TCP (and other transports) queuing latency becomes so much better than today that operators will want to deploy the network part of L4S to enable new products and services. Further, the network part is simple to deploy - incrementally with zero-config. Both parts, sender and network, ensure coexistence with other legacy traffic. At the same time L4S solves the long-recognized problem with the future scalability of TCP throughput.
This document describes the L4S architecture, briefly describing the different components and how the work together to provide the aforementioned enhanced Internet service.
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 http://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 14, 2017.
Copyright (c) 2017 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 (http://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.
It is increasingly common for all of a user's applications at any one time to require low delay: interactive Web, Web services, voice, conversational video, interactive video, instant messaging, online gaming, remote desktop and cloud-based applications. In the last decade or so, much has been done to reduce propagation delay by placing caches or servers closer to users. However, queuing remains a major, albeit intermittent, component of latency. When present it typically doubles the path delay from that due to the base speed-of-light. Low loss is also important because, for interactive applications, losses translate into even longer retransmission delays.
It has been demonstrated that, once access network bit rates reach levels now common in the developed world, increasing capacity offers diminishing returns if latency (delay) is not addressed. Differentiated services (Diffserv) offers Expedited Forwarding [RFC3246] for some packets at the expense of others, but this is not applicable when all (or most) of a user's applications require low latency.
Therefore, the goal is an Internet service with ultra-Low queueing Latency, ultra-Low Loss and Scalable throughput (L4S) - for all traffic. A service for all traffic will need none of the configuration or management baggage (traffic policing, traffic contracts) associated with favouring some packets over others. This document describes the L4S architecture for achieving that goal.
It must be said that queuing delay only degrades performance infrequently [Hohlfeld14]. It only occurs when a large enough capacity-seeking (e.g. TCP) flow is running alongside the user's traffic in the bottleneck link, which is typically in the access network. Or when the low latency application is itself a large capacity-seeking flow (e.g. interactive video). At these times, the performance improvement must be so remarkable that network operators will be motivated to deploy it.
Active Queue Management (AQM) is part of the solution to queuing under load. AQM improves performance for all traffic, but there is a limit to how much queuing delay can be reduced by solely changing the network; without addressing the root of the problem.
The root of the problem is the presence of standard TCP congestion control (Reno [RFC5681]) or compatible variants (e.g. TCP Cubic [I-D.ietf-tcpm-cubic]). We shall call this family of congestion controls 'Classic' TCP. It has been demonstrated that if the sending host replaces Classic TCP with a 'Scalable' alternative, when a suitable AQM is deployed in the network the performance under load of all the above interactive applications can be stunningly improved. For instance, queuing delay under heavy load with the example DCTCP/DualQ solution cited below is roughly 1 millisecond (1 ms) at the 99th percentile without losing link utilization. This compares with 5 to 20 ms on average with a Classic TCP and current state-of-the-art AQMs such as fq_CoDel [I-D.ietf-aqm-fq-codel] or PIE [RFC8033]. Also, with a Classic TCP, 5 ms of queuing is usually only possible by losing some utilization.
It has been convincingly demonstrated [DCttH15] that it is possible to deploy such an L4S service alongside the existing best efforts service so that all of a user's applications can shift to it when their stack is updated. Access networks are typically designed with one link as the bottleneck for each site (which might be a home, small enterprise or mobile device), so deployment at a single node should give nearly all the benefit. The L4S approach requires a number of mechanisms in different parts of the Internet to fulfill its goal. This document presents the L4S architecture, by describing the different components and how they interact to provide the scalable low-latency, low-loss, Internet service.
There are three main components to the L4S architecture (illustrated in Figure 1):
(2) (1)
.-------^------. .--------------^-------------------.
,-(3)-----. ______
; ________ : L4S --------. | |
:|Scalable| : _\ ||___\_| mark |
:| sender | : __________ / / || / |______|\ _________
:|________|\; | |/ --------' ^ \1| |
`---------'\_| IP-ECN | Coupling : \|priority |_\
________ / |Classifier| : /|scheduler| /
|Classic |/ |__________|\ --------. ___:__ / |_________|
| sender | \_\ || | |||___\_| mark/|/
|________| / || | ||| / | drop |
Classic --------' |______|
Figure 1: Components of an L4S Solution: 1) Isolation in separate network queues; 2) Packet Identification Protocol; and 3) Scalable Sending Host
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119]. In this document, these words will appear with that interpretation only when in ALL CAPS. Lower case uses of these words are not to be interpreted as carrying RFC-2119 significance. COMMENT: Since this will be an information document, This should be removed.
The L4S architecture is composed by the following elements.
Protocols:The L4S architecture encompass the two protocol changes that we describe next:
Network components:The Dual Queue Coupled AQM has been specified as generically as possible [I-D.briscoe-aqm-dualq-coupled] as a 'semi-permeable' membrane without specifying the particular AQMs to use in the two queues. An informational appendix of the draft is provided for pseudocode examples of different possible AQM approaches. Initially a zero-config variant of RED called Curvy RED was implemented, tested and documented. The aim is for designers to be free to implement diverse ideas. So the brief normative body of the draft only specifies the minimum constraints an AQM needs to comply with to ensure that the L4S and Classic services will coexist. For instance, a variant of PIE called Dual PI Squared [PI2] has been implemented and found to perform better over a wide range of conditions, so it has been documented in a second appendix of [I-D.briscoe-aqm-dualq-coupled].
Host mechanisms: The L4S architecture includes a number of mechanisms in the end host that we enumerate next:
All the following approaches address some part of the same problem space as L4S. In each case, it is shown that L4S complements them or improves on them, rather than being a mutually exclusive alternative:
A transport layer that solves the current latency issues will provide new service, product and application opportunities.
With the L4S approach, the following existing applications will immediately experience significantly better quality of experience under load in the best effort class:
The significantly lower queuing latency also enables some interactive application functions to be offloaded to the cloud that would hardly even be usable today:
The above two applications have been successfully demonstrated with L4S, both running together over a 40 Mb/s broadband access link loaded up with the numerous other latency sensitive applications in the previous list as well as numerous downloads. A panoramic video of a football stadium can be swiped and pinched so that on the fly a proxy in the cloud generates a sub-window of the match video under the finger-gesture control of each user. At the same time, a virtual reality headset fed from a 360 degree camera in a racing car has been demonstrated, where the user's head movements control the scene generated in the cloud. In both cases, with 7 ms end-to-end base delay, the additional queuing delay of roughly 1 ms is so low that it seems the video is generated locally. See https://riteproject.eu/dctth/ for videos of these demonstrations.
Using a swiping finger gesture or head movement to pan a video are extremely demanding applications—far more demanding than VoIP. Because human vision can detect extremely low delays of the order of single milliseconds when delay is translated into a visual lag between a video and a reference point (the finger or the orientation of the head).
If low network delay is not available, all fine interaction has to be done locally and therefore much more redundant data has to be downloaded. When all interactive processing can be done in the cloud, only the data to be rendered for the end user needs to be sent. Whereas, once applications can rely on minimal queues in the network, they can focus on reducing their own latency by only minimizing the application send queue.
The following use-cases for L4S are being considered by various interested parties:
{ToDo: This section TBA - currently, bullet points only.}
Incremental deployment parts.
Possible deployment sequences.
Prioritizing the most-likely bottlenecks in the various use-cases (access links, downstream and upstream, broadband, mobile, DC, etc).
Deployment incentives: Immediate vs. deferred benefits.
This specification contains no IANA considerations.
Because the L4S service can serve all traffic that is using the capacity of a link, it should not be necessary to police access to the L4S service. In contrast, Diffserv only works if some packets get less favourable treatement than others. So it has to use traffic policers to limit how much traffic can be favoured, In turn, traffic policers require traffic contracts between users and networks as well as pairwise between networks. Because L4S will lack all this management complexity, it is more likely to work end-to-end.
During early deployment (and perhaps always), some networks will not offer the L4S service. These networks do not need to police or re-mark L4S traffic - they just forward it unchanged as best efforts traffic, as they would already forward traffic with ECT(1) today. At a bottleneck, such networks will introduce some queuing and dropping. When a scalable congestion control detects a drop it will have to respond as if it is a Classic congestion control (see item 3-1 in Appendix A). This will ensure safe interworking with other traffic at the 'legacy' bottleneck, but it will degrade the L4S service to no better (but never worse) than classic best efforts, whenever a legacy (non-L4S) bottleneck is encountered on a path.
Certain network operators might choose to restict access to the L4S class, perhaps only to customers who have paid a premium. Their packet classifer (item 2 in Figure 1) could identify such customers against some other field (e.g. source address range) as well as ECN. If only the ECN L4S identifier matched, but not the source address (say), the classifier could direct these packets (from non-paying customers) into the Classic queue. Allowing operators to use an additional local classifier is intended to remove any incentive to bleach the L4S identifier. Then at least the L4S ECN identifier will be more likely to survive end-to-end even though the service may not be supported at every hop. Such arrangements would only require simple registered/not-registered packet classification, rather than the managed application-specific traffic policing against customer-specific traffic contracts that Diffserv requires.
The L4S service does rely on self-constraint - not in terms of limiting capacity usage, but in terms of limiting burstiness. It is hoped that standardisation of dynamic behaviour (cf. TCP slow-start) and self-interest will be sufficient to prevent transports from sending excessive bursts of L4S traffic, given the application's own latency will suffer most from such behaviour.
Whether burst policing becomes necessary remains to be seen. Without it, there will be potential for attacks on the low latency of the L4S service. However it may only be necessary to apply such policing reactively, e.g. punitively targeted at any deployments of new bursty malware.
As mentioned in Section 5.2, L4S should remove the need for low latency Diffserv classes. However, those Diffserv classes that give certain applications or users priority over capacity, would still be applicable. Then, within such Diffserv classes, L4S would often be applicable to give traffic low latency and low loss. WIthin such a class, the bandwidth available to a user or application is often limited by a rate policer. Similarly, in the default Diffserv class, rate policers are used to partition shared capacity.
A classic rate policer drops any packets exceeding a set rate, usually also giving a burst allowance (variant exist where the policer re-marks non-compliant traffic to a discard-eligible Diffserv codepoint, so they may be dropped elsewhere during contention). In networks that deploy L4S and use rate policers, it will be preferable to deploy a policer designed to be more friendly to the L4S service,
This might be achieved by setting a threshold where ECN marking is introduced, such that it is just under the policed rate or just under the burst allowance where drop is introduced. This could be applied to various types of policer, e.g. [RFC2697], [RFC2698] or the local (non-ConEx) variant of the ConEx congestion policer [I-D.briscoe-conex-policing]. Otherwise, whenever L4S traffic encounters a rate policer, it will experience drops and the source will fall back to a Classic congestion control, thus losing all the benefits of L4S.
Further discussion of the applicability of L4S to the various Diffserv classes, and the design of suitable L4S rate policers.
Receiving hosts can fool a sender into downloading faster by suppressing feedback of ECN marks (or of losses if retransmissions are not necessary or available otherwise). [RFC3540] proposes that a TCP sender could pseudorandomly set either of ECT(0) or ECT(1) in each packet of a flow and remember the sequence it had set, termed the ECN nonce. If the receiver supports the nonce, it can prove that it is not suppressing feedback by reflecting its knowledge of the sequence back to the sender. The nonce was proposed on the assumption that receivers might be more likely to cheat congestion control than senders (although senders also have a motive to cheat).
If L4S uses the ECT(1) codepoint of ECN for packet classification, it will have to obsolete the experimental nonce. As far as is known, the ECN Nonce has never been deployed, and it was only implemented for a couple of testbed evaluations. It would be nearly impossible to deploy now, because any misbehaving receiver can simply opt-out, which would be unremarkable given all receivers currently opt-out.
Other ways to protect TCP feedback integrity have since been developed. For instance:
Thanks to Wes Eddy, Karen Nielsen and David Black for their useful review comments.
| [RFC2119] | Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997. |
This list contains a list of features, mechanisms and modifications from currently defined behaviour for scalable Transport protocols so that they can be safely deployed over the public Internet. This list of requirements was produced at an ad hoc meeting during IETF-94 in Prague [TCPPrague].
One of such scalable transport protocols is DCTCP, currently specified in [I-D.ietf-tcpm-dctcp]. In its current form, DCTCP is specified to be deployable in controlled environments and deploying it in the public Internet would lead to a number of issues, both from the safety and the performance perspective. In this section, we describe the modifications and additional mechanisms that are required for its deployment over the global Internet. We use DCTCP as a base, but it is likely that most of these requirements equally apply to other scalable transport protocols.
We next provide a brief description of each required feature.
Requirement #4.1: Fall back to Reno/Cubic congestion control on packet loss.
Description: In case of packet loss, the scalable transport MUST react as classic TCP (whatever the classic version of TCP is running in the host, e.g. Reno, Cubic).
Motivation: As part of the safety conditions for deploying a scalable transport over the public Internet is to make sure that it behaves properly when some or all the network devices connecting the two endpoints that implement the scalable transport have not been upgraded. In particular, it may be the case that some of the switches along the path between the two endpoints may only react to congestion by dropping packets (i.e. no ECN marking). It is important that in these cases, the scalable transport react to the congestion signal in the form of a packet drop similarly to classic TCP.
In the particular case of DCTCP, the current DCTCP specification states that "It is RECOMMENDED that an implementation deal with loss episodes in the same way as conventional TCP." For safe deployment in the public Internet of a scalable transport, the above requirement needs to be defined as a MUST.
Packet loss, while rare, may also occur in the case that the bottleneck is L4S capable. In this case, the sender may receive a high number of packets marked with the CE bit set and also experience a loss. Current DCTCP implementations react differently to this situation. At least one implementation reacts only to the drop signal (e.g. by halving the CWND) and at least another DCTCP implementation reacts to both signals (e.g. by halving the CWND due to the drop and also further reducing the CWND based on the proportion of marked packet). We believe that further experimentation is needed to understand what is the best behaviour for the public Internet, which may or not be one of the existent implementations.
Requirement #4.2: Fall back to Reno/Cubic congestion control on classic ECN bottlenecks.
Description: The scalable transport protocol SHOULD/MAY? behave as classic TCP with classic ECN if the path contains a legacy bottleneck which marks both ect(0) and ect(1) in the same way as drop (non L4S, but ECN capable bottleneck).
Motivation: Similarly to Requirement #3.1, this requirement is a safety condition in case L4S-capable endpoints are communicating over a path that contains one or more non-L4S but ECN capable switches and one of them happens to be the bottleneck. In this case, the scalable transport will attempt to fill in the buffer of the bottleneck switch up to the marking threshold and produce a small sawtooth around that operation point. The result is that the switch will set its operation point with the buffer full and all other non-scalable transports will be starved (as they will react reducing their CWND more aggressively than the scalable transport).
Scalable transports then MUST be able to detect the presence of a classic ECN bottleneck and fall back to classic TCP/classic ECN behaviour in this case.
Discussion: It is not clear at this point if it is possible to design a mechanism that always detect the aforementioned cases. One possibility is to base the detection on an increase on top of a minimum RTT, but it is not yet clear which value should trigger this. Having a delay based fall back response on L4S may as well be beneficial for preserving low latency without legacy network nodes. Even if it possible to design such a mechanism, it may well be that it would encompass additional complexity that implementers may consider unnecessary. The need for this mechanism depends on the extent of classic ECN deployment.
Requirement #4.3: Reduce RTT dependence
Description: Scalable transport congestion control algorithms MUST reduce or eliminate the RTT bias within the range of RTTs available.
Motivation: Classic TCP's throughput is known to be inversely proportional to RTT. One would expect flows over very low RTT paths to nearly starve flows over larger RTTs. However, because Classic TCP induces a large queue, it has never allowed a very low RTT path to exist, so far. For instance, consider two paths with base RTT 1ms and 100ms. If Classic TCP induces a 20ms queue, it turns these RTTs into 21ms and 120ms leading to a throughput ratio of about 1:6. Whereas if a Scalable TCP induces only a 1ms queue, the ratio is 2:101. Therefore, with small queues, long RTT flows will essentially starve.
Scalable transport protocol MUST then accommodate flows across the range of RTTs enabled by the deployment of L4S service over the public Internet.
Requirement #4.4: Scaling down the congestion window.
Description: Scalable transports MUST be responsive to congestion when RTTs are significantly smaller than in the current public Internet.
Motivation: As currently specified, the minimum CWND of TCP (and the scalable extensions such as DCTCP), is set to 2 MSS. Once this minimum CWND is reached, the transport protocol ceases to react to congestion signals (the CWND is not further reduced beyond this minimum size).
L4S mechanisms reduce significantly the queueing delay, achieving smaller RTTs over the Internet. For the same CWND, smaller RTTs imply higher transmission rates. The result is that when scalable transport are used and small RTTs are achieved, the minimum value of the CWND currently defined in 2 MSS may still result in a high transmission rate for a large number of common scenarios. For example, as described in [TCP-sub-mss-w], consider a residential setting with an broadband Internet access of 40Mbps. Suppose now a number of equal TCP flows running in parallel with the Internet access link being the bottleneck. Suppose that for these flows, the RTT is 6ms and the MSS is 1500B. The minimum transmission rate supported by TCP in this scenario is when CWND is set to 2 MSS, which results in 4Mbps for each flow. This means that in this scenario, if the number of flows is higher than 10, the congestion control ceases to be responsive and starts to build up a queue in the network.
In order to address this issue, the congestion control mechanism for scalable transports MUST be responsive for the new range of RTT resulting from the decrease of the queueing delay.
There are several ways how this can be achieved. One possible sub-MSS window mechanism is described in [TCP-sub-mss-w].
In addition to the safety requirements described before, there are some optimizations that while not required for the safe deployment of scalable transports over the public Internet, would results in an optimized performance. We describe them next.
Optimization #5.1: Setting ECT in SYN, SYN/ACK and pure ACK packets.
Description: Scalable transport SHOULD set the ECT bit in SYN, SYN/ACK and pure ACK packets.
Motivation: Failing to set the ECT bit in SYN, SYN/ACK or ACK packets results in these packets being more likely dropped during congestion events. Dropping SYN and SYN/ACK packets is particularly bad for performance as the retransmission timers for these packets are large. [RFC3168] prevents from marking these packets due to security reasons. The arguments provided should be revisited in the the context of L4S and evaluate if avoiding marking these packets is still the best approach.
Optimization #5.2: Faster than additive increase.
Description: Scalable transport MAY support faster than additive increase in the congestion avoidance phase.
Motivation: As currently defined, DCTCP supports additive increase in congestion avoidance phase. It would be beneficial for performance to update the congestion control algorithm to increase the CWND more than 1 MSS per RTT during the congestion avoidance phase. In the context of L4S such mechanism, must also provide fairness with other classes of traffic, including classic TCP and possibly scalable TCP that uses additive increase.
Optimization #5.3: Faster convergence to fairness.
Description: Scalable transport SHOULD converge to a fair share allocation of the available capacity as fast as classic TCP or faster.
Motivation: The time required for a new flow to obtain its fair share of the capacity of the bottleneck when the there are already ongoing flows using up all the bottleneck capacity is higher in the case of DCTCP than in the case of classic TCP (about a factor of 1,5 and 2 larger according to [Alizadeh-stability]). This is detrimental in general, but it is very harmful for short flows, which performance can be worse than the one obtained with classic TCP. for this reason it is desirable that scalable transport provide convergence times no larger than classic TCP.
The following table includes all the itmes that should be standardized to provide a full L4S architecture.
The table is too wide for the ASCII draft format, so it has been split into two, with a common column of row index numbers on the left.
The columns in the second part of the table have the following meanings:
| Req # | Requirement | Reference |
|---|---|---|
| 0 | ARCHITECTURE | |
| 1 | L4S IDENTIFIER | [I-D.briscoe-tsvwg-ecn-l4s-id] |
| 2 | DUAL QUEUE AQM | [I-D.briscoe-aqm-dualq-coupled] |
| 3 | Suitable ECN Feedback | [I-D.ietf-tcpm-accurate-ecn], [I-D.stewart-tsvwg-sctpecn]. |
| SCALABLE TRANSPORT - SAFETY ADDITIONS | ||
| 4-1 | Fall back to Reno/Cubic on loss | [I-D.ietf-tcpm-dctcp] |
| 4-2 | Fall back to Reno/Cubic if classic ECN bottleneck detected | |
| 4-3 | Reduce RTT-dependence | |
| 4-4 | Scaling TCP's Congestion Window for Small Round Trip Times | [TCP-sub-mss-w] |
| SCALABLE TRANSPORT - PERFORMANCE ENHANCEMENTS | ||
| 5-1 | Setting ECT in SYN, SYN/ACK and pure ACK packets | draft-bagnulo-tsvwg-generalized-ECN |
| 5-2 | Faster-than-additive increase | |
| 5-3 | Less drastic exit from slow-start |
| # | WG | TCP | DCTCP | DCTCP-bis | TCP Prague | SCTP Prague | RMCAT Prague |
|---|---|---|---|---|---|---|---|
| 0 | tsvwg? | Y | Y | Y | Y | Y | Y |
| 1 | tsvwg? | Y | Y | Y | Y | ||
| 2 | aqm? | n/a | n/a | n/a | n/a | n/a | n/a |
| 3 | tcpm | Y | Y | Y | Y | n/a | n/a |
| 4-1 | tcpm | Y | Y | Y | Y | Y | |
| 4-2 | tcpm/ iccrg? | Y | Y | ? | |||
| 4-3 | tcpm/ iccrg? | Y | Y | Y | ? | ||
| 4-4 | tcpm | Y | Y | Y | Y | Y | ? |
| 5-1 | tsvwg | Y | Y | Y | Y | n/a | n/a |
| 5-2 | tcpm/ iccrg? | Y | Y | Y | ? | ||
| 5-3 | tcpm/ iccrg? | Y | Y | Y | ? |