| Network Working Group | Y. Nishida |
| Internet-Draft | GE Global Research |
| Intended status: Standards Track | P. Natarajan |
| Expires: September 10, 2015 | Cisco Systems |
| A. Caro | |
| BBN Technologies | |
| P. Amer | |
| University of Delaware | |
| K. Nielsen | |
| Ericsson | |
| March 9, 2015 |
SCTP-PF: Quick Failover Algorithm in SCTP
draft-ietf-tsvwg-sctp-failover-10.txt
One of the major advantages of SCTP is the support of multi-homed communication. A multi-homed SCTP end-point has the ability to withstand network failures by migrating the traffic from an inactive network to an active one. However, if the failover operation as specified in RFC4960 is followed, there can be a significant delay in the migration to the active destination addresses, thus severely reducing the effectiveness of the SCTP failover operation.
This document complements RFC4960 by the introduction of a new path state, the Potentially Failed (PF) path state, and an associated new failover operation to apply during a network failure. The algorithm defined is called SCTP Potentially Failed Algorithm, SCTP-PF for short. In addition, the document complements RFC4960 by introducing alternative switchover operation modes for the data transfer path management after the recovery of a failed primary path. These modes can allow improvements in the performance of the operation in some network environments. The implementation of the additional switchover operation modes is an optional part of SCTP-PF.
The procedures defined in the document require only minimal modifications to the current specification. The procedures are sender-side only and do not impact the SCTP receiver.
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 10, 2015.
Copyright (c) 2015 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.
The Stream Control Transmission Protocol (SCTP) as specified in [RFC4960] supports multihoming at the transport layer -- an SCTP endpoint can bind to multiple IP addresses. SCTP's multihoming features include failure detection and failover procedures to provide network interface redundancy and improved end-to-end fault tolerance.
In SCTP's current failure detection procedure, the sender must experience Path.Max.Retrans (PMR) number of consecutive failed timer-based retransmissions on a destination address before detecting a path failure. The sender fails over to an alternate active destination address only after failure detection. Until detecting the failover, the sender continues to transmit data on the failed path, which degrades the SCTP performance. Concurrent Multipath Transfer (CMT) [IYENGAR06] is an proposed extension to SCTP that allows the sender to transmit data on multiple paths simultaneously. Research [NATARAJAN09] shows that the current failure detection procedure worsens CMT performance during failover and can be significantly improved by employing a better failover algorithm.
This document specifies an alternative failure detection and failover procedure, the SCTP Potentially Failed algorithm, that improves the performance of SCTP multi-homed operation during a failover.
For multi-homed SCTP the operation after the recovery of a failed path equally well impacts the performance of the protocol. With the procedures specified in [RFC4960], SCTP will, after a failover from the primary path, switch back to the primary path for data transfer as soon as this path becomes available again. From a performance perspective, as confirmed in research [CARO02], such a switchback of the data transmission path is not optimal in general. As an optional alternative to the switchback operation of [RFC4960], this document specifies the Permanent Failover procedures proposed by [CARO02].
Additional discussion for alternative approaches that do not require modifications to [RFC4960], as well as discussion of path bouncing effects that might be caused by frequent switchover, are provided in the Appendices.
While the Potentially Failed algorithm primarily is motivated for improvement of the SCTP multi-homed operation, the feature applies also to SCTP single-homed operation. Here the algorithm serves to provide increased failure detection on idle associations, whereas the failover or switchback aspects of the algorithm will not be activated. This is discussed in more detail in Appendix C.
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].
This section describes issues in the SCTP as specified in [RFC4960] to be fixed by the approach described in this document.
An SCTP endpoint can support multiple IP addresses. Each SCTP endpoint exchanges the list of its usable addresses during the initial negotiation with its peer. Then the endpoints select one address from the peer's list and use this as the primary destination address. During normal transmission, an SCTP endpoint sends all user data to the primary destination address. Also, it sends packets containing a HEARTBEAT chunk to all idle destination addresses at a certain interval to check the reachability of these destination addresses. Idle destination addresses normally include all non-primary destination addresses.
If a sender has multiple active destination addresses, it can retransmit data to an non-primary destination address, if the transmission to the primary times out.
When a sender receives an acknowledgment for DATA or HEARTBEAT chunks sent to one of the destination addresses, it considers that destination address to be active and clears the error counter for the destination address. If it fails to receive acknowledgments, the error count for the destination address is increased. If the error counter exceeds the tunable protocol parameter Path.Max.Retrans (PMR), the SCTP endpoint considers the destination address to be inactive.
The failover process of SCTP is initiated when the primary path becomes inactive (the error counter for the primary path exceeds Path.Max.Retrans). If the primary path is marked inactive, SCTP chooses a new destination address from one of the active destinations and starts using this as the destination address for sending data. If the primary path becomes active again, SCTP reverts to using the primary destination address for subsequent data transmissions and stop using the non-primary one.
One issue with this failover process defined in [RFC4960] is that it usually takes a significant amount of time before SCTP switches to the new destination address. Let's say the primary path on a multi-homed host becomes unavailable and the RTO value for the primary path at that time is around 1 second, it usually takes over 60 seconds before SCTP starts to use the non-primary path for initial data transmission. This is because the recommended value for Path.Max.Retrans in the [RFC4960] is 5, which requires 6 consecutive timeouts before the failover takes place. Before SCTP switches to the non-primary address, SCTP keeps trying to send packets to the primary address and only retransmitted packets are sent to the non-primary address and thus can be received by the receiver. This slow failover process can cause significant performance degradation and is not acceptable in some situations.
Another issue with RFC4960 failover and switchback operation is that once the primary path becomes active again, the traffic is unconditionally switched back to use this path. This is not optimal in some situations. This is further discussed in Section 4.3.
To address the issues described in Section 3, this document extends SCTP path management scheme by adding the Potentially Failed state and associated protocol operation. The algorithm is called SCTP Potentially Failed algorithm. SCTP-PF for short. The resulting SCTP path management operation is called SCTP Potentially Failed operation.
The introduction of the Potentially Failed state stems from the following two observations about SCTP's failure detection procedure:
From the above observations it is clear that tuning the PMR value involves the following trade off -- a lower value improves performance but increases the chances of spurious failure detection, whereas a higher value degrades performance and reduces spurious failure detection in a wide range of path conditions. Thus, tuning the association's PMR value is an incomplete solution to address the performance impact during failure.
SCTP-PF defined in this document introduces the new Potentially Failed (PF) destination address state in SCTP's path management procedure. The new Potentially Failed (PF) destination address state applies to SCTP single-homed operation as well as to SCTP multi-homed operation. The PF state was originally proposed to improve CMT performance [NATARAJAN09]. The PF state is an intermediate state between the Active and Failed states. SCTP's failure detection procedure is modified to include the PF state. The new failure detection algorithm assumes that loss detected by a timeout implies either severe congestion or failure en-route. After a number of consecutive timeouts on a path, the sender is unsure, and marks the corresponding destination address as in the PF state. A PF destination address is not used for data transmission except when it is the only destination address available (e.g., for single-homed SCTP) or in other special cases (discussed below). The new failure detection algorithm requires only sender-side changes.
The SCTP-PF operation is specified as follows:
When choosing among multiple destination address in active state the following considerations are given:
In all cases, the sender MUST NOT change the state of chosen destination address, whether this state be active or PF, and it MUST NOT clear the error counter of the destination address as a result of choosing the destination address for data transmission.
The sender MUST NOT change the state and the error counter of any destination address regardless of whether it has been chosen for transmission or not.
In a situation with complete disruption of the communication in between the SCTP Endpoints, the aggressive HEARTBEAT transmissions of SCTP-PF on destination addresses in PF state may make the association enter dormant state faster than a standard [RFC4960] SCTP implementation given the same setting of Path.Max.Retrans (PMR) and Association.Max.Retrans (AMR). For example, an SCTP association with two destination addresses typically would reach dormant state in half the time of an [RFC4960] SCTP implementation in such situations. This is because a SCTP PF sender will send HEARTBEATS and data retransmissions in parallel with RTO intervals when there are multiple destinations addresses in PF state. This argument pressumes that RTO << HB.interval of [RFC4960]. One could use higher values of PMR, which makes the dormant state situations less likely to happen. The downside of increasing the PMR value is that destination address failure detections and notifications of such events to ULP is weakened.
A design goal of SCTP-PF is that it should provide the same level of disruption tolerance as an [RFC4960] SCTP implementation with the same Path.Max.Retrans (PMR) and Association.Max.Retrans (AMR) setting. For this reason, SCTP-PF SHOULD perform the following operations during dormant state, while this is an implementation decision in [RFC4960].
An SCTP-PF implementation MAY keep the operation during dormant state an implementation decision, but it should be careful not to compromise the fault tolerance of the SCTP operation.
The above prescriptions for SCTP-PF dormant state handling SHOULD NOT be coupled to the value of the PFMR, but solely to the activation of SCTP-PF logic in an SCTP implementation. It is further noted that also a standard [RFC4960] SCTP implementation can use this mode of operation to improve the fault tolerance (which some implementations already do).
This section describes an OPTIONAL switchback feature called Permanent Failover which is beneficiary to deploy in certain situations.
In [RFC4960], an SCTP sender migrates the traffic back to the original primary destination address once this address becomes active again. As the CWND towards the original primary destination address has to be rebuilt once data transfer resumes, the switch back to use the original primary address is not always optimal. Indeed [CARO02] shows that the switch back to the original primary may degrade SCTP performance compared to continuing data transmission on the same path, especially, but not only, in scenarios where this path's characteristics are better. In order to mitigate this performance degradation, the Permanent Failover operation was proposed in [CARO02]. When SCTP changes the destination address due to failover, Permanent Failover operation allows SCTP sender to continue data transmission on the new working path even when the old primary destination address becomes active again. This is achieved by having SCTP perform a switch over of the primary path to the alternative working path rather than having SCTP switch back data transfer to the (previous) primary path.
The manner of switch over operation that is most optimal in a given scenario depends on the relative quality of a set primary path versus the quality of alternative paths available as well as it depends on the extent to which it is desired for the mode of operation to enforce traffic distribution over a number of network paths. I.e., load distribution of traffic from multiple SCTP associations may be sought to be enforced by distribution of the set primary paths with [RFC4960] switchback operation. However as [RFC4960] switchback behavior is suboptimal in certain situations, especially in scenarios where a number of equally good paths are available, it is recommended for SCTP to support also, as alternative behavior, the Permanent Failover switch over modes of operation.
The Permanent Failover operation requires only sender side changes. The details are:
To support optimal operation in a wider range of network scenarios, it it proposed for an SCTP-PF implementation to implement Permanent Failover operation as an optional feature. The implementation of the Permanent Failover feature is optional for an SCTP-PF implementation. For an SCTP implementation that implements Permanent Failover, this specification RECOMMENDS that the standard RFC4960 switchback operation is retained as the default operation.
This section describes how the socket API defined in [RFC6458] is extended to provide a way for the application to control and observe the SCTP-PF behavior.
Please note that this section is informational only.
A socket API implementation based on [RFC6458] is, by means of the existing SCTP_PEER_ADDR_CHANGE event, extended to provide the event notification when a peer address enters or leaves the potentially failed state as well as the socket API implementation is extended to expose the potentially failed state of a peer address in the existing SCTP_GET_PEER_ADDR_INFO structure.
Furthermore, two new read/write socket options for the level IPPROTO_SCTP and the name SCTP_PEER_ADDR_THLDS and SCTP_EXPOSE_POTENTIALLY_FAILED_STATE are defined as described below. The first socket option is used to control the values of the PFMR and PSMR parameters described in Section 4. The second one controls the exposition of the potentially failed path state.
Support for the SCTP_PEER_ADDR_THLDS and SCTP_EXPOSE_POTENTIALLY_FAILED_STATE socket options need also to be added to the function sctp_opt_info().
As defined in [RFC6458], the SCTP_PEER_ADDR_CHANGE event is provided if the status of a peer address changes. In addition to the state changes described in [RFC6458], this event is also provided, if a peer address enters or leaves the potentially failed state. The notification as defined in [RFC6458] uses the following structure:
struct sctp_paddr_change {
uint16_t spc_type;
uint16_t spc_flags;
uint32_t spc_length;
struct sockaddr_storage spc_aaddr;
uint32_t spc_state;
uint32_t spc_error;
sctp_assoc_t spc_assoc_id;
}
[RFC6458] defines the constants SCTP_ADDR_AVAILABLE, SCTP_ADDR_UNREACHABLE, SCTP_ADDR_REMOVED, SCTP_ADDR_ADDED, and SCTP_ADDR_MADE_PRIM to be provided in the spc_state field. This document defines in addition to that the new constant SCTP_ADDR_POTENTIALLY_FAILED, which is reported if the affected address becomes potentially failed.
The SCTP_GET_PEER_ADDR_INFO socket option defined in [RFC6458] can be used to query the state of a peer address. It uses the following structure:
struct sctp_paddrinfo {
sctp_assoc_t spinfo_assoc_id;
struct sockaddr_storage spinfo_address;
int32_t spinfo_state;
uint32_t spinfo_cwnd;
uint32_t spinfo_srtt;
uint32_t spinfo_rto;
uint32_t spinfo_mtu;
};
[RFC6458] defines the constants SCTP_UNCONFIRMED, SCTP_ACTIVE, and SCTP_INACTIVE to be provided in the spinfo_state field. This document defines in addition to that the new constant SCTP_POTENTIALLY_FAILED, which is reported if the peer address is potentially failed.
Applications can control the SCTP-PF behavior by getting or setting the number of consecutive timeouts before a peer address is considered potentially failed or unreachable and before the primary path is changed automatically. This socket option uses the level IPPROTO_SCTP and the name SCTP_PEER_ADDR_THLDS.
The following structure is used to access and modify the thresholds:
struct sctp_paddrthlds {
sctp_assoc_t spt_assoc_id;
struct sockaddr_storage spt_address;
uint16_t spt_pathmaxrxt;
uint16_t spt_pathpfthld;
uint16_t spt_pathcpthld;
};
Applications can control the exposure of the potentially failed path state in the SCTP_PEER_ADDR_CHANGE event and the SCTP_GET_PEER_ADDR_INFO as described in Section 5.1. The default value is implementation specific.
This socket option uses the level IPPROTO_SCTP and the name SCTP_EXPOSE_POTENTIALLY_FAILED_STATE.
The following structure is used to control the exposition of the potentially failed path state:
struct sctp_assoc_value {
sctp_assoc_t assoc_id;
uint32_t assoc_value;
};
Security considerations for the use of SCTP and its APIs are discussed in [RFC4960] and [RFC6458]. The logic described here is for sender-side only enabled by configuration and does not have any impacts on protocol messages on the wire. No new chunk type or new field parameter is not required in this document.
This document does not create any new registries or modify the rules for any existing registries managed by IANA.
Initially this work looked to entail some changes of the Congestion Control (CC) operation of SCTP and for this reason the work was proposed as Experimental. These intended changes of the CC operation have since been judged to be irrelevant and are no longer part of the specification. As the specification entails no other potential harmful features, consensus exists in the WG to bring the work forward as PS.
Initially concerns have been expressed about the possibility for the mechanism to introduce path bouncing with potential harmful network impacts. These concerns are believed to be unfounded. This issue is addressed in Appendix B.
It is noted that the feature specified by this document is implemented by multiple SCTP SW implementations and furthermore that various variants of the solution have been deployed in Telco signaling environments for several years with good results.
| [RFC2119] | Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997. |
| [RFC4960] | Stewart, R., "Stream Control Transmission Protocol", RFC 4960, September 2007. |
This section lists alternative approaches for the issues described in this document. Although these approaches do not require to update RFC4960, we do not recommend them from the reasons described below.
Smaller values for Path.Max.Retrans shorten the failover duration. In fact, this is recommended in some research results [JUNGMAIER02] [GRINNEMO04] [FALLON08]. For example, if when Path.Max.Retrans=0, SCTP switches to another destination address on a single timeout. This smaller value for Path.Max.Retrans can results in spurious failover, which might be a problem.
Unlike SCTP-PF, the interval for heartbeat packets is governed by 'HB.interval' even during failover process. 'HB.interval' is usually set in the order of seconds (recommended value is 30 seconds). When the primary path becomes inactive, the next HEARTBEAT can be transmitted only seconds later. Meanwhile, the primary path may have recovered. In such situations, post failover, an endpoint is forced to wait on the order of seconds before the endpoint can resume transmission on the primary path. However, using smaller value for 'HB.interval' might help this situation, but it will be the waste of bandwidth in most cases.
In addition, smaller Path.Max.Retrans values also affect 'Association.Max.Retrans' values. When the SCTP association's error count (sum of error counts on all ACTIVE paths) exceeds Association.Max.Retrans threshold, the SCTP sender considers the peer endpoint unreachable and terminates the association. Therefore, Section 8.2 in [RFC4960] recommends that Association.Max.Retrans value should not be larger than the summation of the Path.Max.Retrans of each of the destination addresses, else the SCTP sender considers its peer reachable even when all destinations are INACTIVE. To avoid such inconsistent behavior an SCTP implementation SHOULD reduce Association.Max.Retrans accordingly whenever it reduces Path.Max.Retrans. However, smaller Association.Max.Retrans value increases chances of association termination during minor congestion events.
As several research results indicate, we can also shorten the duration of failover process by adjusting RTO related parameters [JUNGMAIER02] [FALLON08]. During failover process, RTO keeps being doubled. However, if we can choose smaller value for RTO.max, we can stop the exponential growth of RTO at some point. Also, choosing smaller values for RTO.initial or RTO.min can contribute to keep RTO value small.
Similar to reducing Path.Max.Retrans, the advantage of this approach is that it requires no modification to the current specification, although it needs to ignore several recommendations described in the Section 15 of [RFC4960]. However, this approach requires to have enough knowledge about the network characteristics between end points. Otherwise, it can introduce adverse side-effects such as spurious timeouts.
The methods described in the document can accelerate the failover process. Hence, they might introduce the path bouncing effect where the sender keeps changing the data transmission path frequently. This sounds harmful to the data transfer, however several research results indicate that there is no serious problem with SCTP in terms of path bouncing effect [CARO04] [CARO05].
There are two main reasons for this. First, SCTP is basically designed for multipath communication, which means SCTP maintains all path related parameters (CWND, ssthresh, RTT, error count, etc) per each destination address. These parameters cannot be affected by path bouncing. In addition, when SCTP migrates the data transfer to another path, it starts with the minimal or the initial CWND. Hence, there is little chance for packet reordering or duplicating.
Second, even if all communication paths between the end-nodes share the same bottleneck, the SCTP-PF results in a behavior already allowed by [RFC4960].
For a single-homed SCTP association the only tangible effect of the activation of SCTP-PF operation is enhanced failure detection in terms of potential notification of the PF state of the sole destination address as well as, for idle associations, more rapid entering, and notification, of inactive state of the destination address and more rapid end-point failure detection. It is believed that neither of these effects are harmful, provided adequate dormant state operation is implemented, and furthermore that they may be particularly useful for applications that deploys multiple SCTP associations for load balancing purposes. The early notification of the PF state may be used for preventive measures as the entering of the PF state can be used as a warning of potential congestion. Depending on the PMR value, the aggressive HEARTBEAT transmission in PF state may speed up the end-point failure detection (exceed of AMR threshold on the sole path error counter) on idle associations in case where relatively large HB.interval value compared to RTO (e.g. 30secs) is used.