Load Sharing for the Stream Control Transmission Protocol (SCTP)
University of Delaware,
Computer and Information Sciences DepartmentNewarkDelaware19716U.S.A.+1-302-831-1944amer@cis.udel.eduhttp://www.eecis.udel.edu/~amer/
HAW Hamburg,
Informatics DepartmentBerliner Tor 720099 HamburgHamburgGermany+49-40-42875-8104+49-40-42875-8309martin.becke@haw-hamburg.dehttp://www.scimbe.de/about.htmlSimula Research
Laboratory, Network Systems GroupMartin Linges vei 171364 FornebuAkershusNorway+47-6782-8200+47-6782-8201dreibh@simula.nohttp://www.iem.uni-due.de/~dreibh/
University of Delaware,
Computer and Information Sciences DepartmentNewarkDelaware19716U.S.A.nekiz@udel.eduhttp://www.eecis.udel.edu/~nekiz/
Franklin and Marshall College,
Mathematics and Computer SciencePO Box 3003LancasterPennsylvania17604-3003U.S.A.+1-717-358-4774jiyengar@fandm.eduhttp://www.fandm.edu/jiyengar/Cisco Systems425 East Tasman DriveSan JoseCalifornia95134U.S.A.prenatar@cisco.comAdara NetworksChapinSouth Carolina29036U.S.A.randall@lakerest.net
Muenster University of Applied SciencesStegerwaldstrasse 3948565 SteinfurtNordrhein-WestfalenGermanytuexen@fh-muenster.dehttps://www.fh-muenster.de/fb2/personen/professoren/tuexen/Internet-DraftThe Stream Control Transmission Protocol (SCTP) supports multi-homing for
providing network fault tolerance. However, mainly one path is used for data
transmission. Only timer-based retransmissions are carried over other paths
as well.This document describes how multiple paths can be used simultaneously for
transmitting user messages.One of the important features of the Stream Control Transmission Protocol
(SCTP), which is currently specified in , is network
fault tolerance. This feature is for example required for Reliable Server
Pooling (RSerPool, ). Therefore, transmitting
messages over multiple paths is supported, but only for redundancy. So
does not specify how to use multiple paths
simultaneously.This document overcomes this limitation by specifying how multiple paths
can be used simultaneously. This has several benefits:
Improved bandwidth usage.Better availability check with real user messages compared to
HEARTBEAT-based information.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 .Basic requirement for applying SCTP load sharing is the Concurrent
Multipath Transfer (CMT) extension of SCTP, which utilises multiple
paths simultaneously. We denote CMT-enabled SCTP as CMT-SCTP throughout
this document. CMT-SCTP is introduced in
and in more detail in ,
some illustrative examples of chunk handling are provided in
.
CMT-SCTP provides three modifications to standard SCTP (split Fast
Retransmissions, appropriate congestion window growth and delayed
SACKs), which are described in the following subsections.Paths with different latencies lead to overtaking of DATA chunks.
This leads to gap reports, which are handled by Fast Retransmissions.
However, due to the fact that multiple paths are used simultaneously,
these Fast Retransmissions are usually useless and furthermore lead to a
decreased congestion window size.To avoid unnecessary Fast Retransmissions, the sender has to keep
track of the path each DATA chunk has been sent on and consider
transmission paths before performing Fast Retransmissions. That is, on
reception of a SACK, the sender MUST identify the highest acknowledged
TSN on each path. A chunk SHOULD only be considered as missing if its
TSN is smaller than the highest acknowledged TSN on its path. Section
3.1 of contains an illustrated example.The congestion window adaptation algorithm for SCTP
allows increasing the congestion window only when a new cumulative ack (CumAck)
is received by a sender. When SACKs with unchanged CumAcks are generated (due to
reordering) and later arrive at a sender, the sender does not modify its
congestion window. Since a CMT-SCTP receiver naturally observes reordering, many
SACKs are sent containing new gap reports but not new CumAcks. When these gaps
are later acked by a new CumAck, congestion window growth occurs, but only for
the data newly acked in the most recent SACK. Data previously acked through gap
reports will not contribute to congestion window growth, in order to prevent
sudden increases in the congestion window resulting in bursts of data being
sent.To overcome the problems described above, the congestion window growth has to
be handled as follows :
The sender SHOULD keep track of the earliest non-retransmitted outstanding
TSN per path.The sender SHOULD keep track of the earliest retransmitted
outstanding TSN per path.The in-order delivery per path SHOULD be deduced.The congestion window of a path SHOULD be increased when the earliest
non-retransmitted outstanding TSN of this path is advanced
('Pseudo CumAck') OR when the earliest retransmitted outstanding TSN of this
path is advanced ('RTX Pseudo CumAck').Section 3.2 of contains an illustrated
example of appropriate congestion window handling for CMT-SCTP.Standard SCTP sends a SACK as soon as an
out-of-sequence TSN has been received. Delayed Acknowledgements are only
allowed if the received TSNs are in sequence. However, due to the load
balancing of CMT-SCTP, DATA chunks may overtake each other. This leads
to a high number of out-of-sequence TSNs, which have to be acknowledged
immediately. Clearly, this behaviour increases the overhead traffic
(usually nearly one SACK chunk for each received packet containing a
DATA chunk).Delayed Acknowledgements for CMT-SCTP are handled as follows:
In addition to , delaying of SACKs SHOULD
*also* be applied for out-of-sequence TSNs.A receiver MUST maintain a counter for the number of DATA chunks
received before sending a SACK. The value of the counter is stored
into each SACK chunk (FIXME: add details; needs reservation of flags
bits by IANA). After transmitting a SACK, the counter MUST be reset
to 0. Its initial value MUST be 0.The SACK handling procedure for a missing TSN M is extended as follows:
If all newly acknowledged TSNs have been transmitted
over the same path:
If there are newly acknowledged TSNs L and H so that L
<= M <= H, the missing count of TSN M SHOULD
be incremented by one (like for standard SCTP according to
).Else if all newly acknowledged TSNs N satisfy the
condition M <= N, the missing count of TSN M SHOULD
be incremented by the number of TSNs reported in the SACK
chunk.Otherwise (that is, there are newly acknowledged TSNs on
different paths), the missing count of TSN M SHOULD be
incremented by one (like for standard SCTP according to
).Section 3.3 of contains an illustrated
example of Delayed Acknowledgements for CMT-SCTP.
Before sending/receiving NR-SACKs (see ), both peer endpoints MUST agree on
using NR-SACKs. This agreement MUST be negotiated during association establishment.
NR-SACK is an extension to the core SCTP, and SCTP extensions that an endpoint
supports are reported to the peer endpoint in Supported Extensions
Parameter during association establishment (see Section 4.2.7 of
.)
The Supported Extensions Parameter consists of a list of non-standard Chunk Types
that are supported by the sender.
An endpoint supporting the NR-SACK extension MUST list the NR-SACK
chunk in the Supported Extensions Parameter carried in the INIT
or INIT-ACK chunk, depending on whether the endpoint initiates or
responds to the initiation of the association.
If the NR-SACK chunk type ID is listed in the Chunk Types List of the
Supported Extensions Parameter, then the receiving endpoint MUST assume that
the NR-SACK chunk is supported by the sending endpoint.
Both endpoints MUST support NR-SACKs for either endpoint to send an NR-SACK.
If an endpoint establishes an
association with a remote endpoint that does not list NR-SACK in the Supported
Extensions Parameter carried in INIT chunk, then both endpoints of
the association MUST NOT use NR-SACKs.
After association establishment, an endpoint MUST NOT renegotiate the use of NR-SACKs.
Once both endpoints indicate during association establishment
that they support the NR-SACK extension, each endpoint SHOULD
acknowledge received DATA chunks with NR-SACK chunks, and not SACK chunks.
That is, throughout an SCTP association, both endpoints SHOULD send either
SACK chunks or NR-SACK chunks, never a mixture of the two.
Table 1 illustrates a new chunk type that will be used to
transfer NR-SACK information.
As the NR-SACK chunk replaces the SACK chunk, many SACK chunk fields
are preserved in the NR-SACK chunk. These preserved fields
have the same semantics with the corresponding SACK chunk fields, as defined in
, Section 3.3.4. The Gap Ack fields from RFC4960 have been renamed as R Gap Ack
to emphasize their renegable nature. Their semantics are unchanged.
For completeness, we describe all fields of the NR-SACK chunk,
including those that are identical in the SACK chunk.
Similar to the SACK chunk, the NR-SACK chunk is sent to a peer endpoint to
(1) acknowledge DATA chunks received in-order, (2) acknowledge DATA chunks
received out-of-order, and (3) identify DATA chunks received more than once (i.e., duplicate.)
In addition, the NR-SACK chunk (4) informs
the peer endpoint of non-renegable out-of-order DATA chunks.
Type: 8 bits
This field holds the IANA defined chunk type for
NR-SACK chunk. The suggested value of this field for IANA is 0x10.
Chunk Flags: 8 bits
Currently not used. It is recommended a sender set all bits to zero on
transmit, and a receiver ignore this field.
Chunk Length: 16 bits (unsigned integer) [Same as SACK chunk]
This value represents the size of the chunk in bytes including the
Chunk Type, Chunk Flags, Chunk Length, and Chunk Value fields.
Cumulative TSN Ack: 32 bits (unsigned integer) [Same as SACK chunk]
The value of the Cumulative TSN Ack is the last TSN received before a break
in the sequence of received TSNs occurs. The next TSN value following the
Cumulative TSN Ack has not yet been received at the endpoint sending the NR-SACK.
Advertised Receiver Window Credit (a_rwnd):
32 bits (unsigned integer) [Same as SACK chunk]
Indicates the updated receive buffer space in bytes of
the sender of this NR-SACK, see Section 6.2.1 of
for details.
Number of (R)enegable Gap Ack Blocks (N): 16 bits (unsigned integer)
Indicates the number of Renegable Gap Ack Blocks included in this NR-SACK.
Number of (N)on(R)enegable Gap Ack Blocks (M): 16 bits (unsigned integer)
Indicates the number of Non-Renegable Gap Ack Blocks included in
this NR-SACK.
Number of Duplicate TSNs (X): 16 bits [Same as SACK chunk]
Contains the number of duplicate TSNs the endpoint has
received. Each duplicate TSN is listed following the NR Gap Ack Block list.
Reserved : 16 bits
Currently not used. It is recommended a sender set all bits to zero on
transmit, and a receiver ignore this field.
(R)enegable Gap Ack Blocks:
The NR-SACK contains zero or more R Gap Ack Blocks.
Each R Gap Ack Block acknowledges a subsequence of renegable out-of-order TSNs.
By definition, all TSNs acknowledged by R Gap Ack Blocks are "greater than" the value of the
Cumulative TSN Ack.
Because of TSN numbering wraparound, comparisons and all arithmetic
operations discussed in this document are based on
"Serial Number Arithmetic" as described in Section 1.6 of
.
R Gap Ack Blocks are repeated for each R Gap Ack Block up to 'N' defined in
the Number of R Gap Ack Blocks field. All DATA chunks with TSNs >=
(Cumulative TSN Ack + R Gap Ack Block Start) and <=
(Cumulative TSN Ack + R Gap Ack Block End) of each R Gap Ack Block are assumed
to have been received correctly, and are renegable.
R Gap Ack Block Start: 16 bits (unsigned integer)
Indicates the Start offset TSN for this R Gap Ack Block. This number
is set relative to the cumulative TSN number defined in Cumulative TSN Ack field.
To calculate the actual start TSN number, the Cumulative TSN Ack is added to
this offset number. The calculated TSN identifies the first TSN
in this R Gap Ack Block that has been received.
R Gap Ack Block End: 16 bits (unsigned integer)
Indicates the End offset TSN for this R Gap Ack Block. This number
is set relative to the cumulative TSN number defined in the Cumulative TSN Ack field.
To calculate the actual TSN number, the Cumulative TSN Ack is added to this
offset number. The calculated TSN identifies the TSN of the last DATA chunk received in this R Gap Ack Block.
N(on)R(enegable) Gap Ack Blocks:
The NR-SACK contains zero or more NR Gap Ack Blocks. Each NR
Gap Ack Block acknowledges a continuous subsequence of non-renegable
out-of-order DATA chunks.
If a TSN is nr-gap-acked in any NR-SACK chunk, then all subsequently transmitted
NR-SACKs with a smaller cum-ack value than that TSN SHOULD also nr-gap-ack that TSN.
NR Gap Ack Blocks are repeated for
each NR Gap Ack Block up to 'M' defined in the Number of NR Gap Ack Blocks
field. All DATA chunks with TSNs >= (Cumulative TSN
Ack + NR Gap Ack Block Start) and <= (Cumulative
TSN Ack + NR Gap Ack Block End) of each NR Gap Ack Block are assumed to be received correctly,
and are Non-Renegable.
NR Gap Ack Block Start: 16 bits (unsigned integer)
Indicates the Start offset TSN for this NR Gap Ack Block. This number
is set relative to the cumulative TSN number defined in Cumulative TSN Ack field.
To calculate the actual TSN number, the Cumulative TSN Ack is added to
this offset number. The calculated TSN identifies the first TSN
in this NR Gap Ack Block that has been received.
NR Gap Ack Block End: 16 bits (unsigned integer)
Indicates the End offset TSN for this NR Gap Ack Block.
This number is set relative to the cumulative TSN number defined
in Cumulative TSN Ack field. To calculate the actual TSN number, the Cumulative
TSN Ack is added to this offset number. The calculated TSN identifies the
TSN of the last DATA chunk received in this NR Gap Ack Block.
Note: NR Gap Ack Blocks and R Gap Ack Blocks in an NR-SACK chunk SHOULD acknowledge disjoint sets of
TSNs. That is, an out-of-order TSN SHOULD be listed in either an R Gap Ack Block or an NR Gap Ack Block,
but not the both. R Gap Ack Blocks and NR Gap Ack Blocks together provide the information as do the Gap
Ack Block of a SACK chunk, plus additional information about non-renegability.
If all out-of-order data acked by an NR-SACK are renegable, then the Number of NR Gap Ack Blocks MUST be set to 0.
If all out-of-order data acked by an NR-SACK are non-renegable, then the Number of R Gap Ack Blocks SHOULD be set
to 0. TSNs listed in R Gap Ack Block will be referred as r-gap-acked.
Duplicate TSN: 32 bits (unsigned integer) [Same as SACK chunk]
Indicates a duplicate TSN received since the last NR-SACK was sent.
Exactly 'X' duplicate TSNs SHOULD be reported where 'X' was defined in Number of
Duplicate TSNs field.
Each duplicate TSN is listed in this field as many times as
the TSN was received since the previous NR-SACK was sent.
For example, if a data receiver were to get the TSN 19 three times, the data receiver
would list 19 twice in the outbound NR-SACK. After sending the NR-SACK
if the receiver received one more TSN 19, the receiver would list 19 as a duplicate
once in the next outgoing NR-SACK.
Assume the following DATA chunks have arrived at the receiver.
The above figure shows the list of DATA chunks at the receiver.
TSN denotes the transmission sequence number of the DATA chunk,
SID denotes the stream id to which the DATA chunk belongs,
SSN denotes the sequence number of the DATA chunk within its stream,
and the U bit denotes whether the DATA chunk requires ordered(=0) or unordered(=1)
delivery . Note that TSNs 4,9,10, and 12 have not arrived.
This data can be viewed as three separate streams as follows
(assume each stream begins with SSN=0.)
Note that in this example, the application uses stream 2 for unordered data transfer.
By definition, SSN fields of unordered DATA chunks are ignored.
Stream-0: Stream-1: Stream-2: The NR-SACK to acknowledge the above data SHOULD be constructed as follows
for each of the three cases described below (the a_rwnd is arbitrarily set to 4000):
CASE-1: Minimal Data Receiver Responsibility - no out-of-order deliverable data yet delivered
None of the deliverable out-of-order DATA chunks have been delivered, and
the receiver of the above data does not take responsibility for any of the
received out-of-order DATA chunks. The receiver reserves the right to renege
any or all of the out-of-order DATA chunks.
CASE-2: Minimal Data Receiver Responsibility - all out-of-order deliverable data delivered
In this case, the NR-SACK chunk is being sent after the data receiver has delivered
all deliverable out-of-order DATA chunks to its receiving application(i.e., TSNs 5,6,7,8,13, and 16.)
The receiver reserves the right to renege on all undelivered out-of-order DATA chunks(i.e., TSNs 11,14, and 15.)
CASE-3: Maximal Data Receiver Responsibility
In this special case, all out-of-order data blocks acknowledged are non-renegable.
This case would occur when the data receiver is programmed never to renege, and takes
responsibility to deliver all DATA chunks that arrive out-of-order.
In this case Num of R Gap Ack Blocks is zero indicating all
reported out-of-order TSNs are nr-gap-acked.
The procedures regarding "when" to send an NR-SACK chunk are identical to
the procedures regarding when to send a SACK chunk,
as outlined in Section 6.2 of .
All of the NR-SACK chunk fields identical to the SACK chunk MUST be formed as
described in Section 6.2 of .
It is up to the data receiver whether or not to take responsibility for delivery
of each out-of-order DATA chunk. An out-of-order DATA chunk that has already been
delivered, or that the receiver takes responsibility to deliver (i.e., guarantees
not to renege) is Non Renegable(NR), and SHOULD be included in an NR
Gap Ack Block field of the outgoing NR-SACK. All other out-of-order data is (R)enegable, and SHOULD
be included in R Gap Ack Block field of the outgoing NR-SACK.
Consider three types of data receiver:
Data receiver takes no responsibility for delivery of any out-of-order DATA chunks
Data receiver takes responsibility for all out-of-order
DATA chunks that are "deliverable" (i.e., DATA chunks in-sequence within the
stream they belong to, or DATA chunks whose (U)nordered bit is 1)
Data receiver takes responsibility for delivery of all
out-of-order DATA chunks, whether deliverable or not deliverable
The data receiver SHOULD follow the procedures outlined
below for building the NR-SACK.CASE-1:Identify the TSNs received out-of-order.
For these out-of-order TSNs, identify the R Gap Ack Blocks.
Fill the Number of R Gap Ack Blocks (N) field,
R Gap Ack Block #i Start, and R Gap Ack Block #i End where i goes from 1 to N.
Set the Number of NR Gap Ack Blocks (M) field to 0.
CASE-2:Identify the TSNs received out-of-order.
For the received out-of-order TSNs, check the (U)nordered bit of each TSN.
Tag unordered TSNs as NR.
For each stream, also identify the TSNs received out-of-order but are
in-sequence within that stream.
Tag those in-sequence TSNs as NR.
Tag all out-of-order data that is not NR as (R)enegable.
For those TSNs tagged as (R)enegable, identify the (R)enegable Blocks.
Fill the Number of R Gap Ack Blocks(N) field,
R Gap Ack Block #i Start, and R Gap Ack Block #i End where i goes from 1 to N.
For those TSNs tagged as NR, identify the NR Blocks.
Fill the Number of NR Gap Ack Blocks(M) field,
NR Gap Ack Block #i Start, and NR Gap Ack Block #i End where i goes from 1 to M.
CASE-3:Identify the TSNs received out-of-order. All of these TSNs SHOULD be nr-gap-acked.
Set the Number of R Gap Ack Blocks (N) field to 0.
For these out-of-order TSNs, identify the NR Gap Ack Blocks.
Fill the Number of NR Gap Ack Blocks (M) field,
NR Gap Ack Block #i Start, and NR Gap Ack Block #i End where i goes from 1 to M.
RFC4960 states that the SCTP endpoint MUST report as many Gap Ack Blocks
as can fit in a single SACK chunk limited by the current path MTU.
When using NR-SACKs, the SCTP endpoint SHOULD fill as many R Gap Ack Blocks and NR Gap Ack Blocks starting
from the Cumulative TSN Ack value as can fit in a single NR-SACK chunk limited by the current path MTU.
If space remains, the SCTP endpoint SHOULD fill as many Duplicate TSNs as possible starting from
Cumulative TSN Ack value.
When an NR-SACK chunk is received, all of the NR-SACK fields identical to a
SACK chunk SHOULD be processed and handled as in SACK chunk handling outlined in
Section 6.2.1 of .
The NR Gap Ack Block Start(s) and NR Gap Ack Block End(s) are
offsets relative to the cum-ack. To calculate the actual range of nr-gap-acked TSNs,
the cum-ack MUST be added to the Start and End.
For example, assume an incoming NR-SACK chunk's cum-ack is
12 and an NR Gap Ack Block defines the NR Gap Ack Block Start=5,
and the NR Gap Ack Block End=7. This NR Gap Ack block nr-gap-acks
TSNs 17 through 19 inclusive.
Upon reception of an NR-SACK chunk, all TSNs listed in either R Gap Ack Block(s) or NR Gap Ack Block(s)
SHOULD be processed as would be TSNs included in Gap Ack Block(s) of a SACK chunk.
All TSNs in all NR Gap Ack Blocks SHOULD be removed from the
data sender's retransmission queue as their delivery to the receiving
application has either already occurred, or is guaranteed
by the data receiver. Although R Gap Ack Blocks and NR Gap Ack Blocks SHOULD be disjoint sets,
NR-SACK processing SHOULD work if an NR-SACK chunk has a TSN listed in both an R Gap Ack Block and
an NR Gap Ack Block. In this case, the TSN SHOULD be treated as Non-Renegable.
Implementation Note:
Most of NR-SACK processing at the data sender can be implemented by
using the same routines as in SACK that process the cum ack and the gap ack(s),
followed by removal of nr-gap-acked DATA chunks from the retransmission queue.
However, with NR-SACKs, as out-of-order DATA is sometimes
removed from the retransmission queue, the gap ack processing routine should recognize that
the data sender's retransmission queue has some transmitted data removed.
For example, while calculating missing reports, the
gap ack processing routine cannot assume that the highest TSN transmitted
is always at the tail (right edge) of the retransmission queue.
TBD. See , , .TBD. See , , .TBD. See , , .This algorithm ensures quick blocking resolution for ordered data.
TBD. See , .This section discusses CMT's receive buffer related problems during
path failure, and proposes a solution for the same. Link failures arise when a router or a link connecting two routers
fails due to link disconnection, hardware malfunction, or software
error. Overloaded links caused by flash crowds and denial-of-service
(DoS) attacks also degrade end-to-end communication between peer hosts.
Ideally, the routing system detects link failures, and in response,
reconfigures the routing tables and avoids routing traffic via the
failed link. However, existing research highlights problems with
Internet backbone routing that result in long route convergence times.
The pervasiveness of path failures motivated us to study their impact on
CMT, since CMT achieves better throughput via simultaneous data
transmission over multiple end-to-end paths. CMT is an extension to SCTP, and therefore
retains SCTP's failure detection process. A CMT sender uses a tunable
failure detection threshold called Path.Max.Retrans (PMR). When a sender
experiences more than PMR consecutive timeouts while trying to reach an
active destination, the destination is marked as failed. With PMR=5, the
failure detection takes 6 consecutive timeouts or 63s. After every
timeout, the CMT sender continues to transmit new data on the failed
path increasing the chances of receive buffer (rbuf) blocking and
degrading CMT performance during permanent and short-term path failures
.To mitigate the rbuf blocking, we introduce a new destination state
called 'potentially-failed' state in SCTP (and CMT's) failure detection
process .
This solution is based on the rationale that loss detected by a timeout
implies either severe congestion or failure en route.
After a single timeout on a path, a sender is unsure, and marks the
corresponding destination as 'potentially-failed' (PF).
A PF destination is not used for data transmission or retransmission.
CMT's retransmission policies are augmented to include the PF state.
Performance evaluations prove that the PF state significantly reduces
rbuf blocking during failure detection . This section discusses problems with SCTP's SACK mechanism and how
it affects the send buffer and CMT performance. Gap-acks acknowledge DATA chunks that arrive out-of-order to a
transport layer data receiver. A gap-ack in SCTP is advisory, in that,
while it notifies a data sender about the reception of indicated DATA
chunks, the data receiver is permitted to later discard DATA chunks that
it previously had gap-acked. Discarding a previously gap-acked DATA
chunk is known as 'reneging'. Because of the possibility of reneging in
SCTP, any gap-acked DATA chunk MUST NOT be removed from the data
sender's retransmission queue until the DATA chunk is later
CumAcked.Situations exist when a data receiver knows that reneging on a
particular out-of-order DATA chunk will never take place, such as (but
not limited to) after an out-of-order DATA chunk is delivered to the
receiving application. With current SACKs in SCTP, it is not possible
for a data receiver to inform a data sender if or when a particular
out-of-order 'deliverable' DATA chunk has been 'delivered' to the
receiving application. Thus the data sender MUST keep a copy of every
gap-acked out-of-order DATA chunk(s) in the data sender's retransmission
queue until the DATA chunk is CumAcked. This use of the data sender's
retransmission queue is wasteful. The wasted buffer often degrades CMT
performance; the degradation increases when a CMT flow traverses via
paths with disparate end-to-end properties .Non-Renegable Selective Acknowledgments (NR-SACKs)
are a new kind of
acknowledgements, extending SCTP's SACK chunk functionalities. The
NR-SACK chunk is an extension of the existing SACK chunk. Several fields
are identical, including the Cumulative TSN Ack, the Advertised Receiver
Window Credit (a_rwnd), and Duplicate TSNs. These fields have the same
semantics as described in .NR-SACKs also identify out-of-order DATA chunks that a receiver either:
(1) has delivered to its receiving application, or
(2) takes full responsibility to eventually deliver to its receiving application.
These out-of-order DATA chunks are 'non-renegable.'
Non-Renegable data are reported in the NR Gap Ack Block field of the
NR-SACK chunk as described .
We refer to non-renegable selective acknowledgements as 'nr-gap-acks.' When an out-of-order DATA chunk is nr-gap-acked, the data sender no
longer needs to keep that particular DATA chunk in its retransmission
queue, thus allowing the data sender to free up its buffer space sooner
than if the DATA chunk were only gap-acked. NR-SACKs improve send
buffer utilization and throughput for CMT flows
.CMT-SCTP assumes all paths to be disjoint. Since each path
independently uses a TCP-like congestion control, an SCTP association
using N paths over the same bottleneck acquires N times the bandwidth of
a concurrent TCP flow. This is clearly unfair. A reliable detection of
shared bottlenecks is impossible in arbitrary networks like the
Internet. Therefore, , apply the idea of
Resource Pooling to CMT-SCTP. Resource Pooling (RP) denotes 'making a
collection of resources behave like a single pooled resource'
. The modifications of RP-enabled CMT-SCTP, further
denoted as CMT/RP-SCTP, are described in the following subsections.
A detailed description of CMT/RP-SCTP, including congestion control
examples, can be found in , , .TDB.TDB. See , , .TDB. See , , .TDB. See , .
See and
.
A large-scale and realistic Internet testbed platform with support for the multi-homing feature of the underlying SCTP protocol is NorNet. Particularly, it is also a platform for multi-path transport experiments with CMT-SCTP. A description of and introduction to NorNet is provided in , , , . Further information can be found on the project website at https://www.nntb.no.An Open Source simulation model for CMT-SCTP is available for OMNeT++ within the INET Framework. See for the Git repository. For documentation on the model, see .NOTE to RFC-Editor:
"RFCXXXX" is to be replaced by the RFC number you assign this document.NOTE to RFC-Editor:
The suggested values for the chunk type and the chunk parameter types
are tentative and to be confirmed by IANA.This document (RFCXXXX) is the reference for all registrations
described in this section.
The suggested changes are described below.A chunk type has to be assigned by IANA.
It is suggested to use the values given in .
IANA should assign this value from the pool of chunks with the upper
two bits set to '00'.This requires an additional line in the "Chunk Types" registry for SCTP:
The registration table as defined in
for the chunk flags of this chunk type is empty.This document does not add any additional security considerations
in addition to the ones given in .The authors wish to thank
Hakim Adhari,
Phillip Conrad,
Jonathan Leighton,
Ertugrul Yilmaz and
Xing Zhou
for their invaluable comments and support.End-to-End Concurrent Multipath Transfer Using Transport Layer MultihomingConcurrent Multipath Transfer Using SCTP Multihoming Over Independent End-to-End PathsConcurrent Multipath Transfer Using Transport Layer Multihoming: Introducing the Potentially-failed Destination StateNon-Renegable Selective Acknowledgments (NR-SACKs) for SCTPThe Resource Pooling PrincipleImplementation and Evaluation of Concurrent Multipath Transfer for SCTP in the INET FrameworkApplying TCP-Friendly Congestion Control to Concurrent Multipath TransferThroughput Analysis of Non-Renegable Selective Acknowledgments (NR-SACKs) for SCTPTransmission Scheduling Optimizations for Concurrent Multipath TransferOn the Use of Concurrent Multipath Transfer over Asymmetric PathsEvaluation of Concurrent Multipath Transfer over Dissimilar PathsOn the Impact of Congestion Control for Concurrent Multipath Transfer on the Transport LayerOn the Fairness of Transport Protocols in a Multi-Path EnvironmentEvaluation and Optimisation of Multi-Path Transport using the Stream Control Transmission ProtocolNorNet -- A Real-World, Large-Scale Multi-Homing TestbedDesign and Implementation of the NorNet Core Research Testbed for Multi-Homed SystemsNorNet Core – A Multi-Homed Research TestbedINET Framework Git RepositoryThe NorNet Core Testbed – Introduction and Status in August 2014An Experiment Tutorial for the NorNet Core Testbed