RFC 1644
T/TCP -- TCP Extensions for Transactions Functional Specification
Network Working Group R. Braden Request for Comments: 1644 ISI Category: Experimental July 1994 T/TCP -- TCP Extensions for Transactions Functional Specification Status of this Memo This memo describes an Experimental Protocol for the Internet community, and requests discussion and suggestions for improvements. It does not specify an Internet Standard. Distribution is unlimited. Abstract This memo specifies T/TCP, an experimental TCP extension for efficient transaction-oriented (request/response) service. This backwards-compatible extension could fill the gap between the current connection-oriented TCP and the datagram-based UDP. This work was supported in part by the National Science Foundation under Grant Number NCR-8922231. Table of Contents 1. INTRODUCTION .................................................. 2 2. OVERVIEW ..................................................... 3 2.1 Bypassing the Three-Way Handshake ........................ 4 2.2 Transaction Sequences .................................... 6 2.3 Protocol Correctness ..................................... 8 2.4 Truncating TIME-WAIT State ............................... 12 2.5 Transition to Standard TCP Operation ..................... 14 3. FUNCTIONAL SPECIFICATION ..................................... 17 3.1 Data Structures .......................................... 17 3.2 New TCP Options .......................................... 17 3.3 Connection States ........................................ 19 3.4 T/TCP Processing Rules ................................... 25 3.5 User Interface ........................................... 28 4. IMPLEMENTATION ISSUES ........................................ 30 4.1 RFC-1323 Extensions ...................................... 30 4.2 Minimal Packet Sequence .................................. 31 4.3 RTT Measurement .......................................... 31 4.4 Cache Implementation ..................................... 32 4.5 CPU Performance .......................................... 32 4.6 Pre-SYN Queue ............................................ 33 6. ACKNOWLEDGMENTS .............................................. 34 7. REFERENCES ................................................... 34 APPENDIX A. ALGORITHM SUMMARY ................................... 35
Security Considerations .......................................... 38 Author's Address ................................................. 38 1. INTRODUCTION TCP was designed to around the virtual circuit model, to support streaming of data. Another common mode of communication is a client-server interaction, a request message followed by a response message. The request/response paradigm is used by application-layer protocols that implement transaction processing or remote procedure calls, as well as by a number of network control and management protocols (e.g., DNS and SNMP). Currently, many Internet user programs that need request/response communication use UDP, and when they require transport protocol functions such as reliable delivery they must effectively build their own private transport protocol at the application layer. Request/response, or "transaction-oriented", communication has the following features: (a) The fundamental interaction is a request followed by a response. (b) An explicit open or close phase may impose excessive overhead. (c) At-most-once semantics is required; that is, a transaction must not be "replayed" as the result of a duplicate request packet. (d) The minimum transaction latency for a client should be RTT + SPT, where RTT is the round-trip time and SPT is the server processing time. (e) In favorable circumstances, a reliable request/response handshake should be achievable with exactly one packet in each direction. This memo concerns T/TCP, an backwards-compatible extension of TCP to provide efficient transaction-oriented service in addition to virtual-circuit service. T/TCP provides all the features listed above, except for (e); the minimum exchange for T/TCP is three segments. In this memo, we use the term "transaction" for an elementary request/response packet sequence. This is not intended to imply any of the semantics often associated with application-layer transaction processing, like 3-phase commits. It is expected that T/TCP can be used as the transport layer underlying such an application-layer service, but the semantics of T/TCP is limited to transport-layer services such as reliable, ordered delivery and at-most-once
operation. An earlier memo [RFC-1379] presented the concepts involved in T/TCP. However, the real-world usefulness of these ideas depends upon practical issues like implementation complexity and performance. To help explore these issues, this memo presents a functional specification for a particular embodiment of the ideas presented in RFC-1379. However, the specific algorithms in this memo represent a later evolution than RFC-1379. In particular, Appendix A in RFC-1379 explained the difficulties in truncating TIME-WAIT state. However, experience with an implementation of the RFC-1379 algorithms in a workstation later showed that accumulation of TCB's in TIME-WAIT state is an intolerable problem; this necessity led to a simple solution for truncating TIME-WAIT state, described in this memo. Section 2 introduces the T/TCP extensions, and section 3 contains the complete specification of T/TCP. Section 4 discusses some implementation issues, and Appendix A contains an algorithmic summary. This document assumes familiarity with the standard TCP specification [STD-007]. 2. OVERVIEW The TCP protocol is highly symmetric between the two ends of a connection. This symmetry is not lost in T/TCP; for example, T/TCP supports TCP's symmetric simultaneous open from both sides (Section 2.3 below). However, transaction sequences use T/TCP in a highly unsymmetrical manner. It is convenient to use the terms "client host" and "server host" for the host that initiates a connection and the host that responds, respectively. The goal of T/TCP is to allow each transaction, i.e., each request/response sequence, to be efficiently performed as a single incarnation of a TCP connection. Standard TCP imposes two performance problems for transaction-oriented communication. First, a TCP connection is opened with a "3-way handshake", which must complete successfully before data can be transferred. The 3-way handshake adds an extra RTT (round trip time) to the latency of a transaction. The second performance problem is that closing a TCP connection leaves one or both ends in TIME-WAIT state for a time 2*MSL, where MSL is the maximum segment lifetime (defined to be 120 seconds). TIME-WAIT state severely limits the rate of successive transactions between the same (host,port) pair, since a new incarnation of the connection cannot be opened until the TIME-WAIT delay expires. RFC- 1379 explained why the alternative approach, using a different user port for each transaction between a pair of hosts, also limits the
transaction rate: (1) the 16-bit port space limits the rate to 2**16/240 transactions per second, and (2) more practically, an excessive amount of kernel space would be occupied by TCP state blocks in TIME-WAIT state [RFC-1379]. T/TCP solves these two performance problems for transactions, by (1) bypassing the 3-way handshake (3WHS) and (2) shortening the delay in TIME-WAIT state. 2.1 Bypassing the Three-Way Handshake T/TCP introduces a 32-bit incarnation number, called a "connection count" (CC), that is carried in a TCP option in each segment. A distinct CC value is assigned to each direction of an open connection. A T/TCP implementation assigns monotonically increasing CC values to successive connections that it opens actively or passively. T/TCP uses the monotonic property of CC values in initial <SYN> segments to bypass the 3WHS, using a mechanism that we call TCP Accelerated Open (TAO). Under the TAO mechanism, a host caches a small amount of state per remote host. Specifically, a T/TCP host that is acting as a server keeps a cache containing the last valid CC value that it has received from each different client host. If an initial <SYN> segment (i.e., a segment containing a SYN bit but no ACK bit) from a particular client host carries a CC value larger than the corresponding cached value, the monotonic property of CC's ensures that the <SYN> segment must be new and can therefore be accepted immediately. Otherwise, the server host does not know whether the <SYN> segment is an old duplicate or was simply delivered out of order; it therefore executes a normal 3WHS to validate the <SYN>. Thus, the TAO mechanism provides an optimization, with the normal TCP mechanism as a fallback. The CC value carried in non-<SYN> segments is used to protect against old duplicate segments from earlier incarnations of the same connection (we call such segments 'antique duplicates' for short). In the case of short connections (e.g., transactions), these CC values allow TIME-WAIT state delay to be safely discuss in Section 2.3. T/TCP defines three new TCP options, each of which carries one 32-bit CC value. These options are named CC, CC.NEW, and CC.ECHO. The CC option is normally used; CC.NEW and CC.ECHO have special functions, as follows.
(a) CC.NEW
Correctness of the TAO mechanism requires that clients
generate monotonically increasing CC values for successive
connection initiations. These values can be generated using
a simple global counter. There are certain circumstances
(discussed below in Section 2.2) when the client knows that
monotonicity may be violated; in this case, it sends a CC.NEW
rather than a CC option in the initial <SYN> segment.
Receiving a CC.NEW causes the server to invalidate its cache
entry and do a 3WHS.
(b) CC.ECHO
When a server host sends a <SYN,ACK> segment, it echoes the
connection count from the initial <SYN> in a CC.ECHO option,
which is used by the client host to validate the <SYN,ACK>
segment.
Figure 1 illustrates the TAO mechanism bypassing a 3WHS. The
cached CC values, denoted by cache.CC[host], are shown on each
side. The server host compares the new CC value x in segment #1
against x0, its cached value for client host A; this comparison is
called the "TAO test". Since x > x0, the <SYN> must be new and
can be accepted immediately; the data in the segment can therefore
be delivered to the user process B, and the cached value is
updated. If the TAO test failed (x <= x0), the server host would
do a normal three-way handshake to validate the <SYN> segment, but
the cache would not be updated.
TCP A (Client) TCP B (Server)
_______________ ______________
cache.CC[A]
V
[ x0 ]
#1 --> <SYN, data1, CC=x> --> (TAO test OK (x > x0) =>
data1->user_B and
cache.CC[A]= x; )
[ x ]
#2 <-- <SYN, ACK(data1), data2, CC=y, CC.ECHO=x> <--
(data2->user_A;)
Figure 1. TAO: Three-Way Handshake is Bypassed
The CC value x is echoed in a CC.ECHO option in the <SYN,ACK>
segment (#2); the client side uses this option to validate the
segment. Since segment #2 is valid, its data2 is delivered to the
client user process. Segment #2 also carries B's CC value; this
is used by A to validate non-SYN segments from B, as explained in
Section 2.4.
Implementing the T/TCP extensions expands the connection control
block (TCB) to include the two CC values for the connection; call
these variables TCB.CCsend and TCB.CCrecv (or CCsend, CCrecv for
short). For example, the sequence shown in Figure 1 sets
TCB.CCsend = x and TCB.CCrecv = y at host A, and vice versa at
host B. Any segment that is received with a CC option containing
a value SEG.CC different from TCB.CCsend will be rejected as an
antique duplicate.
2.2 Transaction Sequences
T/TCP applies the TAO mechanism described in the previous section
to perform a transaction sequence. Figure 2 shows a minimal
transaction, when the request and response data can each fit into
a single segment. This requires three segments and completes in
one round-trip time (RTT). If the TAO test had failed on segment
#1, B would have queued data1 and the FIN for later processing,
and then it would have returned a <SYN,ACK> segment to A, to
perform a normal 3WHS.
TCP A (Client) TCP B (Server)
_______________ ______________
CLOSED LISTEN
#1 SYN-SENT* --> <SYN,data1,FIN,CC=x> --> CLOSE-WAIT*
(TAO test OK)
(data1->user_B)
<-- LAST-ACK*
#2 TIME-WAIT <-- <SYN,ACK(FIN),data2,FIN,CC=y,CC.ECHO=x>
(data2->user_A)
#3 TIME-WAIT --> <ACK(FIN),CC=x> --> CLOSED
(timeout)
CLOSED
Figure 2: Minimal T/TCP Transaction Sequence
T/TCP extensions require additional connection states, e.g., the
SYN-SENT*, CLOSE-WAIT*, and LAST-ACK* states shown in Figure 2.
Section 3.3 describes these new connection states.
To obtain the minimal 3-segment sequence shown in Figure 2, the
server host must delay acknowledging segment #1 so the response
may be piggy-backed on segment #2. If the application takes
longer than this delay to compute the response, the normal TCP
retransmission mechanism in TCP B will send an acknowledgment to
forestall a retransmission from TCP A. Figure 3 shows an example
of a slow server application. Although the sequence in Figure 3
does contain a 3-way handshake, the TAO mechanism has allowed the
request data to be accepted immediately, so that the client still
sees the minimum latency.
TCP A (Client) TCP B (Server)
_______________ ______________
CLOSED LISTEN
#1 SYN-SENT* --> <SYN,data1,FIN,CC=x> --> CLOSE-WAIT*
(TAO test OK =>
data1->user_B)
(timeout)
#2 FIN-WAIT-1 <-- <SYN,ACK(FIN),CC=y,CC.ECHO=x> <-- CLOSE-WAIT*
#3 FIN-WAIT-1 --> <ACK(SYN),FIN,CC=x> --> CLOSE-WAIT
#4 TIME-WAIT <-- <ACK(FIN),data2,FIN,CC=y> <-- LAST-ACK
(data2->user_A)
#5 TIME_WAIT --> <ACK(FIN),CC=x> --> CLOSED
(timeout)
CLOSED
Figure 3: Acknowledgment Timeout in Server
2.3 Protocol Correctness
This section fills in more details of the TAO mechanism and
provides an informal sketch of why the T/TCP protocol works.
CC values are 32-bit integers. The TAO test requires the same
kind of modular arithmetic that is used to compare two TCP
sequence numbers. We assume that the boundary between y < z and z
< y for two CC values y and z occurs when they differ by 2**31,
i.e., by half the total CC space.
The essential requirement for correctness of T/TCP is this:
CC values must advance at a rate slower than 2**31 [R1]
counts per 2*MSL
where MSL denotes the maximum segment lifetime in the Internet.
The requirement [R1] is easily met with a 32-bit CC. For example,
it will allow 10**6 transactions per second with the very liberal
MSL of 1000 seconds [RFC-1379]. This is well in excess of the
transaction rates achievable with current operating systems and
network latency.
Assume for the present that successive connections from client A
to server B contain only monotonically increasing CC values. That
is, if x(i) and x(i+1) are CC values carried in two successive
initial <SYN> segments from the same host, then x(i+1) > x(i).
Assuming the requirement [R1], the CC space cannot wrap within the
range of segments that can be outstanding at one time. Therefore,
those successive <SYN> segments from a given host that have not
exceeded their MSL must contain an ordered set of CC values:
x(1) < x(2) < x(3) ... < x(n),
where the modular comparisons have been replaced by simple
arithmetic comparisons. Here x(n) is the most recent acceptable
<SYN>, which is cached by the server. If the server host receives
a <SYN> segment containing a CC option with value y where y >
x(n), that <SYN> must be newer; an antique duplicate SYN with CC
value greater than x(n) must have exceeded its MSL and vanished.
Hence, monotonic CC values and the TAO test prevent erroneous
replay of antique <SYN>s.
There are two possible reasons for a client to generate non-
monotonic CC values: (a) the client may have crashed and
restarted, causing the generated CC values to jump backwards; or
(b) the generated CC values may have wrapped around the finite
space. Wraparound may occur because CC generation is global to
all connections. Suppose that host A sends a transaction to B,
then sends more than 2**31 transactions to other hosts, and
finally sends another transaction to B. From B's viewpoint, CC
will have jumped backward relative to its cached value.
In either of these two cases, the server may see the CC value jump
backwards only after an interval of at least MSL since the last
<SYN> segment from the same client host. In case (a), client host
restart, this is because T/TCP retains TCP's explicit "Quiet Time"
of an MSL interval [STD-007]. In case (b). wrap around, [R1]
ensures that a time of at least MSL must have passed before the CC
space wraps around. Hence, there is no possibility that a TAO
test will succeed erroneously due to either cause of non-
monotonicity; i.e., there is no chance of replays due to TAO.
However, although CC values jumping backwards will not cause an
error, it may cause a performance degradation due to unnecessary
3WHS's. This results from the generated CC values jumping
backwards through approximately half their range, so that all
succeeding TAO tests fail until the generated CC values catch up
to the cached value. To avoid this degradation, a client host
sends a CC.NEW option instead of a CC option in the case of either
system restart or CC wraparound. Receiving CC.NEW forces a 3WHS,
but when this 3WHS completes successfully the server cache is
updated to the new CC value. To detect CC wraparound, the client
must cache the last CC value it sent to each server. It therefore
maintains cache.CCsent[B] for each server B. If this cached value
is undefined or if it is larger than the next CC value generated
at the client, then the client sends a CC.NEW instead of a CC
option in the next SYN segment.
This is illustrated in Figure 4, which shows the scenario for the
first transaction from A to B after the client host A has crashed
and recovered. A similar sequence occurs if x is not greater than
cache.CCsent[B], i.e., if there is a wraparound of the generated
CC values. Because segment #1 contains a CC.NEW option, the
server host invalidates the cache entry and does a 3WHS; however,
it still sets B's TCB.CCrecv for this connection to x. TCP B uses
this CCrecv value to validate the <ACK> segment (#3) that
completes the 3WHS. Receipt of this segment updates cache.CC[A],
since the cache entry was previously undefined. (If a 3WHS always
updated the cache, then out-of-order SYN segments could cause the
cached value to jump backwards, possibly allowing replays).
Finally, the CC.ECHO option in the <SYN,ACK> segment #2 defines
A's cache.CCsent entry.
This algorithm delays updating cache.CCsent[] until the <SYN> has
been ACK'd. This allows the undefined cache.CCsent value to used
as a a "first-time switch" to reliable resynchronization of the
cached value at the server after a crash or wraparound.
When we use the term "cache", we imply that the value can be
discarded at any time without introducing erroneous behavior
although it may degrade performance.
(a) If a server host receives an initial <SYN> from client A but
has no cached value cache.CC[A], the server simply forces a
3WHS to validate the <SYN> segment.
(b) If a client host has no cached value cache.CCsent[B] when it
needs to send an initial <SYN> segment, the client simply
sends a CC.NEW option in the segment. This forces a 3WHS at
the server.
TCP A (Client) TCP B (Server)
_______________ ______________
cache.CCsent[B] cache.CC[A]
V V
(Crash and restart)
[ ?? ] [ x0 ]
#1 --> <SYN, data1,CC.NEW=x> --> (invalidate cache;
queue data1;
3-way handshake)
[ ?? ] [ ?? ]
#2 <-- <SYN, ACK(data1),CC=y,CC.ECHO=x> <--
(cache.CCsent[B]= x;)
[ x ] [ ?? ]
#3 --> <ACK(SYN),CC=x> --> data1->user_B;
cache.CC[A]= x;
[ x ] [ x ]
Figure 4. Client Host Restarting
So far, we have considered only correctness of the TAO mechanism
for bypassing the 3WHS. We must also protect a connection against
antique duplicate non-SYN segments. In standard TCP, such
protection is one of the functions of the TIME-WAIT state delay.
(The other function is the TCP full-duplex close semantics, which
we need to preserve; that is discussed below in Section 2.5). In
order to achieve a high rate of transaction processing, it must be
possible to truncate this TIME-WAIT state delay without exposure
to antique duplicate segments [RFC-1379].
For short connections (e.g., transactions), the CC values assigned
to each direction of the connection can be used to protect against
antique duplicate non-SYN segments. Here we define "short" as a
duration less than MSL. Suppose that there is a connection that
uses the CC values TCB.CCsend = x and TCB.CCrecv = y. By the
requirement [R1], neither x nor y can be reused for a new
connection from the same remote host for a time at least 2*MSL.
If the connection has been in existence for a time less than MSL,
then its CC values will not be reused for a period that exceeds
MSL, and therefore all antique duplicates with that CC value must
vanish before it is reused. Thus, for "short" connections we can
guard against antique non-SYN segments by simply checking the CC
value in the segment againsts TCB.CCrecv. Note that this check
does not use the monotonic property of the CC values, only that
they not cycle in less than 2*MSL. Again, the quiet time at
system restart protects against errors due to crash with loss of
state.
If the connection duration exceeds MSL, safety from old duplicates
still requires a TIME-WAIT delay of 2*MSL. Thus, truncation of
TIME-WAIT state is only possible for short connections. (This
problem has also been noticed by Shankar and Lee [ShankarLee93]).
This difference in behavior for long and for short connections
does create a slightly complex service model for applications
using T/TCP. An application has two different strategies for
multiple connections. For "short" connections, it should use a
fixed port pair and use the T/TCP mechanism to get rapid and
efficient transaction processing. For connections whose durations
are of the order of MSL or longer, it should use a different user
port for each successive connection, as is the current practice
with unmodified TCP. The latter strategy will cause excessive
overhead (due to TCB's in TIME-WAIT state) if it is applied to
high-frequency short connections. If an application makes the
wrong choice, its attempt to open a new connection may fail with a
"busy" error. If connection durations may range between long and
short, an application may have to be able to switch strategies
when one fails.
2.4 Truncating TIME-WAIT State
Truncation of TIME-WAIT state is necessary to achieve high
transaction rates. As Figure 2 illustrates, a standard
transaction leaves the client end of the connection in TIME-WAIT
state. This section explains the protocol implications of
truncating TIME-WAIT state, when it is allowed (i.e., when the
connection has been in existence for less than MSL). In this
case, the client host should be able to interrupt TIME-WAIT state
to initiate a new incarnation of the same connection (i.e., using
the same host and ports). This will send an initial <SYN>
segment.
It is possible for the new <SYN> to arrive at the server before
the retransmission state from the previous incarnation is gone, as
shown in Figure 5. Here the final <ACK> (segment #3) from the
previous incarnation is lost, leaving retransmission state at B.
However, the client received segment #2 and thinks the transaction
completed successfully, so it can initiate a new transaction by
sending <SYN> segment #4. When this <SYN> arrives at the server
host, it must implicitly acknowledge segment #2, signalling
success to the server application, deleting the old TCB, and
creating a new TCB, as shown in Figure 5. Still assuming that the
new <SYN> is known to be valid, the server host marks the new
connection half-synchronized and delivers data3 to the server
application. (The details of how this is accomplished are
presented in Section 3.3.)
The earlier discussion of the TAO mechanism assumed that the
previous incarnation was closed before a new <SYN> arrived at the
server. However, TAO cannot be used to validate the <SYN> if
there is still state from the previous incarnation, as shown in
Figure 5; in this case, it would be exceedingly awkward to perform
a 3WHS if the TAO test should fail. Fortunately, a modified
version of the TAO test can still be performed, using the state in
the earlier TCB rather than the cached state.
(A) If the <SYN> segment contains a CC or CC.NEW option, the
value SEG.CC from this option is compared with TCB.CCrecv,
the CC value in the still-existing state block of the
previous incarnation. If SEG.CC > TCB.CCrecv, the new <SYN>
segment must be valid.
(B) Otherwise, the <SYN> is an old duplicate and is simply
discarded.
Truncating TIME-WAIT state may be looked upon as composing an
extended state machine that joins the state machines of the two
incarnations, old and new. It may be described by introducing new
intermediate states (which we call I-states), with transitions
that join the two diagrams and share some state from each. I-
states are detailed in Section 3.3.
Notice also segment #2' in Figure 5. TCP's mechanism to recover
from half-open connections (see Figure 10 of [STD-007]) cause TCP
A to send a RST when 2' arrives, which would incorrectly make B
think that the previous transaction did not complete successfully.
The half-open recovery mechanism must be defeated in this case, by
A ignoring segment #2'.
TCP A (Client) TCP B (Server)
_______________ ______________
CLOSED LISTEN
#1 --> <...,FIN,CC=x> --> LAST-ACK*
#2 <-- <...ACK(FIN),data2,FIN,CC=y,CC.ECHO=x> <--- LAST-ACK*
TIME-WAIT
(data2->user_A)
#3 TIME-WAIT --> <ACK(FIN),CC=x> --> X (DROP)
(New Active Open) (New Passive Open)
#4 SYN-SENT* --> <SYN, data3,CC=z> ...
LISTEN-LA
#2' (discard) <-- <...ACK(FIN),data2,FIN,CC=y> <--- (retransmit)
#4 SYN-SENT* ... <SYN,data3,CC=z> --> ESTABLISHED*
SYN OK (see text) =>
{Ack seg #2;
Delete old TCB;
Create new TCB;
data3 -> user_B;
cache.CC[A]= z;}
Figure 5: Truncating TIME-WAIT State: SYN as Implicit ACK
2.5 Transition to Standard TCP Operation
T/TCP includes all normal TCP semantics, and it will continue to
operate exactly like TCP when the particular assumptions for
transactions do not hold. There is no limit on the size of an
individual transaction, and behavior of T/TCP should merge
seamlessly from pure transaction operation as shown in Figure 2,
to pure streaming mode for sending large files. All the sequences
shown in [STD-007] are still valid, and the inherent symmetry of
TCP is preserved.
Figure 6 shows a possible sequence when the request and response
messages each require two segments. Segment #2 is a non-SYN
segment that contains a TCP option. To avoid compatibility
problems with existing TCP implementations, the client side should
send segment #2 only if cache.CCsent[B] is defined, i.e., only if
host A knows that host B plays the new game.
TCP A (Client) TCP B (Server)
_______________ ______________
CLOSED LISTEN
#1 SYN-SENT* --> <SYN,data1,CC=x> --> ESTABLISHED*
(TAO test OK =>
data1-> user)
#2 SYN-SENT* --> <data2,FIN,CC=x> --> CLOSE-WAIT*
(data2-> user)
CLOSE-WAIT*
#3 FIN-WAIT-2 <-- <SYN,ACK(FIN),data3,CC=y,CC.ECHO=x> <--
(data3->user)
#4 TIME_WAIT <-- <ACK(FIN),data4,FIN,CC=y> <-- LAST-ACK*
(data4->user)
#5 TIME-WAIT --> <ACK(FIN),CC=x> --> CLOSED
Figure 6. Multi-Packet Request/Response Sequence
Figure 7 shows a more complex example, one possible sequence with
TAO combined with simultaneous open and close. This may be
compared with Figure 8 of [STD-007].
TCP A TCP B
_______________ ______________
CLOSED CLOSED
#1 SYN-SENT* --> <SYN,data1,FIN,CC=x> ...
#2 CLOSING* <-- <SYN,data2,FIN,CC=y> <-- SYN-SENT*
(TAO test OK =>
data2->user_A
#3 CLOSING* --> <FIN,ACK(FIN),CC=x,CC.ECHO=y> ...
#1' ... <SYN,data1,FIN,CC=x> --> CLOSING*
(TAO test OK =>
data1->user_B)
#4 TIME-WAIT <-- <FIN,ACK(FIN),CC=y,CC.ECHO=x> <-- CLOSING*
#5 TIME-WAIT --> <ACK(FIN),CC=x> ...
#3' ... <FIN,ACK(FIN),CC=x,CC.ECHO=y> --> TIME-WAIT
#6 TIME-WAIT <-- <ACK(FIN),CC=y> <--- TIME-WAIT
#5' TIME-WAIT ... <ACK(FIN),CC=x> --> TIME-WAIT
(timeout) (timeout)
CLOSED CLOSED
Figure 7: Simultaneous Open and Close
3. FUNCTIONAL SPECIFICATION 3.1 Data Structures A connection count is an unsigned 32-bit integer, with the value zero excluded. Zero is used to denote an undefined value. A host maintains a global connection count variable CCgen, and each connection control block (TCB) contains two new connection count variables, TCB.CCsend and TCB.CCrecv. Whenever a TCB is created for the active or passive end of a new connection, CCgen is incremented by 1 and placed in TCB.CCsend of the TCB; however, if the previous CCgen value was 0xffffffff (-1), then the next value should be 1. TCB.CCrecv is initialized to zero (undefined). T/TCP adds a per-host cache to TCP. An entry in this cache for foreign host fh includes two CC values, cache.CC[fh] and cache.CCsent[fh]. It may include other values, as discussed in Sections 4.3 and 4.4. According to [STD-007], a TCP is not permitted to send a segment larger than the default size 536, unless it has received a larger value in an MSS (Maximum Segment Size) option. This could constrain the client to use the default MSS of 536 bytes for every request. To avoid this constraint, a T/TCP may cache the MSS option values received from remote hosts, and we allow a TCP to use a cached MSS option value for the initial SYN segment. When the client sends an initial <SYN> segment containing data, it does not have a send window for the server host. This is not a great difficulty; we simply define a default initial window; our current suggestion is 4K. Such a non-zero default should be be conditioned upon the existence of a cached connection count for the foreign host, so that data may be included on an initial SYN segment only if cache.CC[foreign host] is non-zero. In TCP, the window is dynamically adjusted to provide congestion control/avoidance [Jacobson88]. It is possible that a particular path might not be able to absorb an initial burst of 4096 bytes without congestive losses. If this turns out to be a problem, it should be possible to cache the congestion threshold for the path and use this value to determine the maximum size of the initial packet burst created by a request. 3.2 New TCP Options Three new TCP options are defined: CC, CC.NEW, and CC.ECHO. Each carries a connection count SEG.CC. The complete rules for sending and processing these options are given in Section 3.4 below.
CC Option
Kind: 11
Length: 6
+--------+--------+--------+--------+--------+--------+
|00001011|00000110| Connection Count: SEG.CC |
+--------+--------+--------+--------+--------+--------+
Kind=11 Length=6
This option may be sent in an initial SYN segment, and it may
be sent in other segments if a CC or CC.NEW option has been
received for this incarnation of the connection. Its SEG.CC
value is the TCB.CCsend value from the sender's TCB.
CC.NEW Option
Kind: 12
Length: 6
+--------+--------+--------+--------+--------+--------+
|00001100|00000110| Connection Count: SEG.CC |
+--------+--------+--------+--------+--------+--------+
Kind=12 Length=6
This option may be sent instead of a CC option in an initial
<SYN> segment (i.e., SYN but not ACK bit), to indicate that the
SEG.CC value may not be larger than the previous value. Its
SEG.CC value is the TCB.CCsend value from the sender's TCB.
CC.ECHO Option
Kind: 13
Length: 6
+--------+--------+--------+--------+--------+--------+
|00001101|00000110| Connection Count: SEG.CC |
+--------+--------+--------+--------+--------+--------+
Kind=13 Length=6
This option must be sent (in addition to a CC option) in a
segment containing both a SYN and an ACK bit, if the initial
SYN segment contained a CC or CC.NEW option. Its SEG.CC value
is the SEG.CC value from the initial SYN.
A CC.ECHO option should be sent only in a <SYN,ACK> segment and
should be ignored if it is received in any other segment.
3.3 Connection States
T/TCP requires new connection states and state transitions.
Figure 8 shows the resulting finite state machine; see [RFC-1379]
for a detailed development. If all state names ending in stars
are removed from Figure 8, the state diagram reduces to the
standard TCP state machine (see Figure 6 of [STD-007]), with two
exceptions:
* STD-007 shows a direct transition from SYN-RECEIVED to FIN-
WAIT-1 state when the user issues a CLOSE call. This
transition is suspect; a more accurate description of the
state machine would seem to require the intermediate SYN-
RECEIVED* state shown in Figure 8.
* In STD-007, a user CLOSE call in SYN-SENT state causes a
direct transition to CLOSED state. The extended diagram of
Figure 8 forces the connection to open before it closes,
since calling CLOSE to terminate the request in SYN-SENT
state is normal behavior for a transaction client. In the
case that no data has been sent in SYN-SENT state, it is
reasonable for a user CLOSE call to immediately enter CLOSED
state and delete the TCB.
Each of the new states in Figure 8 bears a starred name, created
by suffixing a star onto a standard TCP state. Each "starred"
state bears a simple relationship to the corresponding "unstarred"
state.
o SYN-SENT* and SYN-RECEIVED* differ from the SYN-SENT and
SYN-RECEIVED state, respectively, in recording the fact that
a FIN needs to be sent.
o The other starred states indicate that the connection is
half-synchronized (hence, a SYN bit needs to be sent).
________ g ________
| |<------------| |
| CLOSED |------------>| LISTEN |
|________| h ------|________|
| / | |
| / i| j|
| / | |
a| a'/ | _V______ ________
| / j | |ESTAB- | e' | CLOSE- |
| / -----------|-->| LISHED*|------------>| WAIT*|
| / / | |________| |________|
| / / | | | | |
| / / | | c| d'| c|
____V_V_ / _______V | __V_____ | __V_____
| SYN- | b' | SYN- |c | |ESTAB- | e | | CLOSE- |
| SENT |------>|RECEIVED|---|->| LISHED|----------|->| WAIT |
|________| |________| | |________| | |________|
| | | | | |
| | | | __V_____ |
| | | | | LAST- | |
d'| d'| d'| d| | ACK* | |
| | | | |________| |
| | | | | |
| | ______V_ | ________ |c' |d
| k | | FIN- | | e''' | | | |
| -------|-->| WAIT-1*|---|------>|CLOSING*| | |
| / | |________| | |________| | |
| / | | | | | |
| / | c'| | c'| | |
___V___ / ____V___ V_____V_ ____V___ V____V__
| SYN- | b'' | SYN- | c | FIN- | e'' | | | LAST- |
| SENT* |---->|RECEIVD*|---->| WAIT-1 |---->|CLOSING | | ACK |
|________| |________| |________| |________| |________|
| | |
f| f| f'|
___V____ ____V___ ___V____
| FIN- | e |TIME- | T | |
| WAIT-2 |---->| WAIT |-->| CLOSED |
|________| |________| |________|
Figure 8A: Basic T/TCP State Diagram
________________________________________________________________
| |
| Label Event / Action |
| _____ ________________________ |
| |
| a Active OPEN / create TCB, snd SYN |
| a' Active OPEN / snd SYN |
| b rcv SYN [no TAO]/ snd ACK(SYN) |
| b' rcv SYN [no TAO]/ snd SYN,ACK(SYN) |
| b'' rcv SYN [no TAO]/ snd SYN,FIN,ACK(SYN) |
| c rcv ACK(SYN) / |
| c' rcv ACK(SYN) / snd FIN |
| d CLOSE / snd FIN |
| d' CLOSE / snd SYN,FIN |
| e rcv FIN / snd ACK(FIN) |
| e' rcv FIN / snd SYN,ACK(FIN) |
| e'' rcv FIN / snd FIN,ACK(FIN) |
| e''' rcv FIN / snd SYN,FIN,ACK(FIN) |
| f rcv ACK(FIN) / |
| f' rcv ACK(FIN) / delete TCB |
| g CLOSE / delete TCB |
| h passive OPEN / create TCB |
| i (= b') rcv SYN [no TAO]/ snd SYN,ACK(SYN) |
| j rcv SYN [TAO OK] / snd SYN,ACK(SYN) |
| k rcv SYN [TAO OK] / snd SYN,FIN,ACK(SYN) |
| T timeout=2MSL / delete TCB |
| |
| |
| Figure 8B. Definition of State Transitions |
|________________________________________________________________|
This simple correspondence leads to an alternative state model,
which makes it easy to incorporate the new states in an existing
implementation. Each state in the extended FSM is defined by the
triplet:
(old_state, SENDSYN, SENDFIN)
where 'old_state' is a standard TCP state and SENDFIN and SENDSYN
are Boolean flags see Figure 9. The SENDFIN flag is turned on (on
the client side) by a SEND(... EOF=YES) call, to indicate that a
FIN should be sent in a state which would not otherwise send a
FIN. The SENDSYN flag is turned on when the TAO test succeeds to
indicate that the connection is only half synchronized; as a
result, a SYN will be sent in a state which would not otherwise
send a SYN.
________________________________________________________________
| |
| New state: Old_state: SENDSYN: SENDFIN: |
| __________ __________ ______ ______ |
| |
| SYN-SENT* => SYN-SENT FALSE TRUE |
| |
| SYN-RECEIVED* => SYN-RECEIVED FALSE TRUE |
| |
| ESTABLISHED* => ESTABLISHED TRUE FALSE |
| |
| CLOSE-WAIT* => CLOSE-WAIT TRUE FALSE |
| |
| LAST-ACK* => LAST-ACK TRUE FALSE |
| |
| FIN-WAIT-1* => FIN-WAIT-1 TRUE FALSE |
| |
| CLOSING* => CLOSING TRUE FALSE |
| |
| |
| Figure 9: Alternative State Definitions |
|________________________________________________________________|
Here is a more complete description of these boolean variables.
* SENDFIN
SENDFIN is turned on by the SEND(...EOF=YES) call, and turned
off when FIN-WAIT-1 state is entered. It may only be on in
SYN-SENT* and SYN-RECEIVED* states.
SENDFIN has two effects. First, it causes a FIN to be sent
on the last segment of data from the user. Second, it causes
the SYN-SENT[*] and SYN-RECEIVED[*] states to transition
directly to FIN-WAIT-1, skipping ESTABLISHED state.
* SENDSYN
The SENDSYN flag is turned on when an initial SYN segment is
received and passes the TAO test. SENDSYN is turned off when
the SYN is acknowledged (specifically, when there is no RST
or SYN bit and SEG.UNA < SND.ACK).
SENDSYN has three effects. First, it causes the SYN bit to
be set in segments sent with the initial sequence number
(ISN). Second, it causes a transition directly from LISTEN
state to ESTABLISHED*, if there is no FIN bit, or otherwise
to CLOSE-WAIT*. Finally, it allows data to be received and
processed (passed to the application) even if the segment
does not contain an ACK bit.
According to the state model of the basic TCP specification [STD-
007], the server side must explicitly issued a passive OPEN call,
creating a TCB in LISTEN state, before an initial SYN may be
accepted. To accommodate truncation of TIME-WAIT state within
this model, it is necessary to add the five "I-states" shown in
Figure 10. The I-states are: LISTEN-LA, LISTEN-LA*, LISTEN-CL,
LISTEN-CL*, and LISTEN-TW. These are 'bridge states' between two
successive the state diagrams of two successive incarnations.
Here D is the duration of the previous connection, i.e., the
elapsed time since the connection opened. The transitions labeled
with lower-case letters are taken from Figure 8.
Fortunately, many TCP implementations have a different user
interface model, in which the use can issue a generic passive open
("listen") call; thereafter, when a matching initial SYN arrives,
a new TCB in LISTEN state is automatically generated. With this
user model, the I-states of Figure 10 are unnecessary.
For example, suppose an initial SYN segment arrives for a
connection that is in LAST-ACK state. If this segment carries a
CC option and if SEG.CC is greater than TCB.CCrecv in the existing
TCB, the "q" transition shown in Figure 10 can be made directly
from the LAST-ACK state. That is, the previous TCB is processed
as if an ACK(FIN) had arrived, causing the user to be notified of
a successful CLOSE and the TCB to be deleted. Then processing of
the new SYN segment is repeated, using a new TCB that is generated
automatically. The same principle can be used to avoid
implementing any of the I-states.
______________________________
| P: Passive OPEN / |
| |
| Q: Rcv SYN, special TAO test | d'| d|
| (see text) / Delete TCB, | ________ ___V____ |
| create TCB, snd SYN | |LISTEN- | P | LAST- | |
| | | LA* |<-----| ACK* | |
| Q': (same as Q) if D < MSL | |________| |________| |
| | | | | |
| R: Rcv ACK(FIN) / Delete TCB,| Q| c'| c'| |
| create TCB | | | | |
| | | ___V____ V______V
| S': Active OPEN if D < MSL / | | |LISTEN- | P | LAST- |
| Delete TCB, create TCB, | | | LA |<-----| ACK |
| snd SYN. | | |________| |________|
|______________________________| | | | |
| Q| R| f|
________ ________ | | | |
e''' | | P |LISTEN- | | | V V
---->|CLOSING*|----->| CL* | | | LISTEN CLOSED
|________| |________| | |
| | Q| | |
c'| c'| V V V
| | ESTABLISHED*
____V___ V_______
e'' | | P |LISTEN- |
---->|CLOSING |------>| CL |
|________| |________|
| R| Q|
f| V V
| LISTEN ESTABLISHED*
____V___ _________
e |TIME- | P | LISTEN- |
---->| WAIT |------------->| TW |
|________| |_________|
/ | | | |
S'/ T| T| Q'| |S'
| _____V_ h _____V__ | V
| | |-------->| | | SYN-SENT
| | CLOSED |<--------| LISTEN | |
| |________| ------|________| |
| | / | j| |
| a| a'/ i| V V
| | / | ESTABLISHED*
V V V V
SYN-SENT ...
Figure 10: I-States for TIME-WAIT Truncation
3.4 T/TCP Processing Rules This section summarizes the rules for sending and processing the T/TCP options. INITIALIZATION I1: All cache entries cache.CC[*] and cache.CCsent[*] are undefined (zero) when a host system initializes, and CCgen is set to a non-zero value. I2: A new TCB is initialized with TCB.CCrecv = 0 and TCB.CCsend = current CCgen value; CCgen is then incremented. If the result is zero, CCgen is incremented again. SENDING SEGMENTS S1: Sending initial <SYN> Segment An initial <SYN> segment is sent with either a CC option or a CC.NEW option. If cache.CCsent[fh] is undefined or if TCB.CCsend < cache.CCsent[fh], then the option CC.NEW(TCB.CCsend) is sent and cache.CCsent[fh] is set to zero. Otherwise, the option CC(TCB.CCsend) is sent and cache.CCsent[fh] is set to CCsend. S2: Sending <SYN,ACK> Segment If the sender's TCB.CCrecv is non-zero, then a <SYN,ACK> segment is sent with both a CC(TCB.CCsend) option and a CC.ECHO (TCB.CCrecv) option. S3: Sending Non-SYN Segment A non-SYN segment is sent with a CC(TCB.CCsend) option if the TCB.CCrecv value is non-zero, or if the state is SYN- SENT or SYN-SENT* and cache.CCsent[fh] is non-zero (this last is required to send CC options in the segments following the first of a multi-segment request message; see segment #2 in Figure 6). RECEIVING INITIAL <SYN> SEGMENT Suppose that a server host receives a segment containing a SYN bit but no ACK bit in LISTEN, SYN-SENT, or SYN-SENT* state.
R1.1:If the <SYN> segment contains a CC or CC.NEW option,
SEG.CC is stored into TCB.CCrecv of the new TCB.
R1.2:If the segment contains a CC option and if the local cache
entry cache.CC[fh] is defined and if
SEG.CC > cache.CC[fh], then the TAO test is passed and the
connection is half-synchronized in the incoming direction.
The server host replaces the cache.CC[fh] value by SEG.CC,
passes any data in the segment to the user, and processes
a FIN bit if present.
Acknowledgment of the SYN is delayed to allow piggybacking
on a response segment.
R1.3:If SEG.CC <= cache.CC[fh] (the TAO test has failed), or if
cache.CC[fh] is undefined, or if there is no CC option
(but possibly a CC.NEW option), the server host proceeds
with normal TCP processing. If the connection was in
LISTEN state, then the host executes a 3-way handshake
using the standard TCP rules. In the SYN-SENT or SYN-
SENT* state (i.e., the simultaneous open case), the TCP
sends ACK(SYN) and enters SYN-RECEIVED state.
R1.4:If there is no CC option (but possibly a CC.NEW option),
then the server host sets cache.CC[fh] undefined (zero).
Receiving an ACK for a SYN (following application of rule
R1.3) will update cache.CC[fh], by rule R3.
Suppose that an initial <SYN> segment containing a CC or CC.NEW
option arrives in an I-state (i.e., a state with a name of the
form 'LISTEN-xx', where xx is one of TW, LA, L8, CL, or CL*):
R1.5:If the state is LISTEN-TW, then the duration of the
current connection is compared with MSL. If duration >
MSL then send a RST:
<SEQ=0><ACK=SEG.SEQ+SEG.LEN><CTL=RST,ACK>
drop the packet, and return.
R1.6:Perform a special TAO test: compare SEG.CC with
TCB.CCrecv.
If SEG.CC is greater, then processing is performed as if
an ACK(FIN) had arrived: signal the application that the
previous close completed successfully and delete the
previous TCB. Then create a new TCB in LISTEN state and
reprocess the SYN segment against the new TCB.
Otherwise, silently discard the segment.
RECEIVING <SYN,ACK> SEGMENT
Suppose that a client host receives a <SYN,ACK> segment for a
connection in SYN-SENT or SYN-SENT* state.
R2.1:If SEG.ACK is not acceptable (see [STD-007]) and
cache.CCsent[fh] is non-zero, then simply drop the segment
without sending a RST. (The new SYN that the client is
(re-)transmitting will eventually acknowledge any
outstanding data and FIN at the server.)
R2.2:If the segment contains a CC.ECHO option whose SEG.CC is
different from TCB.CCsend, then the segment is
unacceptable and is dropped.
R2.3:If cache.CCsent[fh] is zero, then it is set to TCB.CCsend.
R2.4:If the segment contains a CC option, its SEG.CC is stored
into TCB.CCrecv of the TCB.
RECEIVING <ACK> SEGMENT IN SYN-RECEIVED STATE
R3.1:If a segment contains a CC option whose SEG.CC differs
from TCB.CCrecv, then the segment is unacceptable and is
dropped.
R3.2:Otherwise, a 3-way handshake has completed successfully at
the server side. If the segment contains a CC option and
if cache.CC[fh] is zero, then cache.CC[fh] is replaced by
TCB.CCrecv.
RECEIVING OTHER SEGMENT
R4: Any other segment received with a CC option is
unacceptable if SEG.CC differs from TCB.CCrecv. However,
a RST segment is exempted from this test.
OPEN REQUEST
To allow truncation of TIME-WAIT state, the following changes
are made in the state diagram for OPEN requests (see Figure
10):
O1.1:A new passive open request is allowed in any of the
states: LAST-ACK, LAST-ACK*, CLOSING, CLOSING*, or TIME-
WAIT. This causes a transition to the corresponding I-
state (see Figure 10), which retains the previous state,
including the retransmission queue and timer.
O1.2 A new active open request is allowed in TIME-WAIT or
LISTEN-TW state, if the elapsed time since the current
connection opened is less than MSL. The result is to
delete the old TCB and create a new one, send a new SYN
segment, and enter SYN-SENT or SYN-SENT* state (depending
upon whether or not the SYN segment contains a FIN bit).
Finally, T/TCP has a provision to improve performance for the case
of a client that "sprays" transactions rapidly using many
different server hosts and/or ports. If TCB.CCrecv in the TCB is
non-zero (and still assuming that the connection duration is less
than MSL), then the TIME-WAIT delay may be set to min(K*RTO,
2*MSL). Here RTO is the measured retransmission timeout time and
the constant K is currently specified to be 8.
3.5 User Interface
STD-007 defines a prototype user interface ("transport service")
that implements the virtual circuit service model [STD-007,
Section 3.8]. One addition to this interface in required for
transaction processing: a new Boolean flag "end-of-file" (EOF),
added to the SEND call. A generic SEND call becomes:
Send
Format: SEND (local connection name, buffer address,
byte count, PUSH flag, URGENT flag, EOF flag [,timeout])
The following text would be added to the description of SEND in
[STD-007]:
If the EOF (End-Of-File) flag is set, any remaining queued
data is pushed and the connection is closed. Just as with the
CLOSE call, all data being sent is delivered reliably before
the close takes effect, and data may continue to be received
on the connection after completion of the SEND call.
Figure 8A shows a skeleton sequence of user calls by which a
client could initiate a transaction. The SEND call initiates a
transaction request to the foreign socket (host and port)
specified in the passive OPEN call. The predicate "recv_EOF"
tests whether or not a FIN has been received on the connection;
this might be implemented using the STATUS command of [STD-007],
or it might be implemented by some operating-system-dependent
mechanism. When recv_EOF returns TRUE, the connection has been
completely closed and the client end of the connection is in
TIME-WAIT state.
__________________________________________________________________
| |
| |
| OPEN(local_port, foreign_socket, PASSIVE) -> conn_name; |
| |
| SEND(conn_name, request_buffer, length, |
| PUSH=YES, URG=NO, EOF=YES); |
| |
| while (not recv_EOF(conn_name)) { |
| |
| RECEIVE(conn_name, reply_buffer, length) -> count; |
| |
| <Process reply_buffer.> |
| } |
| |
| |
| Figure 8A: Client Side User Interface |
|__________________________________________________________________|
If a client is going to send a rapid series of such requests to
the same foreign_socket, it should use the same local_port for
all. This will allow truncation of TIME-WAIT state. Otherwise,
it could leave local_port wild, allowing TCP to choose successive
local ports for each call, realizing that each transaction may
leave behind a significant control block overhead in the kernel.
Figure 8B shows a basic sequence of server calls. The server
application waits for a request to arrive and then reads and
processes it until a FIN arrives (recv_EOF returns TRUE). At this
time, the connection is half-closed. The SEND call used to return
the reply completes the close in the other direction. It should
be noted that the use of SEND(... EOF=YES) in Figure 4B instead of
a SEND, CLOSE sequence is only an optimization; it allows
piggybacking the FIN in order to minimize the number of segments.
It should have little effect on transaction latency.
__________________________________________________________________
| |
| |
| OPEN(local_port, ANY_SOCKET, PASSIVE) -> conn_name; |
| |
| <Wait for connection to open.> |
| |
| STATUS(conn_name) -> foreign_socket |
| |
| while (not recv_EOF(conn_name)) { |
| |
| RECEIVE(conn_name, request_buffer, length) -> count; |
| |
| <Process request_buffer.> |
| } |
| |
| <Compute reply and store into reply_buffer.> |
| |
| SEND(conn_name, reply_buffer, length, |
| PUSH=YES, URG=NO, EOF=YES); |
| |
| |
| Figure 8B: Server Side User Interface |
|__________________________________________________________________|
4. IMPLEMENTATION ISSUES
4.1 RFC-1323 Extensions
A recently-proposed set of TCP enhancements [RFC-1323] defines a
Timestamps option, which carries two 32-bit timestamp values.
This option is used to accurately measure round-trip time (RTT).
The same option is also used in a procedure known as "PAWS"
(Protect Against Wrapped Sequence) to prevent erroneous data
delivery due to a combination of old duplicate segments and
sequence number reuse at very high bandwidths. The approach to
transactions specified in this memo is independent of the RFC-1323
enhancements, but implementation of RFC-1323 is desirable for all
TCP's.
The RFC-1323 extensions share several common implementation issues
with the T/TCP extensions. Both require that TCP headers carry
options. Accommodating options in TCP headers requires changes in
the way that the maximum segment size is determined, to prevent
inadvertent IP fragmentation. Both require some additional state
variable in the TCB, which may or may not cause implementation
difficulties.
4.2 Minimal Packet Sequence Most TCP implementations will require some small modifications to allow the minimal packet sequence for a transaction shown in Figure 2. Many TCP implementations contain a mechanism to delay acknowledgments of some subset of the data segments, to cut down on the number of acknowledgment segments and to allow piggybacking on the reverse data flow (typically character echoes). To obtain minimal packet exchanges for transactions, it is necessary to delay the acknowledgment of some control bits, in an analogous manner. In particular, the <SYN,ACK> segment that is to be sent in ESTABLISHED* or CLOSE-WAIT* state should be delayed. Note that the amount of delay is determined by the minimum RTO at the transmitter; it is a parameter of the communication protocol, independent of the application. We propose to use the same delay parameter (and if possible, the same mechanism) that is used for delaying data acknowledgments. To get the FIN piggy-backed on the reply data (segment #3 in Figure 2), thos implementations that have an implied PUSH=YES on all SEND calls will need to augment the user interface so that PUSH=NO can be set for transactions. 4.3 RTT Measurement Transactions introduce new issues into the problem of measuring round trip times [Jacobson88]. (a) With the minimal 3-segment exchange, there can be exactly one RTT measurement in each direction for each transaction. Since dynamic estimation of RTT cannot take place within a single transaction, it must take place across successive transactions. Therefore, cacheing the measured RTT and RTT variance values is essential for transaction processing; in normal virtual circuit communication, such cacheing is only desirable. (b) At the completion of a transaction, the values for RTT and RTT variance that are retained in the cache must be some average of previous values with the values measured during the transaction that is completing. This raises the question of the time constant for this average; quite different dynamic considerations hold for transactions than for file transfers, for example. (c) An RTT measurement by the client will yield the value:
T = RTT + min(SPT, ATO),
where SPT (server processing time) was defined in the
introduction, and ATO is the timeout period for sending a
delayed ACK. Thus, the measured RTT includes SPT, which may
be arbitrarily variable; however, the resulting variability
of the measured T cannot exceed ATO. (In a popular TCP
implementation, for example, ATO = 200ms, so that the
variance of SPT makes a relatively small contribution to the
variance of RTT.)
(d) Transactions sample the RTT at random times, which are
determined by the client and the server applications rather
than by the network dynamics. When there are long pauses
between transactions, cached path properties will be poor
predictors of current values in the network.
Thus, the dynamics of RTT measurement for transactions differ from
those for virtual circuits. RTT measurements should work
correctly for very short connections but reduce to the current TCP
algorithms for long-lasting connections. Further study is this
issue is needed.
4.4 Cache Implementation
This extension requires a per-host cache of connection counts.
This cache may also contain values of the smoothed RTT, RTT
variance, congestion avoidance threshold, and MSS values.
Depending upon the implementation details, it may be simplest to
build a new cache for these values; another possibility is to use
the routing cache that should already be included in the host
[RFC-1122].
Implementation of the cache may be simplified because it is
consulted only when a connection is established; thereafter, the
CC values relevant to the connection are kept in the TCB. This
means that a cache entry may be safely reused during the lifetime
of a connection, avoiding the need for locking.
4.5 CPU Performance
TCP implementations are customarily optimized for streaming of
data at high speeds, not for opening or closing connections.
Jacobson's Header Prediction algorithm [Jacobson90] handles the
simple common cases of in-sequence data and ACK segments when
streaming data. To provide good performance for transactions, an
implementation might be able to do an analogous "header
prediction" specifically for the minimal request and the response
segments.
The overhead of UDP provides a lower bound on the overhead of
TCP-based transaction processing. It will probably not be
possible to reach this bound for TCP transactions, since opening a
TCP connection involves creating a significant amount of state
that is not required by UDP.
McKenney and Dove [McKenney92] have pointed out that transaction
processing applications of TCP can stress the performance of the
demultiplexing algorithm, i.e., the algorithm used to look up the
TCB when a segment arrives. They advocate the use of hash-table
techniques rather than a linear search. The effect of
demultiplexing on performance may become especially acute for a
transaction client using the extended TCP described here, due to
TCB's left in TIME-WAIT state. A high rate of transactions from a
given client will leave a large number of TCB's in TIME-WAIT
state, until their timeout expires. If the TCP implementation
uses a linear search for demultiplexing, all of these control
blocks must be traversed in order to discover that the new
association does not exist. In this circumstance, performance of
a hash table lookup should not degrade severely due to
transactions.
4.6 Pre-SYN Queue
Suppose that segment #1 in Figure 4 is lost in the network; when
segment #2 arrives in LISTEN state, it will be ignored by the TCP
rules (see [STD-007] p.66, "fourth other text and control"), and
must be retransmitted. It would be possible for the server side
to queue any ACK-less data segments received in LISTEN state and
to "replay" the segments in this queue when a SYN segment does
arrive. A data segment received with an ACK bit, which is the
normal case for existing TCP's, would still a generate RST
segment.
Note that queueing segments in LISTEN state is different from
queueing out-of-order segments after the connection is
synchronized. In LISTEN state, the sequence number corresponding
to the left window edge is not yet known, so that the segment
cannot be trimmed to fit within the window before it is queued.
In fact, no processing should be done on a queued segment while
the connection is still in LISTEN state. Therefore, a new "pre-
SYN queue" would be needed. A timeout would be required, to flush
the Pre-SYN Queue in case a SYN segment was not received.
Although implementation of a pre-SYN queue is not difficult in BSD
TCP, its limited contribution to throughput probably does not
justify the effort.
6. ACKNOWLEDGMENTS
I am very grateful to Dave Clark for pointing out bugs in RFC-1379
and for helping me to clarify the model. I also wish to thank Greg
Minshall, whose probing questions led to further elucidation of the
issues in T/TCP.
7. REFERENCES
[Jacobson88] Jacobson, V., "Congestion Avoidance and Control", ACM
SIGCOMM '88, Stanford, CA, August 1988.
[Jacobson90] Jacobson, V., "4BSD Header Prediction", Comp Comm
Review, v. 20, no. 2, April 1990.
[McKenney92] McKenney, P., and K. Dove, "Efficient Demultiplexing
of Incoming TCP Packets", ACM SIGCOMM '92, Baltimore, MD, October
1992.
[RFC-1122] Braden, R., Ed., "Requirements for Internet Hosts --
Communications Layers", STD-3, RFC-1122, USC/Information Sciences
Institute, October 1989.
[RFC-1323] Jacobson, V., Braden, R., and D. Borman, "TCP Extensions
for High Performance, RFC-1323, LBL, USC/Information Sciences
Institute, Cray Research, February 1991.
[RFC-1379] Braden, R., "Transaction TCP -- Concepts", RFC-1379,
USC/Information Sciences Institute, September 1992.
[ShankarLee93] Shankar, A. and D. Lee, "Modulo-N Incarnation
Numbers for Cache-Based Transport Protocols", Report CS-TR-3046/
UIMACS-TR-93-24, University of Maryland, March 1993.
[STD-007] Postel, J., "Transmission Control Protocol - DARPA
Internet Program Protocol Specification", STD-007, RFC-793,
USC/Information Sciences Institute, September 1981.
APPENDIX A. ALGORITHM SUMMARY This appendix summarizes the additional processing rules introduced by T/TCP. We define the following symbols: Options CC(SEG.CC): TCP Connection Count (CC) Option CC.NEW(SEG.CC): TCP CC.NEW option CC.ECHO(SEG.CC): TCP CC.ECHO option Here SEG.CC is option value in segment. Per-Connection State Variables in TCB CCsend: CC value to be sent in segments CCrecv: CC value to be received in segments Elapsed: Duration of connection Global Variables: CCgen: CC generator variable cache.CC[fh]: Cache entry: Last CC value received. cache.CCsent[fh]: Cache entry: Last CC value sent. PSEUDO-CODE SUMMARY: Passive OPEN => { Create new TCB; } Active OPEN => { <Create new TCB> CCrecv = 0; CCsend = CCgen; If (CCgen == 0xffffffff) then Set CCgen = 1; else Set CCgen = CCgen + 1. <Send initial {SYN} segment (see below)> } Send initial {SYN} segment => { If (cache.CCsent[fh] == 0 OR CCsend < cache.CCsent[fh] ) then { Include CC.NEW(CCsend) option in segment; Set cache.CCsent[fh] = 0;
}
else {
Include CC(CCsend) option in segment;
Set cache.CCsent[fh] = CCsend;
}
}
Send {SYN,ACK} segment => {
If (CCrecv != 0) then
Include CC(CCsend), CC.ECHO(CCrecv) options in segment.
}
Receive {SYN} segment in LISTEN, SYN-SENT, or SYN-SENT* state => {
If state == LISTEN then {
CCrecv = 0;
CCsend = CCgen;
If (CCgen == 0xffffffff) then Set CCgen = 1;
else Set CCgen = CCgen + 1.
}
If (Segment contains CC option OR
Segment contains CC.NEW option) then
Set CCrecv = SEG.CC.
if (Segment contains CC option AND
cache.CC[fh] != 0 AND
SEG.CC > cache.CC[fh] ) then { /* TAO Test OK */
Set cache.CC[fh] = CCrecv;
<Mark connection half-synchronized>
<Process data and/or FIN and return>
}
If (Segment does not contain CC option) then
Set cache.CC[fh] = 0;
<Do normal TCP processing and return>.
}
Receive {SYN} segment in LISTEN-TW, LISTEN-LA, LISTEN-LA*, LISTEN-CL,
or LISTEN-CL* state => {
If ( (Segment contains CC option AND CCrecv != 0 ) then {
If (state = LISTEN-TW AND Elapsed > MSL ) then
<Send RST, drop segment, and return>.
if (SEG.CC > CCrecv ) then {
<Implicitly ACK FIN and data in retransmission queue>;
<Close and delete TCB>;
<Reprocess segment>.
/* Expect to match new TCB
* in LISTEN state.
*/
}
}
else
<Drop segment>.
}
Receive {SYN,ACK} segment => {
if (Segment contains CC.ECHO option AND
SEG.CC != CCsend) then
<Send a reset and discard segment>.
if (Segment contains CC option) then {
Set CCrecv = SEG.CC.
if (cache.CC[fh] is undefined) then
Set cache.CC[fh] = CCrecv.
}
}
Send non-SYN segment => {
if (CCrecv != 0 OR
(cache.CCsent[fh] != 0 AND
state is SYN-SENT or SYN-SENT*)) then
Include CC(CCsend) option in segment.
}
Receive non-SYN segment in SYN-RECEIVED state => {
if (Segment contains CC option AND RST bit is off) {
if (SEG.CC != CCrecv) then
<Segment is unacceptable; drop it and send an
ACK segment, as in normal TCP processing>.
if (cache.CC[fh] is undefined) then
Set cache.CC[fh] = CCrecv.
}
}
Receive non-SYN segment in (state >= ESTABLISHED) => {
if (Segment contains CC option AND RST bit is off) {
if (SEG.CC != CCrecv) then
<Segment is unacceptable; drop it and send an
ACK segment, as in normal TCP processing>.
}
}
Security Considerations
Security issues are not discussed in this memo.
Author's Address
Bob Braden
University of Southern California
Information Sciences Institute
4676 Admiralty Way
Marina del Rey, CA 90292
Phone: (310) 822-1511
EMail: Braden@ISI.EDU