The TCP protocol [Postel81] was designed to operate reliably over almost any transmission medium regardless of transmission rate, delay, corruption, duplication, or reordering of segments. Production TCP implementations currently adapt to transfer rates in the range of 100 bps to 10**7 bps and round-trip delays in the range 1 ms to 100 seconds. Recent work on TCP performance has shown that TCP can work well over a variety of Internet paths, ranging from 800 Mbit/sec I/O channels to 300 bit/sec dial-up modems [Jacobson88a].
The introduction of fiber optics is resulting in ever-higher transmission speeds, and the fastest paths are moving out of the domain for which TCP was originally engineered. This memo defines a set of modest extensions to TCP to extend the domain of its application to match this increasing network capability. It is based upon and obsoletes RFC-1072 [Jacobson88b] and RFC-1185 [Jacobson90b].
There is no one-line answer to the question: "How fast can TCP go?". There are two separate kinds of issues, performance and reliability, and each depends upon different parameters. We discuss each in turn.
TCP performance depends not upon the transfer rate itself, but rather upon the product of the transfer rate and the round-trip delay. This "bandwidth*delay product" measures the amount of data that would "fill the pipe"; it is the buffer space required at sender and receiver to obtain maximum throughput on the TCP connection over the path, i.e., the amount of unacknowledged data that TCP must handle in order to keep the pipeline full. TCP performance problems arise when the bandwidth*delay product is large. We refer to an Internet path operating in this region as a "long, fat pipe", and a network containing this path as an "LFN" (pronounced "elephan(t)").
High-capacity packet satellite channels (e.g., DARPA's Wideband Net) are LFN's. For example, a DS1-speed satellite channel has a bandwidth*delay product of 10**6 bits or more; this corresponds to 100 outstanding TCP segments of 1200 bytes each. Terrestrial fiber-optical paths will also fall into the LFN class; for example, a cross-country delay of 30 ms at a DS3 bandwidth (45Mbps) also exceeds 10**6 bits.
There are three fundamental performance problems with the current TCP over LFN paths:
The TCP header uses a 16 bit field to report the receive window size to the sender. Therefore, the largest window that can be used is 2**16 = 65K bytes.
To circumvent this problem, Section 2 of this memo defines a new TCP option, "Window Scale", to allow windows larger than 2**16. This option defines an implicit scale factor, which is used to multiply the window size value found in a TCP header to obtain the true window size.
Packet losses in an LFN can have a catastrophic effect on throughput. Until recently, properly-operating TCP implementations would cause the data pipeline to drain with every packet loss, and require a slow-start action to recover. Recently, the Fast Retransmit and Fast Recovery algorithms [Jacobson90c] have been introduced. Their combined effect is to recover from one packet loss per window, without draining the pipeline. However, more than one packet loss per window typically results in a retransmission timeout and the resulting pipeline drain and slow start.
Expanding the window size to match the capacity of an LFN results in a corresponding increase of the probability of more than one packet per window being dropped. This could have a devastating effect upon the throughput of TCP over an LFN. In addition, if a congestion control mechanism based upon some form of random dropping were introduced into gateways, randomly spaced packet drops would become common, possible increasing the probability of dropping more than one packet per window.
To generalize the Fast Retransmit/Fast Recovery mechanism to handle multiple packets dropped per window, selective acknowledgments are required. Unlike the normal cumulative acknowledgments of TCP, selective acknowledgments give the sender a complete picture of which segments are queued at the receiver and which have not yet arrived. Some evidence in favor of selective acknowledgments has been published [NBS85], and selective acknowledgments have been included in a number of experimental Internet protocols -- VMTP [Cheriton88], NETBLT [Clark87], and RDP [Velten84], and proposed for OSI TP4 [NBS85]. However, in the non-LFN regime, selective acknowledgments reduce the number of packets retransmitted but do not otherwise improve performance, making their complexity of questionable value. However, selective acknowledgments are expected to become much more important in the LFN regime.
RFC-1072 defined a new TCP "SACK" option to send a selective acknowledgment. However, there are important technical issues to be worked out concerning both the format and semantics of the SACK option. Therefore, SACK has been omitted from this package of extensions. It is hoped that SACK can "catch up" during the standardization process.
TCP implements reliable data delivery by retransmitting segments that are not acknowledged within some retransmission timeout (RTO) interval. Accurate dynamic determination of an appropriate RTO is essential to TCP performance. RTO is determined by estimating the mean and variance of the measured round-trip time (RTT), i.e., the time interval between sending a segment and receiving an acknowledgment for it [Jacobson88a].
Section 4 introduces a new TCP option, "Timestamps", and then defines a mechanism using this option that allows nearly every segment, including retransmissions, to be timed at negligible computational cost. We use the mnemonic RTTM (Round Trip Time Measurement) for this mechanism, to distinguish it from other uses of the Timestamps option.
Now we turn from performance to reliability. High transfer rate enters TCP performance through the bandwidth*delay product. However, high transfer rate alone can threaten TCP reliability by violating the assumptions behind the TCP mechanism for duplicate detection and sequencing.
An especially serious kind of error may result from an accidental reuse of TCP sequence numbers in data segments. Suppose that an "old duplicate segment", e.g., a duplicate data segment that was delayed in Internet queues, is delivered to the receiver at the wrong moment, so that its sequence numbers falls somewhere within the current window. There would be no checksum failure to warn of the error, and the result could be an undetected corruption of the data. Reception of an old duplicate ACK segment at the transmitter could be only slightly less serious: it is likely to lock up the connection so that no further progress can be made, forcing an RST on the connection.
TCP reliability depends upon the existence of a bound on the lifetime of a segment: the "Maximum Segment Lifetime" or MSL. An MSL is generally required by any reliable transport protocol, since every sequence number field must be finite, and therefore any sequence number may eventually be reused. In the Internet protocol suite, the MSL bound is enforced by an IP-layer mechanism, the "Time-to-Live" or TTL field.
Duplication of sequence numbers might happen in either of two ways:
A TCP sequence number contains 32 bits. At a high enough transfer rate, the 32-bit sequence space may be "wrapped" (cycled) within the time that a segment is delayed in queues.
Suppose that a connection terminates, either by a proper close sequence or due to a host crash, and the same connection (i.e., using the same pair of sockets) is immediately reopened. A delayed segment from the terminated connection could fall within the current window for the new incarnation and be accepted as valid.
Duplicates from earlier incarnations, Case (2), are avoided by enforcing the current fixed MSL of the TCP spec, as explained in Section 5.3 and Appendix B. However, case (1), avoiding the reuse of sequence numbers within the same connection, requires an MSL bound that depends upon the transfer rate, and at high enough rates, a new mechanism is required.
More specifically, if the maximum effective bandwidth at which TCP is able to transmit over a particular path is B bytes per second, then the following constraint must be satisfied for error-free operation:
2**31 / B > MSL (secs) 
The following table shows the value for Twrap = 2**31/B in seconds, for some important values of the bandwidth B:
Network B*8 B Twrap bits/sec bytes/sec secs _______ _______ ______ ______ ARPANET 56kbps 7KBps 3*10**5 (~3.6 days) DS1 1.5Mbps 190KBps 10**4 (~3 hours) Ethernet 10Mbps 1.25MBps 1700 (~30 mins) DS3 45Mbps 5.6MBps 380 FDDI 100Mbps 12.5MBps 170 Gigabit 1Gbps 125MBps 17
It is clear that wrap-around of the sequence space is not a problem for 56kbps packet switching or even 10Mbps Ethernets. On the other hand, at DS3 and FDDI speeds, Twrap is comparable to the 2 minute MSL assumed by the TCP specification [Postel81]. Moving towards gigabit speeds, Twrap becomes too small for reliable enforcement by the Internet TTL mechanism.
The 16-bit window field of TCP limits the effective bandwidth B to 2**16/RTT, where RTT is the round-trip time in seconds [McKenzie89]. If the RTT is large enough, this limits B to a value that meets the constraint  for a large MSL value. For example, consider a transcontinental backbone with an RTT of 60ms (set by the laws of physics). With the bandwidth*delay product limited to 64KB by the TCP window size, B is then limited to 1.1MBps, no matter how high the theoretical transfer rate of the path. This corresponds to cycling the sequence number space in Twrap= 2000 secs, which is safe in today's Internet.
It is important to understand that the culprit is not the larger window but rather the high bandwidth. For example, consider a (very large) FDDI LAN with a diameter of 10km. Using the speed of light, we can compute the RTT across the ring as (2*10**4)/(3*10**8) = 67 microseconds, and the delay*bandwidth product is then 833 bytes. A TCP connection across this LAN using a window of only 833 bytes will run at the full 100mbps and can wrap the sequence space in about 3 minutes, very close to the MSL of TCP. Thus, high speed alone can cause a reliability problem with sequence number wrap-around, even without extended windows.
Watson's Delta-T protocol [Watson81] includes network-layer mechanisms for precise enforcement of an MSL. In contrast, the IP mechanism for MSL enforcement is loosely defined and even more loosely implemented in the Internet. Therefore, it is unwise to depend upon active enforcement of MSL for TCP connections, and it is unrealistic to imagine setting MSL's smaller than the current values (e.g., 120 seconds specified for TCP).
A possible fix for the problem of cycling the sequence space would be to increase the size of the TCP sequence number field. For example, the sequence number field (and also the acknowledgment field) could be expanded to 64 bits. This could be done either by changing the TCP header or by means of an additional option.
Section 5 presents a different mechanism, which we call PAWS (Protect Against Wrapped Sequence numbers), to extend TCP reliability to transfer rates well beyond the foreseeable upper limit of network bandwidths. PAWS uses the TCP Timestamps option defined in Section 4 to protect against old duplicates from the same connection.
The extensions defined in this memo all use new TCP options. We must address two possible issues concerning the use of TCP options: (1) compatibility and (2) overhead.
We must pay careful attention to compatibility, i.e., to interoperation with existing implementations. The only TCP option defined previously, MSS, may appear only on a SYN segment. Every implementation should (and we expect that most will) ignore unknown options on SYN segments. However, some buggy TCP implementation might be crashed by the first appearance of an option on a non-SYN segment. Therefore, for each of the extensions defined below, TCP options will be sent on non-SYN segments only when an exchange of options on the SYN segments has indicated that both sides understand the extension. Furthermore, an extension option will be sent in a <SYN,ACK> segment only if the corresponding option was received in the initial <SYN> segment.
A question may be raised about the bandwidth and processing overhead for TCP options. Those options that occur on SYN segments are not likely to cause a performance concern. Opening a TCP connection requires execution of significant special-case code, and the processing of options is unlikely to increase that cost significantly.
On the other hand, a Timestamps option may appear in any data or ACK segment, adding 12 bytes to the 20-byte TCP header. We believe that the bandwidth saved by reducing unnecessary retransmissions will more than pay for the extra header bandwidth.
There is also an issue about the processing overhead for parsing the variable byte-aligned format of options, particularly with a RISC-architecture CPU. To meet this concern, Appendix A contains a recommended layout of the options in TCP headers to achieve reasonable data field alignment. In the spirit of Header Prediction, a TCP can quickly test for this layout and if it is verified then use a fast path. Hosts that use this canonical layout will effectively use the options as a set of fixed-format fields appended to the TCP header. However, to retain the philosophical and protocol framework of TCP options, a TCP must be prepared to parse an arbitrary options field, albeit with less efficiency.
Finally, we observe that most of the mechanisms defined in this memo are important for LFN's and/or very high-speed networks. For low-speed networks, it might be a performance optimization to NOT use these mechanisms. A TCP vendor concerned about optimal performance over low-speed paths might consider turning these extensions off for low-speed paths, or allow a user or installation manager to disable them.