TCP, the workhorse of the internet
cefboud.com
cefboud.com
The three drawbacks of the original TCP algorithm were the window size (the maximum value is just too small for today's speeds), poor handling of missing packets (addressed by extensions such as selective-ACK), and the fact that it only manages one stream at a time, and some applications want multiple streams that don't block each other. You could use multiple TCP connections, but that adds its own overhead, so SCTP and QUIC were designed to address those issues.
The congestion control algorithm is not part of the on-the-wire protocol, it's just some code on each side of the connection that decides when to (re)send packets to make the best use of the available bandwidth. Anything that implements a reliable stream on top of datagrams needs to implement such an algorithm. The original ones (Reno, Vegas, etc) were very simple but already did a good job, although back then network equipment didn't have large buffers. A lot of research is going into making better algorithms that handle large buffers, large roundtrip times, varying bandwidth needs and also being fair when multiple connections share the same bandwidth.
I'll add that at the time of TCP's writing, the telephone people far outnumbered everyone else in the packet switching vs circuit switching debate. TCP gives you a virtual circuit over a packet switched network as a pair of reliable-enough independent byte streams over IP. This idea, that the endpoints could implement reliability through retransmission came from an earlier French network, Cylades, and ends up being a core principle of IP networks.
Circuit switching packets means carrying metadata about the circuit rather than simply using the destination MAC or IP address to figure out routing along the way. ATM took this to an extreme with nearly 10% protocol overhead (48 bytes of payload in a 53 byte cell) and 22 bytes of wasted space in the last ATM cell for a 1500 byte ethernet packet. That inefficiency is what really hurt. Sadly the ATM legacy lives on in GPON and XGSPON -- EPON / 10GEPON are far better protocols. As a result, GPON and XGSPON require gobs of memory per port for frame reassembly (128 ONUs x 8 priorities x 9KB for jumbo frames = 9MB per port worst case), whereas EPON / 10GEPON do not.
MPLS also has certain issues that are solved by using the IPv6 next header feature which avoids having to push / pop headers (modifying the size of the packet which has implications for buffering and the associated QoS issues making the hardware more complex) in the transport network. MPLS labels made sense at the time of introduction in the early 2000s when transport network hardware was able to utilize a small table to look up the next hop of a frame instead of doing a full route lookup. The hardware constraints of those early days requiring small SRAMs have effectively gone away since modern ASICs have billions of transistors which make on chip route tables sufficient for many use-cases.
I did specify more expensive. Even with ASICs it’s more expensive to scale up.
> ATM took this to an extreme with nearly 10% protocol overhead (48 bytes of payload in a 53 byte cell) and 22 bytes of wasted space in the last ATM cell for a 1500 byte ethernet packet. That inefficiency is what really hurt. Sadly the ATM legacy lives on in GPON and XGSPON -- EPON / 10GEPON are far better protocols. As a result, GPON and XGSPON require gobs of memory per port for frame reassembly (128 ONUs x 8 priorities x 9KB for jumbo frames = 9MB per port worst case), whereas EPON / 10GEPON do not.
ATM was a technology designed and pushed by the traditional voice telecoms long before everything converged on IP. The smaller byte sizes were designed to have less jitter/latency for voice, QoS (also to prioritize voice), and more fine-grained multiplexing for voice traffic, where other data was secondary. ATM was originally designed in the late 1980s before Ethernet won out and bulk data transfer was a novelty; the traditional telecoms wanted something that mapped into their current circuits.
ATM’s overhead for IP was arguably too much, but it wasn’t what killed ATM. It was the fact that speeding the sending of such small packets up to and past gigabit speeds was too expensive, which was made worse by the fact that Ethernet became a commodity due to scale.
> Sadly the ATM legacy lives on in GPON and XGSPON -- EPON / 10GEPON are far better protocols. As a result, GPON and XGSPON require gobs of memory per port for frame reassembly (128 ONUs x 8 priorities x 9KB for jumbo frames = 9MB per port worst case), whereas EPON / 10GEPON do not.
This isn’t quite the same argument. I agree that 10GEPON is “better” and I wish it was THE standard (especially for pure ISPs), but it ignores that most ISPs using XGSPON are multiplexing voip, TV, emergency, and other traditional networks. If anything, it’s what ATM should have been, where they could have had a different priority for small celled VOIP and larger packets for other services. I say this as a person who thinks all this is dumb and consumers should just be given an internet connection at this point - I hope 10GEPON wins out in the end - it is certainly already cheaper. I much more hate the fact that I have to use PPPoE than some memory overhead for the ports reassembling GEM packets, though.
As for MPLS, well yeah it was certainly faster than IP lookups, but the circuits also often result in sub-optimal routing as the MPLS “tunnels” don’t always reflect otherwise ideal physical paths. IIRC, about half of the IPv6 internet’s core is actually routed over MPLS tunnels and it can be a large reason IPv6 routing can often have higher TTLs than IPv4 (because the paths often aren’t as efficient). That being said, we’ll have to see if segment routing takes off, and what approaches stick.
The only way to get what you want is to have dedicated line, which isnt scalable.
What about QoS? Jitter, bandwidth, latency, fairness guarantees? What about queuing delay? What about multiplexing and tunneling? Traffic shaping and engineering? What about long-haul performance? Easy integration with optical circuit networks? etc. ATM addressed these issues, but TCP/IP did not.
All of these things showed up again once you tried to do VOIP and video conferencing, and in core ISPs as well as access networks, and they weren't (and in many cases still aren't) easy to solve.
Also MPLS is basically a virtual circuit network.
And why are they trying to persuade regulators to let them get rid of the remaining (peripheral) part of the old circuit-switched network, i.e., to phase out old-school telephone hardware, requiring all customers to have IP phone hardware?
Modern circuit switched networks use optics rather than the legacy copper circuits which date back to telegraphy.
> poor handling of missing packets
so it was poor at exact thing it was designed for?
) compared to when TCP was invented.
When I started at university the ftp speed from the US during daytime was 500 bytes per second! You don't have many unacknowledged packages in such a connection.
Back then even a 1 megabits/sec connection was super high speed and very expensive.
I'll take flak for saying it, but I feel web developers are partially at fault for laziness on this one. I've often seen them trigger a swath of connections (e.g. for uncoordinated async events), when carefully managed multiplexing over one or a handful will do just fine.
Eg. In prehistoric times I wrote a JavaScript library that let you queue up several downloads over one stream, with control over prioritization and cancelability.
It was used in a GreaseMonkey script on a popular dating website, to fetch thumbnails and other details of all your matches in the background. Hovering over a match would bring up all their photos, and if some hadn't been retrieved yet they'd immediately move to the top of the queue. I intentionally wanted to limit the number of connections, to avoid oversaturating the server or the user's bandwidth. Idle time was used to prefetch all matches on the page (IIRC in a sensible order responsive to your scroll location). If you picked a large enough pagination, then stepped away to top up your coffee, by the time you got back you could browse through all of your recent matches instantly, without waiting for any server roundtrip lag.
It was pretty slick. I realize these days modern stacks give you multiplexing for free, but to put in context this was created in the era before even JQuery was well-known.
Funny story, I shared it with one of my matches and she found it super useful but was a bit surprised that, in a way, I was helping my competition. Turned out OK... we're still together nearly two decades later and now she generously jokes I invented Tinder before it was a thing.
But you have to include giant libraries and kernel can’t see the traffic to better manage timing etc.
It only takes a few thousand lines (easily less than 10k even with zero dependencies and no standard library) to implement QUIC.
Kernel management of transport protocols has zero actual benefit for latency or throughput given proper network stack design. Neither does hardware offload except for crypto offload. Claimed differences are just due to poor network stack design and poor protocol implementation.
Not fully standards compliant since I skipped some irrelevant details like bidirectional streams when I can just make a pair of unidirectional streams, but handles all of the core connection setup and transport logic. It is not actually that complicated. And just to get ahead of it, performance is perfectly comparable.
CDNs became popular back in the old days, when some people thought that if two websites are using jquery-1.2.3.min.js, CDN could cache it and second site would load quicker. These days, browser don't do that, they'll ignore cached assets from other websites because it somehow helps to protect user privacy and they value privacy over performance in this case.
There are some reasons CDNs might be helpful. Edge capability probably is the most important one. Another reason is that serving lots of static data might be a complicated task for a small website, so it makes sense to offload it to a specialised service. These days, CDNs went beyond static data. They can hide your backend, so public user won't know its address and can't DDoS it. They can handle TLS for you. They can filter bots, tor and people from countries you don't like. All in a few clicks in the dashboard, no need to implement complicated solutions.
But nothing you couldn't write yourself in a few days, really.
Generally they only do caching on a local basis and not across the entire world
So it's still up to the developer to provide the correct headers otherwise you don't get too much of a benefit
That said some of them will do some default caching if it is recognized as a static file etc
The Linux kernel supports it but at least when I had tried this those modules were disabled on most distros.
Other applications work just fine with a single TCP connection
If I am using TCP for DNS, for example, and I am retrieving data from a single host such as a DNS cache, I can send multiple queries over a single TCP connection and receive multiple responses over the same single TCP single connection, out of order. No blocking.^1 If the cache (application) supports it, this is much faster than receiving answers sequentially and it's more efficient and polite than opening multiple TCP connections
1. I do this every day outside the browser with DNS over TLS (DoT) using something like streamtcp from NLNet Labs. I'm not sure that QUIC is faster, server support for QUIC is much more limited, but QUIC may have other advantages
I also do it with DNS over HTTPS (DoH), outside the browser, using HTTP/1.1 pipelining, but there I receive answers sequentially. I'm still not convinced that HTTP/2 is faster for this particular use case, i.e., downloading data from a single host using multiple HTTP requests (compared to something like integrating online advertising into websites, for example)
You're missing the point. You have one TCP connection, and the sever sends you response1 and then response2. Now if response1 gets lost or delayed due to network conditions, you must wait for response1 to be retransmitted before you can read response2. That is blocking, no way around it. It has nothing to do with advertising(?), and the other protocols mentioned don't have this drawback.
It’s a trade off, and there’s a surprising amount of application code involved on the receiving side in the application waiting for state to be updated on both channels. I definitely prefer it, but it’s not without its tradeoffs.
This is because DoT allows the DNS server to resolve queries concurrently and send query responses out of order.
However, this is an application layer feature, not a transport layer one. The underlying TCP packets still have to arrive in order and therefore are subject to blocking.
• Full-duplex connections are probably a good idea, but certainly are not the only way, or the most obvious way, to create a reliable stream of data on top of an unreliable datagram layer. TCP's predecessor NCP was half-duplex.
• TCP itself also supports a half-duplex mode—even if one end sends FIN, the other end can keep transmitting as long as it wants. This was probably also a good idea, but it's certainly not the only obvious choice.
• Sequence numbers on messages or on bytes?
• Wouldn't it be useful to expose message boundaries to applications, the way 9P, SCTP, and some SNA protocols do?
• If you expose message boundaries to applications, maybe you'd also want to include a message type field? Protocol-level message-type fields have been found to be very useful in Ethernet and IP, and in a sense the port-number field in UDP is also a message-type field.
• Do you really need urgent data?
• Do servers need different port numbers? TCPMUX is a straightforward way of giving your servers port names, like in CHAOSNET, instead of port numbers. It only creates extra overhead at connection-opening time, assuming you have the moral equivalent of file descriptor passing on your OS. The only limitation is that you have to use different client ports for multiple simultaneous connections to the same server host. But in TCP everyone uses different client ports for different connections anyway. TCPMUX itself incurs an extra round-trip time delay for connection establishment, because the requested server name can't be transmitted until the client's ACK packet, but if you incorporated it into TCP, you'd put the server name in the SYN packet. If you eliminate the server port number in every TCP header, you can expand the client port number to 24 or even 32 bits.
• Alternatively, maybe network addresses should be assigned to server processes, as in Appletalk (or IP-based virtual hosting before HTTP/1.1's Host: header, or, for TLS, before SNI became widespread), rather than assigning network addresses to hosts and requiring port numbers or TCPMUX to distinguish multiple servers on the same host?
• Probably SACK was actually a good idea and should have always been the default? SACK gets a lot easier if you ack message numbers instead of byte numbers.
• Why is acknowledgement reneging allowed in TCP? That was a terrible idea.
• It turns out that measuring round-trip time is really important for retransmission, and TCP has no way of measuring RTT on retransmitted packets, which can pose real problems for correcting a ridiculously low RTT estimate, which results in excessive retransmission.
• Do you really need a PUSH bit? C'mon.
• A modest amount of overhead in the form of erasure-coding bits would permit recovery from modest amounts of packet loss without incurring retransmission timeouts, which is especially useful if your TCP-layer protocol requires a modest amount of packet loss for congestion control, as TCP does.
• Also you could use a "congestion experienced" bit instead of packet loss to detect congestion in the usual case. (TCP did eventually acquire CWR and ECE, but not for many years.)
• The fact that you can't resume a TCP connection from a different IP address, the way you can with a Mosh connection, is a serious flaw that seriously impedes nodes from moving around the network.
• TCP's hardcoded timeout of 5 minutes is also a major flaw. Wouldn't it be better if the application could set that to 1 hour, 90 minutes, 12 hours, or a week, to handle intermittent connectivity, such as with communication satellites? Similarly for very-long-latency datagrams, such as those relayed by single LEO satellites. Together this and the previous flaw have resulted in TCP largely being replaced for its original session-management purpose with new ad-hoc protocols such as HTTP magic cookies, protocols which use TCP, if at all, merely as a reliable datagram protocol.
• Initial sequence numbers turn out not to be a very good defense against IP spoofing, because that wasn't their original purpose. Their original purpose was preventing the erroneous reception of leftover TCP segments from a previous incarnation of the connection that have been bouncing around routers ever since; this purpose would be better served by using a different client port number for each new connection. The ISN namespace is far too small for current LFNs anyway, so we had to patch over the hole in TCP with timestamps and PAWS.
I tend to think that one of the reasons linux containers are needed for network services is that DNS traditionally only returns an IP address (rather than address + port) so each service process needs to have its own IP address, which in linux requires a container or at least a network namespace.
AppleTalk also supported a reliable transaction (basically request-response RPC) protocol (ATP) and a session protocol, which I believe were used for Mac network services (printing, file servers, etc.) Certainly easier than serializing/deserializing byte streams.
I agree that, given the existing design of IP and TCP, you could get much of the benefit of first-class addresses for services by using, for example, DNS-SD, and that is what ZeroConf does. (It is not a coincidence that the DNS-SD RFC was written by a couple of Apple employees.) But, if that's the way you're going to be finding endpoints to initiate connections to, there's no benefit to having separate port numbers and IP addresses. And IP addresses are far scarcer than just requiring a Linux container or a network namespace: there are only 2³² of them. But it is rare to find an IP address that is listening on more than 64 of its 2¹⁶ TCP ports, so in an alternate history where you moved those 16 bits from the port number to the IP address, we would have one thousandth of the IP-address crunch that we do.
Historically, possibly the reason that it wasn't done this way is that port numbers predated the DNS by about 10 years.
The higher level protocols were built on ATP which was message based.
ADSP was a stream protocol that could be used for remote terminal access or other applications where byte streams actually made sense.
> Historically, possibly the reason that it wasn't done this way is that port numbers predated the DNS by about 10 years.
Predated or postdated?
My understanding is that DNS can potentially provide port numbers, but this is not widely used or supported.
Mockapetris's DNS RFCs are from 01983, although I think I've talked to people who installed DNS a year or two before that. Port numbers were first proposed in RFC 38 in 01970 https://datatracker.ietf.org/doc/html/rfc38
> The END and RDY must specify relevant sockets in addition to the link number. Only the local socket name need be supplied
and given actual numbers in RFC 54, also in 01970 https://datatracker.ietf.org/doc/html/rfc54
> Connections are named by a pair of sockets. Sockets are 40 bit names which are known throughout the network. Each host is assigned a private subset of these names, and a command which requests a connection names one socket which is local to the requesting host and one local to the receiver of the request.
> Sockets are polarized; even numbered sockets are receive sockets; odd numbered ones are send sockets. One of each is required to make a connection.
In RFC 129 in 01971 we see discussion about whether socketnames should include host numbers and/or user numbers, still with the low-order bit indicating the socket's gender (emissive or receptive). https://datatracker.ietf.org/doc/html/rfc129
RFC 147 later that year https://datatracker.ietf.org/doc/html/rfc147 discusses within-machine port numbers and how they should or should not relate to the socketnames transmitted in NCP packets:
> Previous network papers postulated that a process running under control of the host's operating system would have access to a number of ports. A port might be a physical input or output device, or a logical I/O device (...)
> A socket has been defined to be the identification of a port for machine to machine communication through the ARPA network. Sockets allocated to each host must be uniquely associated with a known process or be undefined. The name of some sockets must be universally known and associated with a known process operating with a specified protocol. (e.g., a logger socket, RJE socket, a file transfer socket). The name of other sockets might not be universally known, but given in a transmission over a universally known socket, (c. g. the socket pair specified by the transmission over the logger socket under the Initial Connection Protocol (ICP). In any case, communication over the network is from one socket to another socket, each socket being identified with a process running at a known host.
RFC 167 the same year https://datatracker.ietf.org/doc/html/rfc167 proposes that socketnames not be required to be unique network-wide but just within a host. It also points out that you really only need the socketname during the initial connection process, if you have some other way of knowing which packets belong to which connections:
> Although fields will be helpful in dealing with socket number allocation, it is not essential that such field designations be uniform over the network. In all network transactions the 32-bit socket number is handled with its 8-bit host number. Thus, if hosts are able to maintain uniqueness and repeatability internally, socket numbers in the network as a whole will also be unique and repeatable. If a host fails to do so, only connections with that offending host are affected.
> Because the size, use, and character of systems on the network are so varied, it would be difficult if not impossible to come up with an agreed upon particular division of the 32-bit socket number. Hosts have different internal restrictions on the number of users, processes per user, and connections per process they will permit.
> It has been suggested that it may not be necessary to maintain socket uniqueness. It is contended that there is really no significant use made of the socket number after a connection has been established. The only reason a host must now save a socket number for the life of a connection is to include it in the CLOSE of that connection.
RFC 172 in June https://datatracker.ietf.org/doc/html/rfc172 proposes using port 3 for the second version of FTP:
> [6] It seems that socket 1 has been assigned to logger. Socket 3 seems a reasonable choice for File Transfer.
This updates the first version in RFC 114 in April https://datatracker.ietf.org/doc/html/rfc114 which said:
> [16] It seems that socket 1 has been assigned to logger and socket 5 to NETRJS. Socket 3 seems a reasonable choice for the file transfer process.
RFC 196 the same year https://datatracker.ietf.org/doc/html/rfc196 proposes to use port 5 to receive mail and/or print jobs:
> Initial Connection will be as per the Official Initial Connection Protocol, Documents #2, NIC 7101, to a standard socket not yet assigned. A candidate socket number would be socket #5.
In RFC204 in August https://www.rfc-editor.org/rfc/rfc204.html Postel publishes the first list of port number assignments:
> I would like to collect information on the use of socket numbers for "standard" service programs. For example Loggers (telnet servers) Listen on socket 1. What sockets at your host are Listened to by what programs?
> Recently Dick Watson suggested assigning socket 5 for use by a mail-box protocol (RFC196). Does any one object ? Are there any suggestions for a method of assigning sockets to standard programs? Should a subset of the socket numbers be reserved for use by future standard protocols?
> Please phone or mail your answers and commtents to (...)
Amusingly in retrospect, Postel did not include an email address, presumably because they didn't have email working yet.
FTP's assignment to port 3 was confirmed in RFC 265 in November:
> Socket 3 is the standard preassigned socket number on which the cooperating file transfer process at the serving host should "listen". (*)The connection establishment will be in accordance with the standard initial connection protocol, (*)establishing a full-duplex connection.
In May of 01972 Postel published a list as RFC 349 https://www.rfc-editor.org/rfc/rfc349.html:
> I propose that there be a czar (me ?) who hands out official socket numbers for use by standard protocols. This czar should also keep track of and publish a list of those socket numbers where host specific services can be obtained. I further suggest that the initial allocation be as follows:
Sockets Assignment
0-63 Network wide standard functions
64-127 Host specific functions
128-239 Reserved for future use
240-255 Any experimental function
Socket Assignment
1 Telnet
3 File Transfer
5 Remote Job Entry
7 Echo
9 Discard
tcpmux 1/tcp # TCP port service multiplexer
echo 7/tcp
echo 7/udp
discard 9/tcp sink null
discard 9/udp sink null
systat 11/tcp users
daytime 13/tcp
daytime 13/udp
netstat 15/tcp
qotd 17/tcp quote
chargen 19/tcp ttytst source
chargen 19/udp ttytst source
ftp-data 20/tcp
ftp 21/tcp
fsp 21/udp fspd
ssh 22/tcp # SSH Remote Login Protocol
telnet 23/tcp
In December of 01972, Postel published RFC 433 https://www.rfc-editor.org/rfc/rfc433.html, obsoleting the RFC 349 list with a list including chargen and some other interesting services:
Socket Assignment
1 Telnet
3 File Transfer
5 Remote Job Entry
7 Echo
9 Discard
19 Character Generator [e.g. TTYTST]
65 Speech Data Base @ ll-tx-2 (74)
67 Datacomputer @ cca (31)
241 NCP Measurement
243 Survey Measurement
245 LINK
RFC 503 https://www.rfc-editor.org/rfc/rfc503.html from 01973 has a longer list (including systat, datetime, and netstat), but also listing which services were running on which ARPANet hosts, 33 at that time. So RFC 503 contained a list of every server process running on what would later become the internet.
Skipping RFC 604, RFC 739 from 01977 https://www.rfc-editor.org/rfc/rfc739.html is the first one that shows the modern port number assignments (still called "socket numbers") for FTP and Telnet, though those presumably dated back a couple of years at that point:
Specific Assignments:
Decimal Octal Description References
------- ----- ----------- ----------
Network Standard Functions
1 1 Old Telnet [6]
3 3 Old File Transfer [7,8,9]
5 5 Remote Job Entry [10]
7 7 Echo [11]
9 11 Discard [12]
11 13 Who is on or SYSTAT
13 15 Date and Time
15 17 Who is up or NETSTAT
17 21 Short Text Message
19 23 Character generator or TTYTST [13]
21 25 New File Transfer [1,14,15]
23 27 New Telnet [1,16,17]
25 31 Distributed Programming System [18,19]
27 33 NSW User System w/COMPASS FE [20]
29 35 MSG-3 ICP [21]
31 37 MSG-3 Authentication [21]
You say, "My understanding is that DNS can potentially provide port numbers, but this is not widely used or supported." DNS SRV records have existed since 01996 (proposed by Troll Tech and Paul Vixie in RFC 2052 https://www.rfc-editor.org/rfc/rfc2052), but they're really only widely used in XMPP, in SIP, and in ZeroConf, which was Apple's attempt to provide the facilities of AppleTalk on top of TCP/IP.
This 100% !! And basically the reason mosh had to be created in the first place (and it probably wasn't easy.) Unfortunately mosh only solves the problem for ssh. Exposing fixed IP addresses to the application layer probably doesn't help either.
So annoying that TCP tends to break whenever you switch wi-fi networks or switch from wi-fi to cellular. (On iPhones at least you have MPTCP, but that requires server-side support.)
Much of that comes from the original applications being FTP and TELNET.
• Sequence numbers on messages or on bytes?
Bytes, because the whole TCP message might not fit in an IP packet. This is the MTU problem.
• Wouldn't it be useful to expose message boundaries to applications, the way 9P, SCTP, and some SNA protocols do?
Early on, there were some message-oriented, rather than stream-oriented, protocols on top of IP. Most of them died out. RDP was one such. Another was QNet.[2] Both still have assigned IP protocol numbers, but I doubt that a RDP packet would get very far across today's internet.
This was a lack. TCP is not a great message-oriented protocol.
• Do you really need urgent data?
The purpose of urgent data is so that when your slow Teletype is typing away, and the recipient wants it to stop, there's a way to break in. See [1], p. 8.
• It turns out that measuring round-trip time is really important for retransmission, and TCP has no way of measuring RTT on retransmitted packets, which can pose real problems for correcting a ridiculously low RTT estimate, which results in excessive retransmission.
Yes, reliable RTT is a problem.
• Do you really need a PUSH bit? C'mon.
It's another legacy thing to make TELNET work on slow links. Is it even supported any more?
• A modest amount of overhead in the form of erasure-coding bits would permit recovery from modest amounts of packet loss without incurring retransmission timeouts, which is especially useful if your TCP-layer protocol requires a modest amount of packet loss for congestion control, as TCP does.
• Also you could use a "congestion experienced" bit instead of packet loss to detect congestion in the usual case. (TCP did eventually acquire CWR and ECE, but not for many years.)
Originally, there was ICMP Source Quench for that, but Berkley didn't put it in BSD, so nobody used it. Nobody was sure when to send it or what to do when it was received.
• The fact that you can't resume a TCP connection from a different IP address, the way you can with a Mosh connection, is a serious flaw that seriously impedes nodes from moving around the network.
That would require a security system to prevent hijacking sessions.
[1] https://archive.org/stream/rfc854/rfc854.txt_djvu.txt
[2] https://en.wikipedia.org/wiki/List_of_IP_protocol_numbers
- TCP should have been a reliability layer above UDB, not beside it (made P2P harder than it should be, mainly burdening teleconferencing and video games)
- Window size field bytes should have been arbitrary length
- Checksum size field bytes should have been arbitrary length and the algorithm should have been optionally customizable
- Ports should have been unique binary strings of arbitrary length instead of numbers, and not limited in count (as mentioned)
- Streams should have been encrypted by default, with clear transmission as the special case (symmetric key encryption was invented before TCP)
- IP should have connected to an arbitrary peer ID, not a MAC address, for resumable sessions if network changes (maybe only securable with encryption)
- Encrypted streams should not have been on a special port for HTTPS (not TCP's fault)
- IP address field bytes should have been arbitrary length (not TCP's fault)
- File descriptors could have been universal instead of using network sockets, unix sockets, files, pipes and bind/listen/accept/select (not TCP's fault)
- Streams don't actually make sense in the first place, we needed state transfer with arbitrary datagram size and partial sends/ranges (not TCP's fault)
https://news.ycombinator.com/item?id=44713493
I want to add that the author of the article wrote one of the cleanest and most concise summaries of the TCP protocol that I've ever read.
A stream of bytes made sense in the 1970s for remote terminal emulation. It still sort of makes sense for email, where a partial message is useful (though downloading headers in bulk followed by full message on demand probably makes more sense.)
But in 2025 much of communication involves messages that aren't useful if you only get part of them. It's also a pain to have to serialize messages into a byte stream and then deserialize the byte stream into messages (see: gRPC etc.) and the byte stream ordering is costly, doesn't work well with multipathing, and doesn't provide much benefit if you are only delivering complete messages.
TCP without congestion control isn't particularly useful. As you note traditional TCP congestion control doesn't respond well to reordering. Also TCP's congestion control traditionally doesn't distinguish between intentional packet drops (e.g. due to buffer overflow) and packet loss (e.g. due to corruption.) This means, for example that it can't be used directly over networks with wireless links (which is why wi-fi has its own link layer retransmission).
TCP's traditional congestion control is designed to fill buffers up until packets are dropped, leading to undesirable buffer bloat issues.
TCP's traditional congestion control algorithms (additive increase/multiplicative decrease on drop) also have the poor property that your data rate tends to drop as RTT increases.
TCP wasn't designed for hardware offload, which can lead to software bottlenecks and/or increased complexity when you do try to offload it to hardware.
TCP's three-way handshake is costly for one-shot RPCs, and slow start means that short flows may never make it out of slow start, neutralizing benefits from high-speed networks.
TCP is also poor for mobility. A connection breaks when your IP address changes, and there is no easy way to migrate it. Most TCP APIs expose IP addresses at the application layer, which causes additional brittleness.
Additionally, TCP is poorly suited for optical/WDM networks, which support dedicated bandwidth (signal/channel bandwidth as well as data rate), and are becoming more important in datacenters and as interconnects for GPU clusters.
etc.
Thats why it has a congestion control mechanism, its just opinionated.
> until packets are dropped, leading to undesirable buffer bloat issues.
because people have bigger buffers than needed, which means saw tooth throughput.
> TCP wasn't designed for hardware offload,
neither is quic.
> TCP's three-way handshake is costly for one-shot RPCs,
Agreed
> A connection breaks when your IP address changes, and there is no easy way to migrate it.
Also agreed, but then TCP wasn't designed for this.
> TCP is poorly suited for optical/WDM networks,
I don't think thats the case. Loads of 10gig networks are optical, and 40 gig fibre used to be DWDM (ie 4 10 gig channels in one fibre, 100gig might be the same, I've not checked) [Again, this depends on the connection type and SPF pluggy thing. It could be twinax.]
Thats what what RDMA over Converged Ethernet is for, because supposedly its cheaper than inifinband.
It is very easy for a malicious actor anywhere in the network to inject data into a connection. By contrast, it is much harder for a malicious actor to break the legitimate traffic flow ... except for the fact that TCP RST grants any rando the power to upgrade "inject" to "break". This is quite common in the wild for any traffic that does not look like HTTP, even when both endpoints are perfectly healthy.
Blocking TCP RST packets using your firewall will significantly improve reliability, but this still does not project you from more advanced attackers which cause a desynchronization due to forged sequence numbers with nonempty payload.
As a result, it is mandatory for every application to support a full-blown "resume on a separate connection" operation, which is complicated and hairy and also immediately runs into the additional flaw that TCP is very slow to start.
---
While not an outright flaw, I also think it has become clear by now that it is highly suboptimal for "address" and "port" to be separate notions.
If we fixed that issue, what would be different? The operating system still needs to assign different addresses to different sockets. So now... we have 48-bit addresses, and instead of reserving one "address" and giving each socket a 16-bit port, we reserve a /32 block of 65536 addresses and give each socket its own address with a unique 16-bit suffix?
> The Stream Control Transmission Protocol (SCTP) is a computer networking communications protocol in the transport layer of the Internet protocol suite. Originally intended for Signaling System 7 (SS7) message transport in telecommunication, the protocol provides the message-oriented feature of the User Datagram Protocol (UDP) while ensuring reliable, in-sequence transport of messages with congestion control like the Transmission Control Protocol (TCP). Unlike UDP and TCP, the protocol supports multihoming and redundant paths to increase resilience and reliability.
[…]
> SCTP may be characterized as message-oriented, meaning it transports a sequence of messages (each being a group of bytes), rather than transporting an unbroken stream of bytes as in TCP. As in UDP, in SCTP a sender sends a message in one operation, and that exact message is passed to the receiving application process in one operation. In contrast, TCP is a stream-oriented protocol, transporting streams of bytes reliably and in order. However TCP does not allow the receiver to know how many times the sender application called on the TCP transport passing it groups of bytes to be sent out. At the sender, TCP simply appends more bytes to a queue of bytes waiting to go out over the network, rather than having to keep a queue of individual separate outbound messages which must be preserved as such.
> The term multi-streaming refers to the capability of SCTP to transmit several independent streams of chunks in parallel, for example transmitting web page images simultaneously with the web page text. In essence, it involves bundling several connections into a single SCTP association, operating on messages (or chunks) rather than bytes.
* https://en.wikipedia.org/wiki/Stream_Control_Transmission_Pr...
The only good answer is "a reliability layer on top of UDP"; fortunately everybody is now rallying around QUIC as the choice for that.
QUIC is still "their own protocol", just implemented as another protocol nested inside a UDP envelope, the same way that HTTP is another protocol typically nested inside a TCP connection. It makes some sense that they'd piggyback on UDP, since (1) it doesn't require an additional IP protocol header code to be assigned by IANA, (2) QUIC definitely wants to coexist with other services on any given node, and (3) it allows whatever middleware analyses that exist for UDP to apply naturally to QUIC applications.
(Regarding (3) specifically, I imagine NAT in particular requires cooperation from residential gateways, including awareness of both the IP and the TCP/UDP port. Allowing a well-known outer UDP header to surface port information, instead of re-implementing ports somewhere in the QUIC header, means all existing NAT implementations should work unchanged for QUIC.)
Additionally, firewalls are also designed to filter out any weird packets. If the packet doesn't look like you wanted to receive it, it's dropped. It usually does this by tracking open ports just like NAT, therefore many firewalls also don't trust custom protocols.
Some people here will argue that it actually really is, and that everybody experiencing issues is just on a really weird connection or using broken hardware, but those weird connections and bad hardware make up the overwhelming majority of Internet connections these days.
Most firewalls will drop unknown IP protocols. Many will drop a lot of TCP; some drop almost all UDP. This is why so much stuff runs over tcp ports 80 and 443; it's almost always open. QUIC/HTTP/3 encourages opening of udp/443, so it's a good port to run unrelated things over too.
Also, given that SCTP had similar goals to QUIC and never got much deployment or support in OSes and NATs and firewalls and etc. It's a clear win to just use UDP and get something that will just work on a large portion of networks.
If middleware decides to do packet inspection, it better make sure that any behavioral differences (relative to not doing any inspection) is strictly an optimization and does not impact the correctness of the link.
Also, although I'm not a network operator by any stretch, my understanding is that TCP congestion control is primarily a function of the endpoints of the TCP link, not the IP routers along the way. As Wikipedia explains [0]:
> Per the end-to-end principle, congestion control is largely a function of internet hosts, not the network itself.
Classic congestion control is done on the sender alone. The router's job is simply to drop packets when the queue is too large.
Maybe the router supports ECN, so if there's a queue going to the next hop, it will look for protocol specific ECN headers to manipulate.
Some network elements do more than the usual routing work. A traffic shaper might have per-user queues with outbound bandwidth limits. A network accelerator may effectively reterminate TCP in hopes of increasing acheivable bandwidth.
Often, the router has an aggregated connection to the next hop, so it'll use a hash on the addresses in the packet to choose which of the underlying connections to use. That hash could be based on many things, but it's not uncommon to use tcp or udp port numbers if available. This can also be used to chose between equally scored next hops and that's why you often see several different paths during a traceroute. Using port numbers is helpful to balance connections from IP A to IP B over multiple links. If you us an unknown protocol, even if it is multiplexed into ports or similar (like tcp and udp), the different streams will likely always hash onto the same link and you won't be able to exceed the bandwidth of a single link and a damaged or congested link will affect all or none of your connections.
Core routers don't inspect that field, NAT/ISP boxes can. I believe that with two suitable dedicated linux servers it is very possible to send and receive single custom IP packet between them even using 253 or 254 (= Use for experimentation and testing [RFC3692]) as the protocol number
To save a skim (though it's an interesting list!), protocol codes 253 and 254 are suitable "for experimentation and testing".
I do hope we'll have stopped using IPv4 by then... But well, a decade after address exhaustion we are still on it, so who knows?
It uses them a little differently -- in IPv4, there is one protocol per packet, while in IPv6, "protocols" can be chained in a mechanism called extension headers -- but this actually makes the problem of number exhaustion more acute.
There is sometimes drama with it, though. Awhile back, the OpenBSD guys created CARP as a fully open source router failover protocol, but couldn't get an official IP number and ended up using the same one as VRRP. There's also a lot of historical animosity that some companies got numbers for proprietary protocols (eg Cisco got one for its then-proprietary EIGRP).
*The protocol.
You might have more luck with an IPv6 packet.
(It's absolutely worth reading some of those old April Fools' RFCs, by the way [0]. I'm a big fan of RFC 7168, which introduced HTTP response code 418 "I'm a teapot".)
[0]: https://en.wikipedia.org/wiki/April_Fools%27_Day_Request_for...
As soon as you start thinking about having multiple services on a host you end up with the idea of having a service id or "port"
UDP or UDP Lite gives you exactly that at the cost of 8 bytes, so there's no real value in not just putting everything on top of UDP
You left out ICMP, my favourite! (And a lot more important in IPv6 than in v4.)
Another pretty well known protocol that is neither TCP nor UDP is IPsec. (Which is really two new IP protocols.) People really did design proper IP protocols still in the 90s.
> Can I just make up a packet and send it to a host across the Internet?
You should be able to. But if you are on a corporate network with a really strict firewalling router that only forwards traffic it likes, then likely not. There are also really crappy home routers which gives similar problems from the other end of enterpriseness.
NAT also destroyed much of the end-to-end principle. If you don't have a real IP address and relies on a NAT router to forward your data, it needs to be in a protocol the router recognizes.
Anyway, for the past two decades people have grown tired of that and just piles hacks on top of TCP or UDP instead. That's sad. Or who am I kidding? Really it's on top of HTTP. HTTP will likely live on long past anything IP.
Not necessarily. Many protocols can survive being NATed if they don't carry IP/port related information inside their payload. FTP is a famous counterexample - it uses a control channel (TCP21) which contains commands to open data channels (TCP20), and those commands specify IP:port pairs, so, depending on the protocol, a NAT router has to rewrite them and/or open ports dynamically and/or create NAT entries on the fly. A lot of other stuff has no need for that and will happily go through without any rewriting.
The end-to-end principle at the IP layer (i.e. having the IP forwarding layer be agnostic to the transport layer protocols above it) is still violated.
TCP and UDP have port numbers that the NAT software can extract and keep state tables for, so we can send the return traffic to its intended destination.
For unknown IP protocols that is not possible. It may at best act like network diode, which is one way of violating the end-to-end principle.
Even ICMP has a hard time traversing NATs and firewalls these days, for largely bad reasons. Try pinging anything in AWS, for example...
If any host is firewalling out ICMP then it won't be pingable but that does not depend on the hosting provider. AWS is no better or worse than any other in that regard, IME.
A part of the problem with UDP is the lack of good platforms and tooling. Examples as well. I’m trying to help with that, but it’s an uphill battle for sure.
Some of this is hurt by the fact that v8, Node's runtime, has had first class JSON parsing support in but no support for binary protocol parsing. So writing Javascript to parse binary protocols is a lot slower than parsing JSON.
There are many routers that don't care at all about what's going through them. But there aren't any firewalls that don't route anymore (not even at the endpoints).
A bunch of multicast stuff (IGMP, PIM)
A few routing protocols (OSPF, but notably not BGP which just uses TCP, and (usually) not MPLS which just goes over the wire - it sits at the same layer as IP and not above it)
A few VPN/encapsulation solutions like GRE, IP-in-IP, L2TP and probably others I can't remember
As usual, Wikipedia has got you covered, much better than my own recollection: https://en.wikipedia.org/wiki/List_of_IP_protocol_numbers
Behind a NA(P)T, you can obviously only use those protocols that the translator knows how to remap ports for.
They absolutely don't. Routers are layer 3 devices; TCP & UDP are layer 4. The only impact is that the ECMP flow hashes will have less entropy, but that's purely an optimization thing.
Note TCP, UDP and ICMP are nowhere near all the protocols you'll commonly see on the internet — at minimum, SCTP, GRE, L2TP and ESP are reasonably widespread (even a tiny fraction of traffic is still a giant number considering internet scales).
You can send whatever protocol number with whatever contents your heart desires. Whether the other end will do anything useful with it is another question.
Idealized routers are, yes.
Actual IP paths these days usually involve at least one NAT, and these will absolutely throw away anything other than TCP, UDP, and if you're lucky ICMP.
And note the GP talked about "intermediate routers". That's the ones in a telco service site or datacenter by my book.
For example, sending some data on a fresh TCP connection is slow, and the “ramp up time” to the bandwidth of the network is almost entirely determined by the latency.
Amazing speed ups can be achieved in a data centre network by shaving microseconds off the round trip time!
Similarly, many (all?) TCP stacks count segments, not bytes, when determining this ramp up rate. This means that jumbo frames can provide 6x the bandwidth during this period!
If you read about the network design of AWS, they put a lot of effort into low switching latency and enabling jumbo frames.
The real pros do this kind of network tuning, everyone else wonders why they don’t get anywhere near 10 Gbps through a 10 Gbps link.
If the net were designed today it would be some complicated monstrosity where every packet was reminiscent of X.509 in terms of arcane complexity. It might even have JSON in it. It would be incredibly high overhead and we’d see tons of articles about how someone made it fast by leveraging CPU vector instructions or a GPU to parse it.
This is called Eroom’s law, or Moore’s law backwards, and it is very real. Bigger machines let programmers and designers loose to indulge their desire to make things complicated.
Many networking folks including myself consider IPv6 router advertisements and SLAAC to be inferior, in practice, to DHCPv6, and that it would be better if we’d just left IP assignment out of the spec like it was in V4. Right now we have this mess where a lot of nets prefer or require DHCPv6 but some vendors, like apparently Android, refuse to support it.
The rules about how V6 addresses are chopped up and assigned are wasteful and dumb. The entire V4 space could have been mapped onto /32 and an encapsulation protocol made to allow V4 to carry V6, providing a seamless upgrade path that does not require full upgrade of the whole core, but that would have been too logical. Every machine should get like a /96 so it can use 32 bits of space to address apps, VMs, containers, etc. As it stands we waste 64 bits of the space to make SLAAC possible, as near as I can tell. The SLAAC tail must have wagged the dog in that people thought this feature was cool enough to waste 8 bytes per packet.
The V6 header allows extension bits that are never used and blocked by most firewalls. There’s really no point in them existing since middle boxes effectively freeze the base protocol in stone.
Those are some of the big ones.
Basically all they should have done was make IPs 64 or 128 bits and left everything else alone. But I think there was a committee.
As it stands we have what we have and we should just treat V6 as IP128 and ignore the rest. I’m still in favor of the upgrade. V4 is too small, full stop. If we don’t enlarge the addresses we will completely lose end to end connectivity as a supported feature of the network.
You can just SLAAC some more addresses for whatever you want. Although hopefully you don't use more than the ~ARP~ NDP table size on your router; then things get nasty. This should be trivial for VMs, and could be made possible for containers and apps.
> The V6 header allows extension bits that are never used and blocked by most firewalls. [...] Basically all they should have done was make IPs 64 or 128 bits and left everything else alone.
This feels contradictory... IPv4 also had extension headers that were mostly unused and disallowed. V6 changed the header extension mechanism, but offers the same opportunities to try things that might work on one network but probably won't work everywhere.
UDP has its place as well, and if we have more simple and effective solutions like WireGuard’s handshake and encryption on top of it we’d be better off as an industry.
That means we can only have up to 65,534 port addresses per IP address.
they would have looked at you and asked straight out what you hoped to gain by making these things distinguished, because it certainly complicates things.
Otherwise please use the original title, unless it is misleading or linkbait; don't editorialize.
Well ... he seems very motivated. I am more skeptical.
For instance, Google via chrome controls a lot of the internet, even more so via its search engine, AI, youtube and so forth.
Even aside from this people's habits changed. In the 1990s everyone and their Grandma had a website. Nowadays ... it is a bit different. We suddenly have horrible blogging sites such as medium.com, pestering people with popups. Of course we also had popups in the 1990s, but the diversity was simply higher. Everything today is much more streamlined it seems. And top-down controlled. Look at Twitter, owned by a greedy and selfish billionaire. And the US president? Super-selfish too. We lost something here in the last some 25 years.
Personally I found the tone of the article quite genuine and the video at the end made a compelling case for it. Well I figure you commented having actually read it.
Edit: I can't downvote but if I could it probably would have been better than this comment!