To see posts by date, check out the archives

The 14KB in the TCP Initial Window
Tyler Cipriani Posted

Today I went to coffee with a buddy of mine from $DAY_JOB - 2. During coffee we were discussing all the folk theories of “fast” websites. I made some off-the-cuff remark that I didn’t really understand – that you should try to squeeze the most important bits of your website into the first X bytes (X being a value I didn’t remember) for an opaque reason that has to do with TCP congestion algorithms that I’ve never understood.

We spent a good chunk of time looking at Wireshark captures of an HTTP request. It got muddled and we got off-track because we were looking at an HTTPS request and we started talking about TLS1.2, CLIENT-HELLO, and the TLS handshake in general.

I eventually found the source for what I was yammering about.

Due to how TCP estimates the capacity of a connection (i.e. TCP Slow Start), a new TCP connection cannot immediately use the full available bandwidth between the client and the server. Because of this, the server can send up to 10 TCP packets on a new connection (~14KB) in first roundtrip, and then it must wait for client to acknowledge this data before it can grow its congestion window and proceed to deliver more data.

After I got home and ate some ice cream, I did some more reading trying to understand this stuff. I think a lot of what was confusing about the Wireshark stuff we were looking at is it also had TLS1.2 mixed in there. Requesting a plain-old HTTP, port 80 site made looking at packets a bit more sane.

It seems like the size of a TCP packet is negotiated during the SYN, SYN-ACK phase of the TCP three-way handshake. In the Options section of the TCP segment (for a segment that has the SYN-flag set in the TCP segment header) a sender or receiver may indicate a Max Segment Size (MSS). An MSS is derived from a machine’s Maximum Transmission Unit (MTU). In the case of my machine the MTU is set to 1500. I found this via the ip command:

The MSS is derived from the MTU and is (at most?) the MTU minus the 40 bytes used for the TCP header and the IP header, so my machine sent the MSS size of 1460 to the server in the Options section of the SYN TCP segment that it initially sent my webserver.

Client MSS

So the data portion of a TCP segment sent in this particular connection should never exceed 1460 bytes.

While we were looking at Wireshark packet captures we kept noticing a TCP segment size of 1486 in the “length” column with a TCP segment data length of 1432. I noticed that my webserver sent an MSS size of 1432, which I suppose became the limiting factor in how much data could be sent per segment (rather than the 1460 MSS sent by my client – the browser):

Server MSS

Each TCP segment had a header-size of 20 bytes, so that leaves 34 unexplained bytes which (I guess) were part of the IP header.

So why try to squeeze a website into ~14KB? Seems like you should be trying to squeeze it into 1432 bytes.

I found my answer in IETF RFC 6928 – a proposal to increase the TCP Initial Window to 10 segments. The Initial Congestion Window (cwnd) is the maximum number of bytes that a server can send without receiving any acknowledgement that the bytes were received via a client ACK packet. This maximum number of bytes to send is calculated by multiplying some number by MSS. As of RFC 6928 some number is equal to 10 (initially).

I checked the value of cwnd on some established connections on my server:

And it is indeed set to 10.

This means that my server can send cwnd * MSS bytes of information before the client has to acknowledge that anything has been receieved. So, in the case of the Wireshark connection I recorded, my webserver can send 1432 * 10 bytes, or 14320 bytes:

OR around 14KB of data can be sent to a browser before another round-trip from the client to the webserver has to happen.

So, overall, a nice, relaxing coffee and catch-up session with an old coworker.