This isn’t about understanding everything about TCP or reading through TCP/IP Illustrated. It’s about how a little bit of TCP knowledge is essential. Here’s why.
When I was at the Recurse Center, I wrote a TCP stack in Python (and wrote about what happens if you write a TCP stack in Python). This was a fun learning experience, and I thought that was all.
A year later, at work, someone mentioned on Slack “hey I’m publishing messages to NSQ and it’s taking 40ms each time”. I’d already been thinking about this problem on and off for a week, and hadn’t gotten anywhere.
A little background: NSQ is a queue that you send to messages to. The way you publish a message is to make an HTTP request on localhost. It really should not take 40 milliseconds to send a HTTP request to localhost. Something was terribly wrong. The NSQ daemon wasn’t under high CPU load, it wasn’t using a lot of memory, it didn’t seem to be a garbage collection pause. Help.
Then I remembered an article I’d read a week before called In search of performance - how we shaved 200ms off every POST request. In that article, they talk about why every one of their POST requests were taking 200 extra milliseconds. That’s.. weird. Here’s the key paragraph from the post
Delayed ACKs & TCP_NODELAY
Ruby’s Net::HTTP splits POST requests across two TCP packets - one for the headers, and another for the body. curl, by contrast, combines the two if they’ll fit in a single packet. To make things worse, Net::HTTP doesn’t set TCP_NODELAY on the TCP socket it opens, so it waits for acknowledgement of the first packet before sending the second. This behaviour is a consequence of Nagle’s algorithm.
Moving to the other end of the connection, HAProxy has to choose how to acknowledge those two packets. In version 1.4.18 (the one we were using), it opted to use TCP delayed acknowledgement. Delayed acknowledgement interacts badly with Nagle’s algorithm, and causes the request to pause until the server reaches its delayed acknowledgement timeout..
Let’s unpack what this paragraph is saying.
- TCP is an algorithm where you send data in packets
- Their HTTP library was sending POST requests in 2 small packets
Here’s what the rest of the TCP exchange looked like after that:
application: hi! Here’s packet 1.
HAProxy: <silence, waiting for the second packet>
HAProxy: <well I’ll ack eventually but nbd>
application: <well I’m waiting for an ACK maybe there’s network congestion>
HAProxy: ok i’m bored. here’s an ack
application: great here’s the second packet!!!!
HAProxy: sweet. we’re done here
That period where the application and HAProxy are both passive-aggressively waiting for the other to send information? That’s the extra 200ms. The application is doing it because of Nagle’s algorithm, and HAProxy because of delayed ACKs.
Delayed ACKs happen, as far as I understand, by default on every Linux system. So this isn’t an edge case or an anomaly – if you send your data in more than 1 TCP packet, it can happen to you.
in which we become wizards
So I read this article, and forgot about it. But I was stewing about my extra 40ms, and then I remembered.
And I thought – that can’t be my problem, can it? can it??? And I sent an email to my team saying “I think I might be crazy but this might be a TCP problem”.
So I committed a change turning on
TCP_NODELAY for our application, and BOOM.
All of the 40ms delays instantly disappeared. Everything was fixed. I was a wizard.
should we stop using delayed ACKs entirely
The real problem is ACK delays. The 200ms “ACK delay” timer is a bad idea that someone at Berkeley stuck into BSD around 1985 because they didn’t really understand the problem. A delayed ACK is a bet that there will be a reply from the application level within 200ms. TCP continues to use delayed ACKs even if it’s losing that bet every time.
He goes on to comment that ACKs are small and inexpensive, and that the problems caused in practice by delayed ACKs are probably much worse than the problems they solve.
you can’t fix TCP problems without understanding TCP
I used to think that TCP was really low-level and that I did not need to understand it. Which is mostly true! But sometimes in real life you have a bug and that bug is because of something in the TCP algorithm. So it turns out that understanding TCP is important. (as we frequently discuss on this blog, this turns out to be true for a lot of things, like, system calls & operating systems :) :))
This delayed ACKs / TCP_NODELAY interaction is particularly bad – it could affect anyone writing code that makes HTTP requests, in any programming language. You don’t have to be a systems programming wizard to run into this. Understanding a tiny bit about how TCP worked really helped me work through this and recognize that that thing the blog post was describing also might be my problem. I also used strace, though. strace forever.