Performance Guide

Sat Apr 14 20:30:31 CES 2018


What causes slow performance?

Slow performance can be caused by a variety of things from connection speeds to data encoding.

Data Serialization

Assumptions: We are going to assume we already know what the message looks like, and we’ll only look in to how to decode a string, which should illustrate the performance drawbacks of parsing data structures.

Let’s take a very simple message and see it’s faximile over the wire; here is what it would look like in everyone’s favourite data markup language:


This is what it would look like in a protocol buffers by far the most popular binary serialization format

Protocol Buffers
Protocol Buffers

So how would we decode the a string in a JSON message? (this is a super simplified, rough estimation)

JSON Decoding
JSON Decoding

rougly this is what this code is doing:

All we are doing here is reading a VERY simple ascii string and copying it from the network buffer, this example doesn’t include how to corner cases like, malformed json, handling of unicode characters, \ escape sequences and so on, as you start handling those cases the code becomes a lot more costly.

In contrast here is what decoding protocol buffers look like;

Protocol Buffers Decoding
Protocol Buffers Decoding


Let’s assume, we have a proxy that takes a message and dispatches it to another port on the same machine. We’ll modify our message so we know which port to forward our message: {"port": 80,"hello": "World!"}.

The message it self is contrived but the reality isn’t that further off from a real world scenario, change the port to host and that’s roughly how http requests are routed.


The animation above has two distinct domains called user space to kernel space. The reason why that distinction is emphasized in the animation is everytime you read(2) or write(2) the ammount of bytes you’ve processed will cross from one domain to the other.


Let’s look at the example below; first bit is the order and type, second byte 50 is the port number. After reading just 2 bytes, we can lookup which file descriptor corresponds to which port and use sendfile to proxy the rest of the message to the approprite fd without having to read the entire message. That saves the hassle of not having to copy the same message from kernelspace to userspace and back again.

Protocol Buffers Sendfile
Protocol Buffers Sendfile

Wrapping up

Binary serialized data gives us constant read times O(1), where as delimited formats like json, xml or csv you’d have to read either the entire message or a substantial portion of it.

As HTTP/2 spec puts it

Binary protocols are more efficient to parse, more compact “on the wire”, and most importantly, they are much less error-prone, compared to textual protocols like HTTP/1.x, because they often have a number of affordances to “help” with things like whitespace handling, capitalization, line endings, blank lines and so on.

Pros Cons
Binary (pb,capnproto)
  • O(1)read time
  • needs schema definition to read
Delimited (json, xml)
  • self describing
  • O(n)read time

Slow connection

let’s put some context around the Binary Serialization vs Delimited Serialization overhead; these are approximate timing for various operations on a typical PC (source:1 originally:2)

Here are some tools to simulate network latency issues -

  1. Latency Numbers Every Developer Should Know

  2. Teach Yourself Programming in Ten Years