Warp Memory leak test

This tiny repo is intended to be useful for showing memory leaks in Hyper/Warp.

Motivation

In testing out my Warp application (private), I noticed a memory leak. This lead me to look into Hyper and see #1790, which sounds like it's been closed because the issue has not been reproducable by the maintainers. Hopefully, this repository allows someone to reproduce the memory leak consistently.

Getting started

Assuming all of the prerequisites have been installed, from the top of the repo, just run:

$ ./test-me.sh

Environment variables

The following variables are used to control the application and benchmarking.

• HOST - the host name to use (default: 127.0.0.1)
• PORT - the port name to use (default: 3030)
• DURATION - the duration of the load test (default: 3 seconds)
• WORKERS - the number of simultaneous workers (default: 100)

These can be tweaked by adding a .env file in the root of the repository.

Load Testing tool

The benchmark tool uses Hey, though I suspect that any benchmarking tool could be used. More specifically, I noticed that this problem was relatively easy to produce within a very short period of time (under 5 seconds), though it becomes much more obvious over time.

For the mac, install was as simple as brew install hey.

Environment setup

I have a 2015 mac running Macos 10.15.5. I am using rustc 1.44.0 to produce the results.

I use Tokio's "full" feature so that there's nothing missing (e.g. multithreading).

Interpreting numbers

For the following settings (.env):

TRIES=1
HOST=127.0.0.1
PORT=3030
DURATION=1s
WORKERS=1

Running test-me.sh yields something like:

----------------------------------------
Settings:
----------------------------------------
Iterating 1 times
Loading 127.0.0.1:3030
Running for 1s each iteration
Using 1 workers
----------------------------------------
Server running on PID: 72930
[Warp App] Serving on: 127.0.0.1:3030
===============================
Iteration: 1 of 1
-------------------------------
Requests/sec: 5895.6067
Latency (worst case): 0.0005 secs
-------------------------------
Starting memory: 2180 KB
Ending memory: 2644 KB
Memory leaked: 464 KB
===============================

In this case, we're seeing a 464 KB increase in RAM. If we let this run for 100 iterations, some of them will appear to be 0. We're using ps here, and explicitly examining the rss (Resident Set Size) field. Because of that, the numbers are in KB rather than B. So the 0's being reported may actually be more than 0 B.