http-stress-test

README

Problem

black box testing against live systems can cause high load on the tested system, resulting in loss of availability.

It is often assumed that the raw number of request sent to a system correlates to the load (CPU, I/O wait) on this system over time. While this may or may not be true, limiting the number of requests sent per unit of time to a system can result in limiting the overall load on that system. Limiting the number of requests sent per unit of time will also increase scan time, which can result in decreased performance at the receiving end of the scan over a longer period of time than a faster scan would, and can also affect the test results negatively (missed tests, &c).

One of the ways to try and achieve a proper send rate is to assume that there's a dependence between the number of requests sent per unit of time, and response times for those requests. The most common assumption is that an increase in the number of sent requests will lead to an increase in response time. Based on this assumption, some tools may try to adapt the send rate based on response time measurements.

We will explore that assumption in the environment of web application scanning tools used over the Internet.

http-stress-test.go test scenario

$ ./http-stress-test -h
Usage of http-stress-test:
  -body="": request body
  -btype="application/x-www-form-urlencoded": body type (if body is set)
  -duration=3: send duration (s)
  -method="GET": HTTP method
  -rate=50: send rate (req/s)
  -timeout=20: request timeout (s)
  -url="": URL

We're using Go's net/http DefaultClient as a reference client with no extra options. The complete response body is read. DefaultClient supports compression and keep-alive functionality. This is reasonable functionality to have in an HTTP based scanner.

Difference in average response times between requests to different resources is expected. foo.php might execute more code and perform more DB interaction server side that bar.php. Therefor, we limit our tests to one resource.

The tests are performed over the Internet, no WiFi, against a server with no other significant traffic than the test traffic itself.

http-stress-test.go measurements