cowboy_post_bench

POST benchmark to compare between cowboy-1.1.2 and cowboy-2.10.0.

how to run

make

how it works

test unit

• Prepare a binary file of a specific size filled with zeros.
  • Note: It does not consume storage because it is actually a hole file.
• Upload this file to the local cowboy HTTP/1.1 server via loopback device using local curl client.
  • The cowboy HTTP/1.1 server measures the time to read the request body and simply discards the body.
• Repeat 10 times.

test1

• Repeat the test unit in various binary sizes and cowboy1/2.

test2

• Repeat the test unit in various active_n and length options setting with cowboy2 and 1Gbytes binary.

test3

• Repeat the test unit in various socket setopts buffer sizes with patched cowboy2 and 1Gbytes binary.

Tested platforms

• Xubuntu 16.04
• OTP-21.3.8.18
• OPT-26.0

Result summary

• OTP-21
• OTP-26 without JIT
• OTP-26 with JIT

Findings

• Cowboy2 is about 10 times (without JIT) or 7.6 times (with JIT) slower than cowboy1 for uploading 1 1Gbyte file.
• Changing active_n or length has too little effect to solve the problem.
• Changing buffer has large effect, but still slightly slower than Cowboy1.