about

measure how long diff llm clients block on io for sync vs streaming calls

usage

on devbox:

./worker_blocking_analysis.sh ruby <test_script>

<test_script>:

• openai-stream.rb
• openai-sync.rb
• stripe-stream-poll.rb

methodology

run the prompt Write a haiku about OpenAI. on gpt-5 three times in series.

measure wall clock time and subtract user+sys cpu time to calculate blocked time on network i/o.

validate blocked time using strace being spent on epoll_wait and futex.

clients and calls

clients:

• stripe developerai ruby llm client code
• openai ruby sdk link

calls:

• sync
• native streaming
• poll-based streaming

results

Client	Sync Calls (blocked on i/o)	Streaming Calls (blocked on i/o)
OpenAI Native SDK	95.09%	97.36%
Stripe Ruby Client (Poll)	N/A	79.19%