[Router] Replace httpx with aiohttp in vllm_router for enhanced high-concurrency performance

[ikaadil]

FILL IN THE PR DESCRIPTION HERE

FIX #569

BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE

---

• Make sure the code changes pass the pre-commit checks.
• Sign-off your commit by using -s when doing git commit
• Try to classify PRs for easy understanding of the type of changes, such as [Bugfix], [Feat], and [CI].

Detailed Checklist (Click to Expand)

Thank you for your contribution to production-stack! Before submitting the pull request, please ensure the PR meets the following criteria. This helps us maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Please try to classify PRs for easy understanding of the type of changes. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

• [Bugfix] for bug fixes.
• [CI/Build] for build or continuous integration improvements.
• [Doc] for documentation fixes and improvements.
• [Feat] for new features in the cluster (e.g., autoscaling, disaggregated prefill, etc.).
• [Router] for changes to the vllm_router (e.g., routing algorithm, router observability, etc.).
• [Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

• Pass all linter checks. Please use pre-commit to format your code. See README.md for installation.
• The code need to be well-documented to ensure future contributors can easily understand the code.
• Please include sufficient tests to ensure the change is stay correct and robust. This includes both unit tests and integration tests.

DCO and Signed-off-by

When contributing changes to this project, you must agree to the DCO. Commits must include a Signed-off-by: header which certifies agreement with the terms of the DCO.

Using -s with git commit will automatically add this header.

What to Expect for the Reviews

We aim to address all PRs in a timely manner. If no one reviews your PR within 5 days, please @-mention one of YuhanLiu11
, Shaoting-Feng or ApostaC.

[gemini-code-assist]

Summary of Changes

Hello @ikaadil, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request initiates a significant performance upgrade for the vllm_router. By replacing httpx with aiohttp, the router is expected to achieve superior performance under high-concurrency workloads, leveraging aiohttp's asynchronous capabilities. The initial step involves updating the project dependencies to include aiohttp.

Highlights

• Performance Enhancement: The primary objective of this pull request is to enhance the high-concurrency performance of the vllm_router by transitioning its underlying HTTP client from httpx to aiohttp.
• Dependency Update: The aiohttp library (version 3.8.0 or newer) has been added to the src/vllm_router/requirements.txt file, which is a prerequisite for the planned client migration.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request adds aiohttp as a dependency. To fully leverage the performance enhancements, ensure httpx is replaced with aiohttp throughout the codebase, and that the aiohttp session lifecycle is properly managed. Also, remember to remove httpx if it's no longer needed and update the sentry-sdk dependency.

[kobe0938]

After reading the https://oxylabs.io/blog/httpx-vs-requests-vs-aiohttp from #569, it seems that AIOHTTP is indeed faster than HTTPX under asynchronous programming in the benchmarking results.

However, can you give an simple benchmarking for the production stack before and after replacing the httpx library? I suppose you may have a good scenario/ use case for making this replacement. Thanks!

[ikaadil]

After reading the https://oxylabs.io/blog/httpx-vs-requests-vs-aiohttp from #569, it seems that AIOHTTP is indeed faster than HTTPX under asynchronous programming in the benchmarking results.

However, can you give an simple benchmarking for the production stack before and after replacing the httpx library? I suppose you may have a good scenario/ use case for making this replacement. Thanks!

@kobe0938 I ran a simple benchmark by sending 10,000 concurrent requests to compare performance before and after switching from httpx to aiohttp.

🧪 Benchmark Results:

• Before (using httpx): 28.93 seconds
• After (using aiohttp): 6.70 seconds

The improvement was quite significant under asynchronous workloads.

You can find all the details and steps to reproduce the benchmark here.

[ikaadil]

@ikaadil Thanks a lot this is awesome! Can you fix the pre-commit issue and the router E2E tests?

I've added some changes to ensure the pre-commit checks pass. However, I'm unable to run the E2E router tests locally. Are there any instructions available for running the full E2E suite locally(without GPU)?

[YuhanLiu11]

@ikaadil You can follow the commands listed in https://github.com/vllm-project/production-stack/blob/main/.github/workflows/router-e2e-test.yml to run the e2e tests locally. but you will need to have GPU on it to run.

[ikaadil]

@YuhanLiu11 All tests are green. Nothing is broken. The universe is briefly in balance. Ready for merge.

[ikaadil]

@YuhanLiu11 The PR passed all the test cases. Could you please review it?

[YuhanLiu11]

@ikaadil Thanks.
@kobe0938 Can you run this PR with your router stress tests?

[YuhanLiu11]

@ikaadil btw this is a pretty large change in our router core logic, and we'll need to do more comprehensive testing before we merge this PR on our side.

[ikaadil]

@ikaadil btw this is a pretty large change in our router core logic, and we'll need to do more comprehensive testing before we merge this PR on our side.

Thanks @YuhanLiu11, that makes sense. I’m happy to support with any additional testing or fixes needed. Let me know if I can help.

[kobe0938]

@YuhanLiu11

Sure. I conducted some experiment with the router stress-testing code from https://github.com/vllm-project/production-stack/pull/562/files#diff-e83461fc0a9738855618c627d68d8b77941359fe60a4feb06f93e4390377d3ef

Here's the steps:

source /home/ie-user/kobe/ps/bin/activate

current version; hash (1da26b6)

git checkout main
pip install -e .
bash tests/e2e/stress-test.sh

results:

# 200/1000 -> 2.729s
# 2000/10000 -> 54.648s
# 20000/20000 -> (test time out)

proposed version(#589)

git checkout upstream-pr-589
pip install -e .
bash tests/e2e/stress-test.sh

# 200/1000 -> 2.817s
# 2000/10000 -> 27.220s
# 20000/20000 -> 71.398s

We can see that, indeed the aiohttp library will improve the performance during high-current workloads than httpx 2x during 2000 concurrent requests(from 54.648s to 27.220s).

Here's one example of logs

((ps) ) ie-user@g0361:~/kobe/production-stack$ bash tests/e2e/stress-test.sh
[INFO] Checking prerequisites...
[INFO] Router stress test configuration:
[INFO]   Concurrent requests: 2000
[INFO]   Total requests: 10000
[INFO]   Router port: 30080
[INFO]   Backend ports: 8001, 8002
[INFO]   Model: facebook/opt-125m
[INFO] Starting router with round-robin routing (stress test mode)
[INFO] Router started with PID: 1307668
[INFO] Waiting for router to be ready...
[INFO] Router is ready
[INFO] Running stress test with Apache Bench
[INFO] Concurrent: 2000, Total: 10000
This is ApacheBench, Version 2.3 <$Revision: 1879490 $>
Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
Licensed to The Apache Software Foundation, http://www.apache.org/

Benchmarking localhost (be patient)
Completed 1000 requests
Completed 2000 requests
Completed 3000 requests
Completed 4000 requests
Completed 5000 requests
Completed 6000 requests
Completed 7000 requests
Completed 8000 requests
Completed 9000 requests
Completed 10000 requests
Finished 10000 requests


Server Software:        uvicorn
Server Hostname:        localhost
Server Port:            30080

Document Path:          /v1/chat/completions
Document Length:        21 bytes

Concurrency Level:      2000
Time taken for tests:   54.648 seconds
Complete requests:      10000
Failed requests:        0
Non-2xx responses:      10000
Total transferred:      1930000 bytes
Total body sent:        3920000
HTML transferred:       210000 bytes
Requests per second:    182.99 [#/sec] (mean)
Time per request:       10929.546 [ms] (mean)
Time per request:       5.465 [ms] (mean, across all concurrent requests)
Transfer rate:          34.49 [Kbytes/sec] received
                        70.05 kb/s sent
                        104.54 kb/s total

Connection Times (ms)
              min  mean[+/-sd] median   max
Connect:        0   14  18.0      4      63
Processing:   118 9322 3654.3   8204   18354
Waiting:       25 8933 3648.5   7785   17623
Total:        118 9336 3646.5   8239   18357

Percentage of the requests served within a certain time (ms)
  50%   8239
  66%   9501
  75%  10511
  80%  11791
  90%  16048
  95%  16759
  98%  17191
  99%  17494
 100%  18357 (longest request)
[INFO] Stress test completed
[INFO] Checking round-robin routing correctness...
[INFO] Round-robin routing results:
[INFO]   Backend localhost:8001: 5000 requests
[INFO]   Backend localhost:8002: 5000 requests
[INFO]   Total routed: 10000 requests
[INFO]   Backend localhost:8001: 50%
[INFO]   Backend localhost:8002: 50%
[INFO] ✅ Round-robin routing is working correctly (0% difference)
[INFO] Test completed successfully!
[INFO] Cleaning up router processes...