Hi there, great work writing oat++. I am the author of Passenger (https://www.phusionpassenger.com), an application server that powers 650.000 companies and sites world-wide such as Apple, Pixar, Juniper, etc. Passenger is mostly written in C++ and like you, performance is a priority for us.

I've done a fast code review, and I thought I'd share some of my findings with you. These are based on my experience with writing high-performance servers. You can read about some of my experience in these blog posts: https://www.rubyraptor.org/how-we-made-raptor-up-to-4x-faste...

Hopefully you can use my feedback to make oat++ even faster and better. Code review here: https://github.com/oatpp/oatpp/issues/15

I've covered:

- Pools and contention

- Shared_ptr and atomics everywhere

- List vs vector

- Zero-copy architecture

- SpinLock everywhere

- New does not return nullptr?

- More documentation

- More tests

I would just like to say that as someone who has never used Passenger, never used Oat++, and hopes to god he never has to write a line of C++ again this comment (and the linked code review) really warmed my heart. Not only is it a positive/helpful comment in a world with so much focus on tearing down and picking apart but it's all of that from the author of a competing framework. You go even further to review the code and suggest ways to make your "competitor"'s (I use the word lightly in this context but "framework wars" often take on an almost religious zeal) product better.

I will probably never use either of your frameworks directly but I'd like to just say thank you FooBarWidget for this, it made me smile.

Thanks, this is a really good code-review! I appreciate this!

Good point on the shared_pointers. What's your opinion on using Ravenbrook MPS pools and not deal with refcounts at all?

Memory management, and pools.

Memory management is definitely super important and has big impact on the performance. Also there is a memory fragmentation issue that you should keep in mind building server apps. Solving this problems is time consuming and you need to do a lot of experiments in order to tune all this system. So if there is a robust time-proved solution for solving this problems it is definitely worth of considering.

shared_ptr and refcounts - actually it is a compromise (like any other solution). On the one hand you could avoid refcounts, atomics etc. On the other hand you have a framework for the end user. And user may be not so familiar in depth of your framework's thread management and concurrent access to some shared objects.

Also end-user may decide to bring in some other library which for example uses shared_ptr. And what user should do if he has to share some object among std::shared_ptr and your own custom shared-ptr:)?

This is the reason I decided to go with std::shared_ptr (with it's atomic refcounter) as I mentioned in the code-review in the response https://github.com/oatpp/oatpp/issues/15

Now about Ravenbrook MPS pools in particular: It seems to be a robust time-proved solution which may help you to avoid a lot of pain. As for oatpp - oatpp claims to be zero-dependency at the moment. So no external dependencies are considered at the moment. This may be changed at some moment in the future but not for now.

I have no experience with Ravenbrook MPS. But from the documentation it appears to contain both tracing garbage collectors (automatic memory management) as well as manual memory managers.

Not dealing with refcounts implies using a tracing garbage collector. There is a wide range of literature out there about such a comparison, but in general I cannot compare the two at all because of the vast fundamental differences.

The idea of the project is to give user something light, something that could cover a set of basic needs when you develop a web-service (Like rest-framework, basic DI-framework, web-client, connection management, object-mapping etc...) and make it highly customizable at the same time.

So when you need to kickoff something from scratch you don't need to do anything more than a git clone starter-project and start coding. And when it comes time for additional requirements you cant easily substitute ex. any SSL backend, any http-client, add http compression etc... And oat++ provides interfaces to configure there things.

See more: Example how to create web-service with swagger-uI and auto-documented endpoints https://medium.com/oatpp/c-oatpp-web-service-with-swagger-ui...

What is the rationale behind going for from-full-scratch instead of packaging a copy of ASIO with the library? Beating properly-used ASIO performance is going to be a massive uphill battle, especially in multithreaded scenarios.

On that note, benchmarks against a simple ASIO http server are a must if performance is your #1 stated goal.

I agree. I should schedule to add ASIO vs oatpp benchmark.

However performance is not the only thing oatpp has. Among with other features oatpp provides ObjectMapping layer. ObjectMapping layer enables you to do cool stuff like: - autodocument endpoints see https://medium.com/oatpp/c-oatpp-web-service-with-swagger-ui...

- easily implement custom protocols (I will write article about this later)

I think you are misunderstanding my question. I am not suggesting that ASIO fills the role your library does. All it is, really, is just fancied-up cross-patform select().

What I'm wondering is: Is there's a specific reason you chose to reinvent that specific wheel?

Oh ok, thank you for clarification.

During the initial investigation boost::asio appeared to work not very good on MacOS. Problems where on the load higher than 10K concurrent connections.

There was multiple requests to compare networking performance of oatpp to ASIO lately. So I want to make full-scale testing before making statements.

+1 to this. If you are faster than ASIO - consider contributing back. If you are slower, and performance is the goal, what's the point of using your framework?

In case any other ignorant amateurs like me are wondering what an audio driver layer has to do with web services: it turns out ASIO is also a C++ library for network programming (https://think-async.com/).

To elaborate, ASIO is arguably the de-facto standard way to do low-level networking in C++, to the point where it forms the basis for the standard library's upcoming networking library.

> What is the rationale behind going for from-full-scratch instead of packaging a copy of ASIO with the library?

Is there a way of building ASIO without Boost?

Yes, there's a standalone version of ASIO.

Have you noticed oatpp ApiClient (retrofit-like client wrapper)? This is another trick oatpp's ObjectMapper layer can do.

The oatpp-consul client is built using oatpp ApiClient. see ApiClient in the oatpp-consul https://github.com/oatpp/oatpp-consul/blob/master/rest/Clien...

example how to do client requests https://github.com/oatpp/oatpp-examples/issues/3

It is also possible to build your own RequestExecutor based on ex. Curl

Looks pretty cool and easy to use, at least for the synchronous methods. The async methods could use a little bit more documentation, e.g. what is really async about them (what is Action? Some kind of deferred response? Can you store the request objects there for longer than the act() methods scope? Etc).

What I would also like to see there are additional examples which show how streaming the request/response body is done, e.g. for file uploads and downloads or SSE.

Hey, thank you for the feedback! It is really very important to me what seems unclear to user, and what info to add.

I wonder how this stacks against Cutelyst in terms of performance.

Hello and thank you for the question!

The short answer: - I don't know. - Let Techempower to settle this.

I decided to take on Go(net/http) in the initial benchmarks as it shows good performance and even overperformed c++ solutions that claimed to be more performant (in their tests).

But! There multiple things that can be tuned especially with c++ servers. And there are multiple tests which show different aspects of server's performance. The problem is that currently I have no resources to do all these tests and what is more important - I think such tests should be made by an independent third party.

So I think oat++ should be submitted to Techempower to answer all those questions.

You are expected to deliver tests to Techempower yourself. Optimizing for those tests is part of the game. ;)

net/http is known for being quite slow

Yeah, I already looked through requirements. It will take me some time to prepare all that stuff. But I think it's gonna be fun

Did you notice that https://oatpp.io is served by oatpp :)?

I have a question. I want to adopt oatpp.io, but I'm using a synchronous database driver mongo-cxx-driver. Can I still benefit from oat++'s coroutine? Also, when using coroutine, will the server be a single-thread process? or it will be a multiple thread process with each thread using coroutines (to fully utilize the cpu)?

If my server wants to visit third party rest APIs in an asynchronized way, what's the best asynchronized library to choose to make use of coroutine?

Thank you for the question.

-----------------------------------------------------------------------------

// I want to adopt oatpp.io, but I'm using a synchronous database driver mongo-cxx-driver.

- Unfortunately no - if your driver has only synchronous API you can use it with multithreaded oatpp only.

-----------------------------------------------------------------------------

// Also, when using coroutine, will the server be a single-thread process?

// or it will be a multiple thread process with each thread using coroutines (to fully utilize the cpu)

- Yes, it will be a multiple thread process with each thread using coroutines. You can configure how many threads should run the AsyncHttpConnectionHandler. Number of threads is passed to the AsyncHttpConnectionHandler constructor. Also you may tweak it in compile time via option:

"-D OATPP_ASYNC_HTTP_CONNECTION_HANDLER_THREAD_NUM_DEFAULT=<num of threads>"

You have to play with this parameter in order to achieve best performance. In most cases on : - Linux it's value should be (<num of CPUs> - 1). - On MacOS it's value should be 2.

-----------------------------------------------------------------------------

// If my server wants to visit third party rest APIs in an asynchronous way,

// what's the best asynchronized library to choose to make use of coroutine?

- Native way to do it is to use oatpp ApiClient https://oatpp.io/docs/component/api-client

it provides two declarations:

- API_CALL("GET", "/resource", getResource) ///< Example of API call

- API_CALL_ASYNC("GET", "/resource", getResourceAsync) ///< Example of Async API call

Also there is an example project with the full example how to make asynchronous calls to rest APIs https://github.com/oatpp/oatpp-examples/tree/master/tls-libr...

There is no README currently in this example, but I gonna add asap.

On a corporate proxied network I'm getting this response

    server=oatpp/0.18.9
    code=431
    description=Request Header Fields Too Large
    message= Can't parse headers

It seems to work okay on the non-corporate network

There is a limitation on the header size in the oatpp html parser.

Think I should tweak this.

Thank you for the note!

Perhaps my NVidia tablet is out of date but that link comes up secure connection failed under Firefox Focus. I have not done deep dive to study but not a common issue for me with focus. (Edit: might be intermediate cerificates issue.)

Just a small thing I found: your site seems to break scrolling on iOS.

Yeah known issue, I am not a JS master:(

Have you tried using (or building on top of) seastar-framework ?

Thank you for the question,

No, I have not tried seastar. But thank you for the tip, I think there might be things for me to point out.

> return createResponse(Status::CODE_200, "Hello World!");

Is "Status::CODE_200" a useful abstraction over just "200"?

Assuming that's an enum, the compiler will tell you if you miss one case in a switch statement. Also if you mistype 200 for example as 20 or 2000 the compiler will also tell you. None are huge reasons, but it is nice and cheap.

good point

An HTTP status has both a numerical code (200) and a textual reason (OK).

Is this cross platform? It looks like there is direct dependency on Linux socket APIs.

Hello, Thank you for the question!

oatpp has been recently tested on: - Linux (Ubuntu 16, 18. CentOS 7.5) - MacOS

Windows support is not currently available. But I would love to add it. If there someone with windows machine and willing to help, I will provide all needed assistance. Windows support issue: https://github.com/oatpp/oatpp/issues/2

What about TLS 1.3?

Thank you for the question,

Oat++ provides interfaces to substitute SSL backend. One can easily substitute any SSL backend by implementing two class: - Connection - ConnectionProvider

Currently there is an adapter for LibreSSL https://github.com/oatpp/oatpp-libressl/tree/master/server

But if you need TLS 1.3 one may add GnuTLS adapter for example