Not on techempower benchmarks?

[ibsusu]

So I know that this project is specifically for web sockets but the speed is also very much there for normal http. Why isn't it on techempower benchmarks?

[ghost]

Because tech empower are retards.

[ibsusu]

I find this kinda funny. What you've essentially done is get uWebSockets to the point where it has as little overhead as possible versus the kernel networking stack itself but the more I think about it the more I realize that most people's clumsy attempts to utilize uWebSockets will go to waste. In my project I have to talk to the db, I have to write to disk and uws stays out of my way. It's always my crap that slows it down.

[ghost]

The nic is to the web server what the hdd is to the database; you can include them in your test but if those become the bottleneck you're not really measuring the database or the web server you're measuring hardware.

Software runs on CPU-time so if you want to isolate and measure the CPU-time overhead of the web server or database (such as algorithmic complexities, memory allocations, etc) you have to put the CPU-time as the bottleneck.

Sometimes you can do that easily, such as on the raspberry pi where CPU-time is naturally limited.

Tech empower puts their measurements on Intel gold server CPUs where no software will ever bottleneck on CPU-time so they pretty much the up measuring their Nic (which is a constant) so the only difference between the top 30 entries become random noise.

With databases you also have scheduling and packing to take into account but we're not talking about databases here.

[ghost]

Yes performance is always the weakest link of a chain. USING something fast doesn't MAKE everything fast it only unlocks the POTENTIAL to MAKE everything fast.

[jobs-git]

Well, you are free to perform the benchmark yourself to validate which framework you will use. But what they did are still useful, as it gives developers an idea, but you will still have to test in on your production.

[uNetworkingAB]

TechEmpower is useful as a negative. That is, if something scores very low in TechEmpower it is a strong hint of poor performance. But the other way around is not true.

Top-scorers in TechEmpower are almost always:

• Entirely broken in every testable sense, such as failing the majority of h1spec tests (see my github for repo)
• Often times fully hacked together with the most minimal non-compliant parser possible
• Fast for pipelined requests but very slow for normal requests
• Riddled with security flaws, undefined behavior, etc

So TechEmpower can tell you what not to pick, but it can't tell you what to pick

[jobs-git]

So I know that this project is specifically for web sockets but the speed is also very much there for normal http. Why isn't it on techempower benchmarks?

for it to be included, somebody should contribute and maintain the uwebsocket code

[uNetworkingAB]

It already is included. It scores top 3-ish. But like stated in this thread, TechEmpower is a test of shit quality and lacks the most basic credibility:

https://x.com/uNetworkingAB/status/1811425564764610926