Here's what I read: they had a solution written in Node.js; they rewrote the solution in .NET Core; it was 20x faster. It's anecdotal, sure, and Microsoft prompted them to write it, and YMMV, but they did a thing and it had a remarkable result so they've remarked on it. It's not realistic to expect them to take the time to build a representative sample that doesn't expose any of their proprietary business logic; they're running a business and have more important things to do.
Anecdote of my own: I was working on a web bug that had to generate a few v4 UUIDs on every request, and using a version of Node for which the libuuid wrapper wasn't working, so I was using the fastest script-based generator I could find, but it was still too slow. A .NET Core version of the same code handled something like 40x the number of requests on the same hardware.
If nothing else, it demonstrates that using the same solution for all your different problems is A Bad Thing, because there are surely things that .NET Core is not particularly good at either.
The biggest problem with .NET Core imo is that .NET just seems to be permanently uncool, despite how good it is. Maybe one day this industry will be less fashion driven...
They aren't, but there's still a lot of irrational hate for anything even remotely associated with Microsoft, even amongst people who aren't chasing the hype. I can't really think of any other major general purpose platform that has this kind of PR problem.
This, and the AgeOfAscent piece it links to, both read like PR pieces commissioned by Microsoft. Both are absent of actual information to really determine what exactly they were testing, or what this means to anyone else's problem space.
Because, of course, when you get to the fun of benchmarks there's always a faster options. With .NET Core offering them 20,000 requests per second (on a C3.large), that is a terribly low bar to hit. Again, maybe they're doing something amazing, but many frameworks have rates in the seven digits on that sort of hardware. And I know .NET Core can process basic requests in the six digits, so even it is hardly the limiting point, and it comes to the logic.
With respect to the TechEmpower benchmarks, if you look at Round 13 [1] they note that ASP.NET got 859 times faster in their benchmarks since Microsoft dedicated effort to improving it.
Indeed (though it's worth noting that they're talking about the performance of ASP.NET Core versus Mono, which was notoriously disastrous performance wise, and that story got somewhat misrepresented). In no way am I saying that the .NET Core is slow.
But if we want to use that benchmark, doing the most trivial useful thing of all -- serializing a simple object to JSON -- and node.js beat core .net. In the recent iteration it beat it by almost 2x (which makes the "node.js is slow we all know that" bit in the linked piece humorous given that it beats aspnetcore in every test but plaintext, which is an irrelevant test anyways given that they put their service behind nginx).
I don't use node.js. I don't advocate it. I personally think most node.js solutions end up being a spaghetti mess. But these sorts of "the details are hidden but look at the magic we wrought!" benchmarks are worse than useless. They're fools gold for people choosing platforms and thinking this sort of advocacy is guidance to learn from.
Duh, yes, of course the logic makes it much slower than copying a L1 cached static response to the client. But then that is not a use case anyone cares very much about.
The AgeOfAscent story is not a PR piece, the writer (Ben Adams) is a very active contributor to the Kestrel server behind .NET Core. If he's writing about improved performance it's probably because he wrote >1/3rd of the patches.
What a weirdly trite response given that you're essentially repeating what I said. And without specifics the linked piece, and its claims about benchmarks, is utterly meaningless.
"Rewrote inefficient code. Now it's faster. Story at 11!"
The AgeOfAscent story is not a PR piece
Humorously it was likely a "copy a L1 cached static response" type benchmark.
Okay, so it wasn't PR, it was self-aggrandizement (which is effectively PR). Got it. Though in this article it was linked as a performance improvement of "switching" to .NET Core, when really it was a story of terribly inefficient .NET Core code becoming better, though again in a nutshell it is meaningless. Cool.
yea that is caveat for comparing numbers in a simple http payload benchmark -- doesn't really reflect real workloads. I don't think it's meaningless that .NET got better. Perhaps more competition is better, otherwise we'd all be using Java/netty. However, I also see the repetitive work in having different stacks.
> when we started to look at .NET Core in early 2016, it became quite obvious that being able to asynchronously hand off to our queuing service greatly improved throughput. Unfortunately, at the time, Node.js didn’t provide an easy mechanism to do this, while .NET Core had great concurrency capabilities from day one.
This is a mildly maddening article - not because I have any emotional attachment to either platform - but because it never really gets to the heart of the matter. It's like they can tell us, but they'd have to kill us afterwards.
A 20X improvement cannot be attributed to .NET core, it must attributed substantially to some implementation/algorithmic issue that they resolved once they got to the .NET core platform. There's a big problem somewhere with node.js if it is 20X slower. If true, then its simply a bug or series of bugs to be fixed in node.js.
Actually 3rd party libraries and application code is where .NET is likely to be faster unless node lib is wrapping a C library - .NET can easily beat JS an order of magnitude simply by having control of memory layouts and high performance/specialized data structures. Meanwhile JS has crap structures (doesn't even have primitive arrays) and everything is it's dynamic object where you pray the JIT eventually figures out a constant structure and tries to optimize.
> But removing any 3rd party libraries, and external requests. Generally .NET core is faster.
Is that with reference to "faster than .Net regular" or "faster than node.js"?
You're right either way but I'm not sure that "it's just faster" is all that relevant. The "Just a basic request/response with barely anything thing in it" case removes anything of interest.
For any normal non-trivial, non-demo setup, the web server is fine. I would look at the performance of data stores and other backends. The performance of "asynchronously hand off to our queuing service" is most likely the overriding factor in a 20 or 30x increase.
Honestly, add another machine and the problem is solved. youll be able to afford a team of programmers to rewrite your app in any language of your choosing by the time you feel the slow down.
I'm using .NET core for a production app now. Team of 16. It's working beautifully. The data access is the bottleneck in 99% of what I have ever worked on, not the web framework. That is why Rails is so popular.
I'm a .NET fan, but it sounds strongly like these guys knew what they were doing on .NET and not so much on node. It's amazing how many of these rewrite stories actually boil down to "and the new version didn't do some incredibly dumb things the old one did".
If you really wanted throughput and large numbers of simultaneous connections, you'd be looking at Erlang, anyway.
The fact that they were only getting 1k reqs/sec with Node gives me concern. It clearly shows something went wrong there very early on at a very fundamental level. By no means is Node the end-all be-all for performance by any measure, but you should definitely be getting much much higher throughput than 1k reqs/sec.
Simply booting up a single core http server should net you around 4-5k requests per second. Spin up an instance per core and you should be _at least_ in the 10k realm.
>Simply booting up a single core http server should net you around 4-5k requests per second.
There's no such thing as "requests per second" generally. It's requests per second for a specific workload.
So, whether Node can do 4-5k rps with "hello world" doesn't matter much. It's the same engine that needs to also do the further processing for each fuller request.
> The fact that they were only getting 1k reqs/sec with Node gives me concern. It clearly shows something went wrong there very early on at a very fundamental level.
Because the program is performing CPU compute for each request. Node is single threaded.
It's a shame that the Node project killed off the multithreaded web workers pull request. It sorely needs that functionality. Pools of node processes are a poor substitute.
Looking at my Prometheus stats right now, I've got a Node TCP server doing 12k concurrent connections with 3-5% CPU and 130 MB memory. At least 4 DB queries are done per request, sometimes 10 queries. Raygun either has some nonobvious stuff going on, or Microsoft paid them to write PR fluff. After 6 years with Node it seems odd for their performance to be so bad. They should dive into more technical details to explain what specifically was made faster.
You've described a classic I/O bound server application (waiting on the database) that Node handles well. As soon as you introduce serious computations or server side rendering into the request handling performance would fall dramatically.
Care to elaborate on why additional node processes are a poor substitute? They are really easy to reason about, and to manage. What are the downsides? Not trolling, really interested in hearing about pros/cons.
One obvious thing is that multiple processes means no sharing of any kind of state or in-memory cache, so you immediately have to go to an external cache like Redis or whatever, with the additional maintenance and minor performance hit.
As the sibling comment points out, your middleware that has a 30k req/sec hello world may slow down to 500 req/sec on a production server once it's competing with everything else on the thread.
Obviously they did something terribly wrong in the first place and now are showing off being all right. Netty would probably faster. As there're no details at all, I feel free to just guess :-).
I recently found Node much faster then .NET Core for a very specific scenario:
where every connection must be TLS. So, Node / ASP.NET Core have to decrypt traffic from Elasticsearch and encrypt to NGINX. For a minimal workload the whole trip took 50 ms with .NET Core and 20 ms with Node. Obviously there's something wrong with the .NET setup - maybe some setting with Kestrel and TLS.
Anyway, that's my anecdote. And I wouldn't dare writing a block post how node is 2,5 times faster than .NET.
Has anyone any code they can show which performs the same job on both and is close to comparing apples to apples? It'd be interesting to me, not for performance comparison, but as a real world application structure comparison, a kind of more advanced Rosetta Code[1] example.
It looks like they ran their last node.js benchmark around Sep 2015. I'm curious if this was still with node.js v0.x, or if they had already adopted the latest version of io.js (3 or maybe 4).
There is no technical information in this article. As a result it reads like Microsoft marketing and is off-putting to .NET Core, something I'd otherwise be interested in.
> From the questions we received around the specifics of our performance improvements, there seems to be two schools of thought:
> 1. Of course it’s faster, Node is slow
> 2. You must be doing Node wrong, it can be fast
Allow me to offer a third :)
When you rewrite something, it had better be faster! You know a lot more about how it works and how it's used than you knew at the outset. There was a recent post, about going back to Ruby after creating the first version in Clojure, that touches on this point.
The article is an excellent example of sneaky PR practices. Not only it lacks the technical details, it deceives the masses with insane amounts of artificial hype. "How AstroTurf invented astroturfing" would be a better title for it.
Anecdote of my own: I was working on a web bug that had to generate a few v4 UUIDs on every request, and using a version of Node for which the libuuid wrapper wasn't working, so I was using the fastest script-based generator I could find, but it was still too slow. A .NET Core version of the same code handled something like 40x the number of requests on the same hardware.
If nothing else, it demonstrates that using the same solution for all your different problems is A Bad Thing, because there are surely things that .NET Core is not particularly good at either.