|
|
|
|
|
|
by jpgvm
1669 days ago
|
|
|
I think both have different pros. Java has higher peak JIT performance though after Rosalyn was introduced C# has made strides at closing that gap on many metrics.
GC algorithms are more advanced, namely ZGC and Shenandoah. I would elaborate but I don't want to turn this post into an actual essay.
JVM is more configurable. This is a double edged sword, it heavily favors users that have spent the time to really understand the platform but it has created a reputation of the JVM being hard to operate.
JVM class loaders and the new(ish) Jigzaw module system are very powerful, especially for shipping customized/cut-down JVM + stdlib when you need to do distribution.
JVM has better portability. .NET Core has obviously changed the game but other platforms are still treated second class to Windows in many ways.
GraalVM is reviving the guest language ecosystem of the JVM. Clojure, Scala and Kotlin have all remained reasonably strong but there was a dark period there for Jython/JRuby and friends. GraalVM is changing the tide here and I have already been able to leverage this over Nashorn to embed essentially native speed JS code into one of my Java programs. CLR has better primitive types, JVM's project Vahalla may resolve this.
The CLR/.NET tends to introduce features at a quicker pace than the JVM/Java, this I feel is CLR's double edged sword. Async/await being the key example. It's been in the CLR for what feels like forever now vs JVM which is only now just getting close to landing Loom. However I feel like Loom is an infinitely better solution to the problem that leverages the unique advantages of JIT compiled VM languages that rely on relatively little unmanaged code.
The experience on the CLR is much more integrated assuming you are on the well trodden path. i.e VS + Windows 10.
MSIL is much much more pleasant than JVM bytecode in my experience and the community developed tooling for working with it is great.
Optimizing programs on CLR is easier than JVM, it takes comparatively less effort to write allocation free code and get great memory layout using primitive arrays etc for high performance code.
CLR interacts better with native code. JNI is poor in comparison. Project Panama could make this better on the JVM side. This is by no means a complete list of ways the two differ, this is just my experiences having written a decent chunk of code for both. FWIW I prefer the JVM but if I am forced not to use the JVM the CLR is the very next thing I propose. |
|
|
Not really. Maybe a tiny little bit, but not really. Microsoft has discovered that the "other platforms" are very important, because they want to sell Azure, and people there will use linux predominantly. Same as MS invested heavily in WSL to support "linux" workflows.
dotnet on linux might not be completely to parity with dotnet on windows when it comes to platform integration, but it's getting there. dotnet on macos is somewhat similar in this regard and maybe a little bit slower to catch up because may be a little lower in priority than linux.
Anyway, it's not my impression that platforms other than Windows are treated second class. It's just that Microsoft has already invested decades in the Windows implementation, while the linux and macos implementations are rather fresh (although dotnet using mono as a base certainly helped) and therefore less optimized yet in certain regards.
>Java has higher peak JIT performance though after Rosalyn was introduced C# has made strides at closing that gap on many metrics.
That isn't really true. JIT (peak) performance is pretty much comparable between Java and dotnet, maybe even a little better in dotnet/RyuJIT. There are certain areas where one is better than the other.
Roslyn is the C# to MSIL compiler, btw, while the JIT compiler for MSIL is RyuJIT. While Roslyn helps produce MSIL that is (hopefully) great to JIT compile, the actual JIT compilation work happens in RyuJIT.
> GC algorithms are more advanced, namely ZGC and Shenandoah.
Yes, GC is where the JVM really shines - if you go through the trouble of finetuning it to your particular need. At the same time the CLR GC seems to work better by default, but it's a lot less configurable too. In the CLR you basically have only 3 modes of operation: interactive GC (concurrent, trying to avoid GC pauses, with the trade off of less performance, for things like GUI applications), non-concurrent GC (which works well for a lot of CLI type tools and some services, where small GC pauses are less noticeable), and "server" (which is e.g. a lot less eager to "return" memory to the OS, aimed at long running processes that are then usually the only major thing running), and that's basically it, where on java you have tons of available schemes/algorithms and each one is usually heavily configurable as well.
As you already hinted at, dotnet made great strides in the last years to avoid a lot of heap allocations and the GC with value types, Span<> and friends, avoiding boxing a lot of times and bringing stackalloc to managed code, and their use in the standard library. So peak GC performance is a little less important in the CLR than in the JVM (but still very important, of course).
You say "may" a lot, referring to proposed java changes e.g. to add value types, make native code interaction less painful, add some form of async-await. Great. And I hope they will finish up and ship these things eventually, and it will be great implementations. The thing is dotnet already has all these things shipped and in production already, often for quite a while. And while java is still figuring out the design, dotnet can already improve their shipped implementations based on real world experience and data in every new version, and usually does so. Java/JVM is playing catch up here.
>Async/await being the key example. It's been in the CLR for what feels like forever now vs JVM which is only now just getting close to landing Loom. However I feel like Loom is an infinitely better solution to the problem that leverages the unique advantages of JIT compiled VM languages that rely on relatively little unmanaged code.
From what I have seen of Loom so far (which is admittedly not that much) my impression is the direct opposite. To me it so far looks worse conceptually and API-wise than what MS did, while still borrowing a lot of general async-await concepts and even C# and CLR specific concepts, but worse. Don't get me wrong, C# async-await and the CLR support for it has a lot of problematic areas too and is far from perfect and easy, aside from general async-await conceptual problems, but it doesn't seem to me like Loom is avoiding or solving those things, instead adding quite a few of their own warts on top. But as I said, my exposure to Loom is very limited, and it's not a final product yet anyways, so I may be very wrong here.
Loom reminds me of when Java tried to copy a lot of collections-based LINQ stuff but did a terrible job doing so in my humble opinion (they since fixed some of the worst mistakes). Or maybe I am just biased, which is certainly a possibility.