From: ClassCastException on 22 Jun 2010 20:29 On Sun, 20 Jun 2010 22:14:32 -0400, Arne Vajhøj wrote: > On 27-05-2010 21:32, ClassCastException wrote: >> Actually, the other thing we really need is OS/JVM integration so the >> JVM heap plays nicer with paging and multitasking. > > I believe that both work fine with current JVM's. If everything you're doing (except the OS) is in the JVM, it works fine. If what's in the JVM is quite small, it works fine. But if the JVM heap grows to say a gig or two and you have Firefox, Photoshop, and a dozen other big applications open as well, the JVM has a tendency to get swapped out, and when it's used again and a major collection is triggered, ouch. Nonresponsive for literally *minutes*. The G1 collector that's supposed to be in JDK7 (thus bringing this topic full-circle) might alleviate that somewhat, given what I understand about how it will work. But the optimum logically must be if the paging and the garbage collector cooperated in some way, which pretty much requires at the very least adding some kind of specific API support for garbage collection to the OS. Ultimately, though, there are security, stability, portability, and other benefits to standardizing on a VM as user-mode application host and then performance benefits to moving that VM into the OS and much OS functionality (however much is not kernel-mode) into the VM and API library.
From: Tom Anderson on 23 Jun 2010 05:04 On Wed, 23 Jun 2010, ClassCastException wrote: > On Sun, 20 Jun 2010 22:14:32 -0400, Arne Vajh?j wrote: > >> On 27-05-2010 21:32, ClassCastException wrote: >> >>> Actually, the other thing we really need is OS/JVM integration so the >>> JVM heap plays nicer with paging and multitasking. >> >> I believe that both work fine with current JVM's. > > If everything you're doing (except the OS) is in the JVM, it works fine. > If what's in the JVM is quite small, it works fine. But if the JVM heap > grows to say a gig or two and you have Firefox, Photoshop, and a dozen > other big applications open as well, the JVM has a tendency to get > swapped out, and when it's used again and a major collection is > triggered, ouch. Nonresponsive for literally *minutes*. On which subject, has anyone pointed to this: http://www.managedruntime.org/ http://www.networkworld.com/news/2010/061510-azul-systems-searches-for-managed.html?hpg1=bn yet? Azul, who are pretty expert and large-scale java performance, have this idea to make the JVM and OS work a bit more closely together, by adding kernel modules to the OS (so far, just linux) and then modifying the JVM to use them. I don't know about the details, as information is thin on the ground, but it's interesting stuff. Might even be something interesting for Patricia to work on. tom -- There is no latest trend.
First
|
Prev
|
Pages: 1 2 3 4 5 6 Prev: Centralizing instantiation? Next: Display Byte value for GB2123 Character |