Hacker News new | past | comments | ask | show | jobs | submit login
Building Memory-efficient Java Applications (ibm.com)
85 points by markerdmann on Feb 2, 2012 | hide | past | favorite | 15 comments



Very nice presentation. However much of this can be actually discovered and felt first hand just by playing with profilers that support memory profiling (such as yourkit[1]).

After a few hours playing with it, you will never forget the precious lessons on this presentation.

[1] http://www.yourkit.com/


This is probably an 'apple and oranges' type of question, but I haven't used Erlang, nor Java to build large applications.

Does anyone know how Erlang compares to Java in regards to memory usage/performance?


Per object overhead in functional languages is usually lower since polymorphism is handled at the function level rather than the data level. You don't have to store class pointers etc, just the raw data and probably a type tag.

I can't find a reference for erlang's representation but I believe it has only a little more overhead than ocaml which is described here - http://rwmj.wordpress.com/2009/08/04/ocaml-internals/


I know I should thank you by up voting. But also, just want to say that the guide was a very interesting and insightful read.


using their official docs

http://erlang.org/doc/efficiency_guide/advanced.html#9

(I haven't tried this:

    erts_debug:flat_size(dict:new()).
http://stackoverflow.com/questions/1964015/erlang-what-is-mo...


Per-object overhead is generally more reasonable.


You can have a whole Erlang process for a price of (roughly) 100 Java objects.


For those of you who may not be familiar with them, using http://fastutil.dsi.unimi.it/ collections instead of Java collections is essentially the work-around for most things you'll use. Fastutil has custom classes that remove the pointer overhead associated with storing primitives. Unfortunately, Java really falls apart at the primitives aren't objects thing when it comes to performance.


It just goes to show that whatever language you use, you need to understand what the abstractions are doing: what bytes are really going where?


Some (not all) of these cases would be helped massively if the JVM gained an equivalent of the fixnum/flonum types from the LISP world. This has been proposed kind-of seriously in the past, e.g.:

http://blogs.oracle.com/jrose/entry/fixnums_in_the_vm

Has anyone seen anything more recent?



Does anyone know of a similar study of memory efficiency in C# (or the CLR in general)?


Cool! Very good explanations.


TreeMap<Double, Double>, 100 entries, 8.6 kB, 82% overhead

Man, that's just annoying.


That's great, but perhaps they should teach some of this to their colleagues at IBM first.




Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: