5 ms·

Building Memory-efficient Java Applications

85 points by markerdmann 15y ago

dikanggu 15y ago
Cool! Very good explanations.
DanielRibeiro 15y ago
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/ http://www.yourkit.com/
jesperhh 15y ago
Does anyone know of a similar study of memory efficiency in C# (or the CLR in general)?
morsch 15y ago
TreeMap<Double, Double>, 100 entries, 8.6 kB, 82% overhead Man, that's just annoying.
NSMeta 15y ago
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?
- abecedarius 15y ago
  Per-object overhead is generally more reasonable.
- thesz 15y ago
  You can have a whole Erlang process for a price of (roughly) 100 Java objects.
- jamii 15y ago
  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/ http://rwmj.wordpress.com/2009/08/04/ocaml-internals/
  - NSMeta 15y ago
    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.
- gtani 15y ago
  using their official docs http://erlang.org/doc/efficiency_guide/advanced.html#9 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-most-wrong-with-this-trie-implementation http://stackoverflow.com/questions/1964015/erlang-what-is-mo...
insom 15y ago
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?
sjwright 15y ago
That's great, but perhaps they should teach some of this to their colleagues at IBM first.
gtani 15y ago
from 2008. slide 25: experiments on a sampling of different JVMs. More recent : http://vanillajava.blogspot.com/2011/07/java-getting-size-of-object.html http://vanillajava.blogspot.com/2011/07/java-getting-size-of... http://vanillajava.blogspot.com/2011/07/java-how-much-memory-do-different.html http://vanillajava.blogspot.com/2011/07/java-how-much-memory... http://kohlerm.blogspot.com/2008/10/knowledge-about-much-memory-things-need.html http://kohlerm.blogspot.com/2008/10/knowledge-about-much-mem... http://kohlerm.blogspot.com/2009/02/how-to-really-measure-memory-usage-of.html http://kohlerm.blogspot.com/2009/02/how-to-really-measure-me... http://marxsoftware.blogspot.com/2011/12/estimating-java-object-sizes-with.html http://marxsoftware.blogspot.com/2011/12/estimating-java-obj... (i have zillions bookmarks on this)
dasmoth 15y ago
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 http://blogs.oracle.com/jrose/entry/fixnums_in_the_vm Has anyone seen anything more recent?
joshhart 15y ago
For those of you who may not be familiar with them, using http://fastutil.dsi.unimi.it/ 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.

85 points by markerdmann 15y ago

Cool! Very good explanations.

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/ http://www.yourkit.com/

jesperhh 15y ago

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

morsch 15y ago

TreeMap<Double, Double>, 100 entries, 8.6 kB, 82% overhead Man, that's just annoying.

NSMeta 15y ago

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?

abecedarius 15y ago
Per-object overhead is generally more reasonable.
thesz 15y ago
You can have a whole Erlang process for a price of (roughly) 100 Java objects.
jamii 15y ago
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/ http://rwmj.wordpress.com/2009/08/04/ocaml-internals/
- NSMeta 15y ago
  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.
gtani 15y ago
using their official docs http://erlang.org/doc/efficiency_guide/advanced.html#9 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-most-wrong-with-this-trie-implementation http://stackoverflow.com/questions/1964015/erlang-what-is-mo...

insom 15y ago

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?

sjwright 15y ago

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

gtani 15y ago

from 2008. slide 25: experiments on a sampling of different JVMs. More recent : http://vanillajava.blogspot.com/2011/07/java-getting-size-of-object.html http://vanillajava.blogspot.com/2011/07/java-getting-size-of... http://vanillajava.blogspot.com/2011/07/java-how-much-memory-do-different.html http://vanillajava.blogspot.com/2011/07/java-how-much-memory... http://kohlerm.blogspot.com/2008/10/knowledge-about-much-memory-things-need.html http://kohlerm.blogspot.com/2008/10/knowledge-about-much-mem... http://kohlerm.blogspot.com/2009/02/how-to-really-measure-memory-usage-of.html http://kohlerm.blogspot.com/2009/02/how-to-really-measure-me... http://marxsoftware.blogspot.com/2011/12/estimating-java-object-sizes-with.html http://marxsoftware.blogspot.com/2011/12/estimating-java-obj... (i have zillions bookmarks on this)

dasmoth 15y ago

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 http://blogs.oracle.com/jrose/entry/fixnums_in_the_vm Has anyone seen anything more recent?

joshhart 15y ago

For those of you who may not be familiar with them, using http://fastutil.dsi.unimi.it/ 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.