Community
Participate
Working Groups
3.1 M4 From an email from Ed Burnette: You asked me to track down the mail about JDT memory usage. Sorry it took so long. http://www.eclipsepowered.org/archives/2004/11/30/eclipse-grand-challenges/#comment-652 Quote: See these postings for comments on Eclipse memory usage: http://dev.eclipse.org/mhonarc/lists/platform-core-dev/msg00648.html http://dev.eclipse.org/mhonarc/lists/equinox-dev/msg00348.html With a smaller workspace, under a newer Eclipse (3.1M4), the JavaModelCache and things under it were still using 25% of total memory. Most of that was strings. In all, 53% of memory was going into text strings. This is with YourKit 3.2.
Adding my name to the cc list as we are now tracking performance issues more closely. Please remove the performance keyword if this is not a performance bug.
Changed ClassFileInfo and CompilationUnitStructureRequestor to intern the field names, the method selectors and the method parameter types. Note the JavaModelManager intern mechanism is used. Using the VM intern would prevent the strings from being garbagge collected on some VMs.
Noticed that String#substring(...) doesn't make a copy of the underlying char array. As a result, ClassFile handles still point to the whole zip entry path. Changed ClassFile constructor to make a copy of the given name. Similarly changed ImportContainer#getImport(String) to make a coy of the string without the trailing '.*'. Finally changed JavaModelManager#intern(String) to catch other cases. Total memory gain: 440KB out of 3,380 KB (size of the Java model cache in my test workspace)
All ClassFile handles have an element name that ends with ".class". Changed ClassFile to store only the element name without the ".class". This saves another 120KB.
Added WeakHashSetOfCharArray and added JavaModelManager#intern(char[]) that uses this weak hashset. Changed CompilationUnitStructureRequestor to intern the following char arrays: - field info's type name - method info's parameter names, return type and exception types - type info's superclass and superinterfaces This saves another 87KB.
Excellent; if I'm reading this right you've eliminated about 20% of the Java Model cache memory so far.
Thanks ! Along with fixes to bug 89090, bug 89092 and bug 89110, the saved memory in JDT Core is 22.5%. Unfortunately, I think I hit the limit of string sharing. ClassFile's names take the most space now, but interning them has the opposite effect (the weak hashset entries take more space than sharing the file names saves). It is because there is not enough common names in class libraries.
Since there is no other opportunity to share strings in the Java model cache (that would improve memory), I'm marking this bug as fixed. Note that people are welcome to open new bugs if other memory performance problems are noticed in JDT Core.
Verified in I20050330-0500
Created attachment 20123 [details] compress strings using utf-8 encoding I had a weird idea, so I thought I'd try it out. Since most strings are ascii, the normal java string encoding (utf-16) is thus 50% efficient. So I converted ClassFile to store it's strings in utf-8 form using byte arrays. This reduced the memory retained by ClassFile instances by 34% (as measured by YourKit). I'm not sure what this will do to CPU performance, since there is now additional conversion happening, but the memory saved may be worth it for larger projects. Also, I tried converting PackageFragment#names to perform the same trick, but that only effected a 1.2% saving.
