CMS, PLAB-size and fragmentation

Florian Binder java at java4.info
Tue Aug 14 02:27:47 PDT 2012


Hi everybody,

one of our servers (uses CMS with ParNew) promotes a lot of very small 
objects at the young gc to the tenure (survivor is disabled since the 
objects would survive it anyway):

5[4]: 10703/101530/19930
5[6]: 5482/50575/10115

(Sometimes even more)
Therefore I increased the CMSOldPLABMax=131072  (ok, this might be to 
large and 32k would be enough) and decreased the CMSOldPLABMin=8 because 
there are always a few larger objects which have always a different size:

11[134]: 7/8/8
11[148]: 4/944/8
11[150]: 7/9/9
11[156]: 6/24/8
11[158]: 4/16/8
11[160]: 6/8/8
11[164]: 3/8/8
11[166]: 7/64/8
11[224]: 7/8/8

My questions now are:
Does this have any effect on fragmentation of the tenure space? I assume 
the increasing of the maximum would have a positive effect because the 
small objects are more compacted. Is this right?
Are there any other negative effects of changing these parameters?

Thanks a lot,
Flo



More information about the hotspot-gc-use mailing list