Sorry, didn’t pay attention that the size of every object is 1 KB. Then I would suggest to share your test/heap dump to see what happens.
— Denis > On Jan 21, 2017, at 9:10 PM, Denis Magda <[email protected]> wrote: > > Hello Peter, > > This behavior is expected. Look, you serialize 2000 objects * 1000 KB which > gives ~ 2 GB in total. After that you hold the references to these objects by > putting them into an ordinary hash map. The objects will not be removed from > the heap until the map keeps them. > > One the other side, when the same objects are placed into an off-heap cache > they will be collected from the heap because at the end there will be no > reachable references to them. > > — > Denis > >> On Jan 21, 2017, at 2:46 PM, Peter Schmitt <[email protected] >> <mailto:[email protected]>> wrote: >> >> Hello Ignite-Community! >> >> I would like to use the Marshaller from Ignite for de-/serialising objects. >> I would like to re-use it, because it's very fast and I'm using Ignite >> anyway. >> >> I tried it with: >> >> //every thread gets the Marshaller via: >> GridKernalContext ctx = ((IgniteKernal) ignite).context(); >> Marshaller marshaller = ctx.grid().configuration().getMarshaller(); >> >> -> >> >> byte serializedResult[] = marshaller.marshal(obj); >> or >> Object deserializedResult = marshaller.unmarshal(serializedResult, >> ctx.getClass().getClassLoader()); >> >> In a simple test it worked very well (and fast). >> However, once I'm using larger Objects (e.g. 2000 instances each ~1kb after >> marshalling) and store the results in a map, the Heap consumption grows a >> lot (almost 2gb in few seconds). >> If I store the same objects in an Ignite OffHeap-Cache, the (OffHeap) >> memory-consumption is minimal. >> In the debugger I saw that the heap grows once those 2000 instances get >> processed. >> (A manually triggered GC can just release few mb.) >> >> I even tried to re-use the marshaller, but nothing changed. >> At runtime BinaryMarshaller is used which wraps GridBinaryMarshaller. >> >> Any hint to get rid of the leak is appreciated, >> Peter >
