Benchmark scenario with high G1 performance degradation

[Stefan Johansson]

On 2017-04-24 09:29, Jens Wilke wrote:
> On Freitag, 21. April 2017 18:03:45 ICT Jens Wilke wrote:
>> Please mind the core count. My stomach tells me that it could have something
>> to do with the hash table arrays. When you are testing with a system that
>> reports more than 8 cores the allocated arrays will be smaller than in my
>> case, since the cache is doing segmentation.
> Update: I did some tests with different segmentation levels (up to 512) to
> reduce the internal hashtable array size in cache2k. The high CPU consumption
> via OtherRegionsTable::add_reference is still present. So this does not seem
> the problem cause.
One thing you could try is to run with a larger G1HeapRegionSize than 
default. In the G1 log you provided I see that the heap region size is 
4M, so you could try to run with 8M or 16M (-XX:G1HeapRegionSize=8m) to 
see if that improves the situation. The calls to add_reference are 
caused by region to region pointers that G1 needs to keep track of.

Stefan
> Best,
>
> Jens
>