Strange behaviour of G1 GC

[Alexander Bulaev]

Hello!

I am writing to you about strange behaviuor of G1 GC that I have encountered in our production environment.
Sometimes there are happening Full GCs that are cleaning a lot of garbage:
2014-07-24T14:27:57.020+0400: 94749.771: [Full GC (Allocation Failure)  11G->5126M(12G), 13.7944745 secs]

I suppose that this is garbage in the old generation. I expect it to be cleaned during mixed and concurrent GCs, but, according to the logs, the last concurrent phase happened over half an hour earlier prior to that Full GC:
2014-07-24T13:49:43.228+0400: 92455.979: [GC concurrent-mark-start]

And there is no evidence in logs that this concurrent mark has ever ended. Seems like that concurrent GC just hang somewhere.
Same thing with mixed GCs:
2014-07-24T13:42:47.425+0400: 92040.176: [GC pause (G1 Evacuation Pause) (mixed)

Please help me understand this problem and find a solution if possible.
We are using Java 8u5. I can supply these GC logs if needed.
Thanks.

Best regards,
Alexander Bulaev
Java developer, Yandex LLC