OOM on bootstrap after long GC pause

I have a 20-node cluster (i2.4xlarge) with vnodes (default of 256) and 500-700GB per node. SSTable counts are <10 per table. I am attempting to provision additional nodes, but bootstrapping OOMs every time after about 10 hours with a sudden long GC pause:

INFO  [Service Thread] 2015-10-05 23:33:33,373 GCInspector.java:252 - G1 Old Generation GC in 1586126ms.  G1 Old Gen: 49213756976 -> 49072277176;
...
ERROR [MemtableFlushWriter:454] 2015-10-05 23:33:33,380 CassandraDaemon.java:223 - Exception in thread Thread[MemtableFlushWriter:454,5,main]
java.lang.OutOfMemoryError: Java heap space

I have tried increasing max heap to 48G just to get through the bootstrap, to no avail.