Update: turns out that sharding actually does something.
1 shard: 30 seconds slower than Guava 256 shards: 12 seconds slower 4096 shards: 7 seconds slower (This is with AWFE and no persistent cache.) I could bump up the shards again, but I think instead I'll try ThreadLocal. http://gwt-code-reviews.appspot.com/1578804/ -- http://groups.google.com/group/Google-Web-Toolkit-Contributors