Nfs server allocates an inode table for each subvolume. Each inode table allocates a mempool of around 200,000 entries for inode and dentry. This contributes to relatively high memory used by each of subvolumes. One of the possible solutions would be to lower the number of entries in inode and dentry mem-pool.
just to be sure we fix the right bug, this should be seen for every glusterfs daemon because that mem allocation is coming from inode_table_new. There this value is hard-coded. Patch coming soon.
PATCH: http://patches.gluster.com/patch/5575 in master (core: Use lru_limit as count for inode and dentry mempool)