Issue with broadcast Join in Spark 1.6.0 - sql

Question

Spark creates different SQL execution plans in versions 1.5.2 and 1.6.0.
The database contains one large facts table (f) and several small dimension tables (d1, d2).
The dimension tables (d1, d2) are cached in Spark using:
sqlContext.cacheTable("d1");
sqlContext.sql("SELECT * from d1").count();
sqlContext.cacheTable("d2");
sqlContext.sql("SELECT * from d2").count();
The size of tables d1 and d2: ~100MB (the tab STORAGE shows this size for the tables in Spark Web UI).
The Spark SQL is configured by the broadcast parameter (1GB) which is more than the size of the tables:
spark.sql.autoBroadcastJoinThreshold=1048576000
The SQL is:
SELECT d1.name, d2.name, SUM(f.clicks)
FROM f
JOIN d1 on f.d1_id = d1.id
JOIN d2 on f.d2_id = d2.id
WHERE d1.name='A' AND f.date=20160101
GROUP BY d1.name, d2.name
ORDER BY SUM(f.clicks) DESC
The query is executed using ThriftServer.
When we run this query using Spark 1.5.2 the query is executed fast and the execution plan contains BroadcastJoin wth dimensions.
But when we run the same query using Spark 1.6.0 the query is executed very slow and the execution plan contains SortMergeJoin with dimensions.
The environment for both executions is the same (the same Yarn Cluster).
Provided executors count, tasks count per executor and memory of the executors are the same for both executions.
What can I do to configure Spark 1.6.0 for using BroadcastJoin in sql executions?
Thanks.