ASAN maintains stacks for each allocation and free of memory. Impala
sometimes allocates/frees memory from codegen'd code, so this means
that the number of distinct stacks is unbounded. ASAN is storing
these stacks in a hash table with a fixed number of buckets (one million).
As the stacks accumulate, allocations and frees get slower and slower,
because the lookup in this hashtable gets slower. This causes test
execution time to degrade over time. Since backend tests and custom cluster
tests don't have long running daemons, only the end to end tests are
affected.
This adds support for breaking end-to-end test execution into shards,
restarting Impala between each shard. This uses the preexisting shard_tests
pytest functionality introduced for the docker-based tests in IMPALA-6070.
The number of shards is configurable via the EE_TEST_SHARDS environment
variable. By default, EE_TEST_SHARDS=1 and no sharding is used.
Without sharding, an ASAN core job takes about 16-17 hours. With 6 shards,
it takes about 9 hours. It is recommended to always use sharding with ASAN.
Testing:
- Ran core job
- Ran ASAN with EE_TEST_SHARDS=6
Change-Id: I0bdbd79940df2bc7b951efdf0f044e6b40a3fda9
Reviewed-on: http://gerrit.cloudera.org:8080/16155
Reviewed-by: Impala Public Jenkins <impala-public-jenkins@cloudera.com>
Tested-by: Impala Public Jenkins <impala-public-jenkins@cloudera.com>
605e301739