Hello Philip Zeyliger, Dan Hecht, I'd like you to reexamine a change. Please visit
http://gerrit.cloudera.org:8080/8363 to look at the new patch set (#5). Change subject: IMPALA-4456: Address scalability issues of qs_map_lock_ and client_request_state_map_lock_ ...................................................................... IMPALA-4456: Address scalability issues of qs_map_lock_ and client_request_state_map_lock_ The following 2 locks have shown to be frequent points of contention on recent perf runs: - qs_map_lock_ - client_request_state_map_lock_ Since these are process wide locks, any threads waiting on these locks potentially slow down the runtime of a query. I tried to address this previously by converting the client_request_state_map_lock_ to a reader-writer lock. This showed great perf improvements in the general case, however, there were edge cases with big regressions as well. In the general case, strict readers of the map got through so quickly that we were able to see a reduction in the number of client connections created, since this lock was contended for in the context of Thrift threads too. The bad case is when writers were starved trying to register a new query since there were so many readers. Changing the starve option resulted in worse read performance all round. Another approach which is relatively simpler is to shard the locks, which proves to be very effective with no regressions. The maps and locks are sharded to a default of 4 buckets initally. Query IDs are created by using boost::uuids::random_generator. We use the high bits of a query ID to assign queries to buckets. I verified that the distribution of the high bits of a query ID are even across buckets on my local machine: For 10,000 queries sharded across 4 buckets, the distribution was: bucket[0]: 2500 bucket[1]: 2489 bucket[2]: 2566 bucket[3]: 2445 A micro-benchmark is added to measure the improvement in performance. This benchmark creates multiple threads each of which creates a QueryState and accesses it multiple times. We can see improvements in the range 2x - 3.5x. BEFORE: ------------------Benchmark 1: Create and access Query States. Total Time (#Queries: 5 #Accesses: 100) : 1ms Total Time (#Queries: 50 #Accesses: 100) : 8ms Total Time (#Queries: 50 #Accesses: 1000) : 54ms Total Time (#Queries: 500 #Accesses: 100) : 76ms Total Time (#Queries: 500 #Accesses: 1000) : 543ms AFTER: ------------------Benchmark 1: Create and access Query States. Total Time (#Queries: 5 #Accesses: 100) : 2173.59K clock cycles Total Time (#Queries: 50 #Accesses: 100) : 4ms Total Time (#Queries: 50 #Accesses: 1000) : 15ms Total Time (#Queries: 500 #Accesses: 100) : 46ms Total Time (#Queries: 500 #Accesses: 1000) : 151ms This change introduces a ShardedQueryMapTemplate, which is used to replace the QueryExecMgr::qs_map_ and the ImpalaServer::client_request_state_map_, and their corresponding locks. NOTE: This microbenchmark has shown that SpinLock has better performance than boost::mutex for the qs_map_lock_'s, so that change has been made too. TODO: Add benchmark for client_request_state_map_lock_ too. The APIs around that are more complicated, so this patch only includes the benchmarking of qs_map_lock_. TODO 2: Consider adopting the ShardedQueryMapTemplate for the SessionStateMap. Change-Id: I61089090e1095da45a8a64ed3ccc78bd310807f1 --- M be/src/benchmarks/CMakeLists.txt A be/src/benchmarks/process-wide-locks-benchmark.cc M be/src/runtime/query-exec-mgr.cc M be/src/runtime/query-exec-mgr.h M be/src/service/impala-http-handler.cc M be/src/service/impala-server.cc M be/src/service/impala-server.h A be/src/util/sharded-query-map-util.h M be/src/util/uid-util.h 9 files changed, 353 insertions(+), 53 deletions(-) git pull ssh://gerrit.cloudera.org:29418/Impala-ASF refs/changes/63/8363/5 -- To view, visit http://gerrit.cloudera.org:8080/8363 To unsubscribe, visit http://gerrit.cloudera.org:8080/settings Gerrit-Project: Impala-ASF Gerrit-Branch: master Gerrit-MessageType: newpatchset Gerrit-Change-Id: I61089090e1095da45a8a64ed3ccc78bd310807f1 Gerrit-Change-Number: 8363 Gerrit-PatchSet: 5 Gerrit-Owner: Sailesh Mukil <sail...@cloudera.com> Gerrit-Reviewer: Dan Hecht <dhe...@cloudera.com> Gerrit-Reviewer: Philip Zeyliger <phi...@cloudera.com> Gerrit-Reviewer: Sailesh Mukil <sail...@cloudera.com>