Robert Kruszewski created SPARK-61592:
-----------------------------------------
Summary: BytesToBytesMap.safeLookup causes cache line thrashing,
scales badly with increased threads
Key: SPARK-61592
URL: https://issues.apache.org/jira/browse/SPARK-61592
Project: Spark
Issue Type: Improvement
Components: Spark Core
Affects Versions: 4.3.0
Reporter: Robert Kruszewski
BytesToBytesMap.safeLookup updates probe counters on every lookup. Broadcast
hash joins call it from every task on the same read-only UnsafeHashedRelation
map. It incremented the shared fields numKeyLookups and numProbes on each
lookup, so all threads wrote to the same cache line. On a 32-core machine, a
broadcast hash join on three long keys spent about 2 us per probe row on this
contention.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]