Zoltán Borók-Nagy created IMPALA-15649:
------------------------------------------
Summary: Concurrent codegen cache misses on the same key compile
the same module several times
Key: IMPALA-15649
URL: https://issues.apache.org/jira/browse/IMPALA-15649
Project: IMPALA
Issue Type: Improvement
Components: Backend
Reporter: Zoltán Borók-Nagy
With the codegen cache enabled, LlvmCodeGen::FinalizeModule() looks up the
cache and on a miss optimizes, compiles and stores the module (llvm-codegen.cc,
master 9c8bcc9585). When several fragment instances miss on the same key at the
same time, as on the first run of a query after a restart, each of them
optimizes and compiles the same module. CodeGenCache::Store() only skips a
store while another store of the same key is in progress, so a later store
replaces the equivalent entry stored first and evicts it.
*Proposal:* compile a module once per key:
* The first fragment that misses on a key registers as its compiler.
* The others wait for it, with a timeout, then look up the cache again and
compile the module themselves only if it is still missing.
* A store is skipped if the cache already holds an entry with the same function
names hash and at least the same optimization level.
* Profile counters show the number and time of the waits.
*Measured* on TPC-DS SF100 with 5 executors (MT_DOP=0): with a prototype
(timeout flag default 10 s), the first run after a restart was 3.9% faster in
total; warm runs were unchanged.
Prototype: https://gerrit.cloudera.org/#/c/25203/
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
