SF100 Status update (Aug 28 2025 versus April 2025) #43
Description
SF100 Status update
Action items
| issue | summary | status |
|---|---|---|
| Q4 slowdown | Looks like we are falling back on HashJoin[5][RIGHT SEMI (FILTER). April profile, August profile |
needs owner |
| Q21 slowdown | Q21 long execution time 13s now versus ~1.5s. Driven by 284 calls to mixed_left_anti_join. Previously we only had 28 calls. | needs owner |
| Q13 failure | Exceptions.h:66] Line: /workspace/velox/./velox/type/StringView.h:59, Function:StringView, Expression: len >= 0 (-47 vs. 0), Source: RUNTIME, ErrorCode: INVALID_STATE |
@karthikeyann - Fix upstream Review pending facebookincubator#14359 |
| Q16 failure | Exceptions.h:66] Line: /workspace/velox/velox/exec/Driver.cpp:574, Function:operator(), Expression: Operator::getOutput failed for [operator: TableScan, plan node ID: 0]: CUDF failure at:/buildcache/release/_deps/cudf-src/cpp/src/ast/expression_parser.cpp:244: An AST expression was provided non-matching operand types., Source: RUNTIME, ErrorCode: INVALID_STATE |
✅ fixed here |
| Q19 | performance hasn't changed since April, but cudf-polars can do this much faster (Slack thread) | tbd |
CPU performance changes
Velox C++ execution. Since Apr 2025, the overall runtime is similar on Grace (~100 seconds). Q4 is still the longest, Q13 jumped +5s and Q21 dropped -10s. So our CPU target hasn't changed too much since April.
System information
GH200
https://github.com/rapidsai/velox-testing on main
Velox on mattgara/upstream-tpch-bench-fixes