Presto: Reinstate a query optimizer flag for GPU only (ALSO SEE PR #111)

presto/docker/config/template/etc_coordinator/config_native_cpu.properties

@@ -0,0 +1,112 @@
+# Run this node as the cluster coordinator; it schedules and manages queries.
+coordinator=true
+# Do not schedule worker tasks on the coordinator to avoid resource contention.
+node-scheduler.include-coordinator=false
+# Coordinator REST/HTTP port for clients and workers.
+http-server.http.port=8080
+# Embedded service that provides node discovery for workers.
+discovery-server.enabled=true
+# Address workers use to register with the discovery service.
+discovery.uri=http://presto-coordinator:8080
+
+# Set Presto version string to match workers for compatibility in tests.
+presto.version=testversion
+
+# Keep up to 30 rolled log files to bound disk usage.
+log.max-history=30
+# Rotate logs at ~100MB per file for manageable artifacts.
+log.max-size=104857600B
+# Reserve heap headroom per node to reduce full GC and OOM risk.
+memory.heap-headroom-per-node={{ .HeadroomGb }}GB
+
+# Limit pending splits per task to avoid excessive memory usage.
+node-scheduler.max-pending-splits-per-task=2000
+# Cap concurrent splits per node for balanced scheduling.
+node-scheduler.max-splits-per-node=2000
+
+# Optimizer flags
+# Use a default filter factor to estimate the selectivity of filters in queries? Good for GPU. Bad for CPU.
+optimizer.default-filter-factor-enabled=false
+# Use known constraints to simplify plan and filters.
+optimizer.exploit-constraints=true
+# Rewrite large IN lists as joins for performance in some cases.
+optimizer.in-predicates-as-inner-joins-enabled=true
+# Allow partial aggregations to reduce data shuffled across stages.
+optimizer.partial-aggregation-strategy=automatic
+# Prefer partial aggregations when beneficial.
+optimizer.prefer-partial-aggregation=true
+# Default selectivity heuristic for joins when stats are missing.
+optimizer.default-join-selectivity-coefficient=0.1
+# Infer additional range predicates to improve filtering.
+optimizer.infer-inequality-predicates=true
+# Support complex equi-join patterns in the optimizer.
+optimizer.handle-complex-equi-joins=true
+# Add dynamic domain filters to reduce scanned data.
+optimizer.generate-domain-filters=true
+# Upper limit for broadcasted table size to avoid memory blowups.
+# See: https://github.com/prestodb/presto/issues/22161#issuecomment-1994128619
+join-max-broadcast-table-size={{ .JoinMaxBroadcastTableSizeMb }}MB
+
+# Client request timeout to avoid hung queries.
+query.client.timeout=30m
+# Use phased execution policy for improved large query scheduling.
+query.execution-policy=phased
+# Kill queries based on total reservation on blocked nodes to recover memory.
+query.low-memory-killer.policy=total-reservation-on-blocked-nodes
+# Upper limit on query wall time to keep tests bounded.
+query.max-execution-time=30m
+# Keep metadata of up to 1000 queries for UI and debugging.
+query.max-history=1000
+# Memory quotas per node and cluster to protect stability.
+query.max-total-memory-per-node={{ .JavaQueryMaxTotalMemPerNodeGb }}GB
+query.max-total-memory={{ mul .JavaQueryMaxTotalMemPerNodeGb .NumberOfWorkers }}GB
+query.max-memory-per-node={{ .JavaQueryMaxMemPerNodeGb }}GB
+query.max-memory={{ mul .JavaQueryMaxMemPerNodeGb .NumberOfWorkers }}GB
+# Allow deep stage DAGs required by certain benchmark queries.
+query.max-stage-count=1300
+# Retain query info at least this long for diagnostics.
+query.min-expire-age=120.00m
+# Larger scheduling batches for better throughput in benchmarks.
+query.min-schedule-split-batch-size=2000
+# Raise warning threshold to align with higher max stage count.
+query.stage-count-warning-threshold=150
+# Increase serialized plan/query length limit for complex benchmark queries.
+query.max-length=2000000
+
+# Disable dynamic filtering for deterministic benchmarking.
+experimental.enable-dynamic-filtering=false
+# Cap revocable memory per node to avoid overcommit.
+experimental.max-revocable-memory-per-node=50GB
+# Limit disk spill usage per node to bound IO and disk usage.
+experimental.max-spill-per-node=50GB
+# Enable repartitioning improvements for shuffle efficiency.
+experimental.optimized-repartitioning=true
+# Enable dereference and subfield pushdown to reduce scanned data.
+experimental.pushdown-dereference-enabled=true
+experimental.pushdown-subfields-enabled=true
+# Cluster-wide guardrail for spill during a single query per node.
+experimental.query-max-spill-per-node=50GB
+# Disable reserved memory pool to simplify test behavior.
+experimental.reserved-pool-enabled=false
+# Stop spilling when disk usage exceeds this fraction.
+experimental.spiller-max-used-space-threshold=0.7
+# Directory for spill files during execution.
+experimental.spiller-spill-path=/tmp
+
+
+# Min workers before query starts; keep minimal for quick tests.
+query-manager.required-workers=1
+# Maximum wait for required workers to join.
+query-manager.required-workers-max-wait=10s
+
+# Set required configuration for Presto C++ workers as indicated in https://prestodb.io/docs/current/presto_cpp/properties.html#coordinator-properties
+native-execution-enabled=true
+# Disable Java-side hash generation optimizations not used by native workers.
+optimizer.optimize-hash-generation=false
+# Use RE2J regex engine for performance and determinism.
+regex-library=RE2J
+# Enable alternative function signatures for native compatibility.
+use-alternative-function-signatures=true
+
+# Optimize for queries that can run entirely on a single worker.
+single-node-execution-enabled=true

presto/docker/config/template/etc_coordinator/config_native.properties → presto/docker/config/template/etc_coordinator/config_native_gpu.properties

@@ -25,6 +25,8 @@ node-scheduler.max-pending-splits-per-task=2000
 node-scheduler.max-splits-per-node=2000
 
 # Optimizer flags
+# Use a default filter factor to estimate the selectivity of filters in queries? Good for GPU. Bad for CPU.
+optimizer.default-filter-factor-enabled=true
 # Use known constraints to simplify plan and filters.
 optimizer.exploit-constraints=true
 # Rewrite large IN lists as joins for performance in some cases.

presto/docker/docker-compose.native-cpu.yml

@@ -4,7 +4,7 @@ services:
       file: docker-compose.common.yml
       service: presto-base-coordinator
     volumes:
-      - ./config/generated/etc_coordinator/config_native.properties:/opt/presto-server/etc/config.properties
+      - ./config/generated/etc_coordinator/config_native_cpu.properties:/opt/presto-server/etc/config.properties
 
   presto-native-worker-cpu:
     extends:

presto/docker/docker-compose.native-gpu.yml

@@ -4,7 +4,7 @@ services:
       file: docker-compose.common.yml
       service: presto-base-coordinator
     volumes:
-      - ./config/generated/etc_coordinator/config_native.properties:/opt/presto-server/etc/config.properties
+      - ./config/generated/etc_coordinator/config_native_gpu.properties:/opt/presto-server/etc/config.properties
 
   presto-native-worker-gpu:
     extends: