Skip to content

Include configuration setup for Presto benchmark runs #32

New issue
New issue
Closed
Closed
Include configuration setup for Presto benchmark runs#32
Assignees
misiugodfrey

Description

@GregoryKimball
GregoryKimball
opened on Sep 3, 2025

Here are some notes on required configuration details

# hive.properties
hive.file-splittable=false
# This is needed for correctness as we cannot split a single parquet file

# coordinator config
single-node-execution-enabled=true
# This just happens to be the mode we're focusing on while IBM focuses on distributed

# worker config
task.max-drivers-per-task=4
# We've seen in the past that this increases overlap between IO and compute for GPU. I believe this value should be higher for effective utilisation of CPU in case of cpu-only native runs

@devavret is there more to add here?

Activity

  1. GregoryKimball
    assigned
    misiugodfrey
    on Sep 3, 2025

  2. karthikeyann commented on Sep 3, 2025

    @karthikeyann
    karthikeyann
    on Sep 3, 2025
    Contributor

    I have these changes related to Presto setup.

    diff --git a/presto/docker/config/etc_common/jvm.config b/presto/docker/config/etc_common/jvm.config
index e33ab84..8a1aba4 100644
--- a/presto/docker/config/etc_common/jvm.config
+++ b/presto/docker/config/etc_common/jvm.config
@@ -7,3 +7,20 @@
 -XX:+HeapDumpOnOutOfMemoryError
 -XX:+ExitOnOutOfMemoryError
 -Djdk.attach.allowAttachSelf=true
+--add-opens=java.base/java.io=ALL-UNNAMED
+--add-opens=java.base/java.lang=ALL-UNNAMED
+--add-opens=java.base/java.lang.ref=ALL-UNNAMED
+--add-opens=java.base/java.lang.reflect=ALL-UNNAMED
+--add-opens=java.base/java.net=ALL-UNNAMED
+--add-opens=java.base/java.nio=ALL-UNNAMED
+--add-opens=java.base/java.security=ALL-UNNAMED
+--add-opens=java.base/javax.security.auth=ALL-UNNAMED
+--add-opens=java.base/javax.security.auth.login=ALL-UNNAMED
+--add-opens=java.base/java.text=ALL-UNNAMED
+--add-opens=java.base/java.util=ALL-UNNAMED
+--add-opens=java.base/java.util.concurrent=ALL-UNNAMED
+--add-opens=java.base/java.util.concurrent.atomic=ALL-UNNAMED
+--add-opens=java.base/java.util.regex=ALL-UNNAMED
+--add-opens=java.base/jdk.internal.loader=ALL-UNNAMED
+--add-opens=java.base/sun.security.action=ALL-UNNAMED
+--add-opens=java.security.jgss/sun.security.krb5=ALL-UNNAMED
diff --git a/presto/docker/config/etc_coordinator/config_native.properties b/presto/docker/config/etc_coordinator/config_native.properties
index 21db547..a09f4d9 100644
--- a/presto/docker/config/etc_coordinator/config_native.properties
+++ b/presto/docker/config/etc_coordinator/config_native.properties
@@ -8,9 +8,9 @@ discovery.uri=http://presto-coordinator:8080
 presto.version=testversion
 
 # Query execution settings
-query.max-memory=2GB
-query.max-memory-per-node=1GB
-query.max-total-memory-per-node=1GB
+query.max-memory=20GB
+query.max-memory-per-node=10GB
+query.max-total-memory-per-node=10GB
 
 # Set required workers to 1 (minimum allowed) and reduce wait time
 query-manager.required-workers=1

  3. karthikeyann commented on Sep 3, 2025

    @karthikeyann
    karthikeyann
    on Sep 3, 2025
    Contributor

    Exact memory numbers may vary. We should test in different GPUs and arrive at numbers that works best in all machines.

  4. devavret commented on Sep 4, 2025

    @devavret
    devavret
    on Sep 4, 2025
    Contributor

    @devavret is there more to add here?

    These are all the ones that affect GPU execution for now

  5. patdevinwilson commented on Sep 17, 2025

    @patdevinwilson
    patdevinwilson
    on Sep 17, 2025
    Contributor

    Here are my configs for successful runs on GPU and CPU Gh200

    Notably- "cudf.exchange.enabled": true,

    tpch_sf100_cpu_optimized.json
    tpch_sf100_gpu.json
    tpch_sf1000_cpu.json
    tpch_sf1000_gpu.json

    The actual CUDF parquet reader is controlled via Velox command-line flags, not .properties files:

    From /data/nvidia-work/presto/presto-native-execution/zrl_bin/run_worker.sh

    VELOX_ARGS="-velox_cudf_debug=true -velox_cudf_table_scan=true -velox_cudf_enabled=true -velox_cudf_exchange=true -velox_cudf_memory_resource=pool -v=3"

    Presto server is launched with these flags:

    presto_server -etc_dir $WORKER_CONFIG_DIR $VELOX_ARGS

  6. GregoryKimball commented on Sep 25, 2025

    @GregoryKimball
    GregoryKimball
    on Sep 25, 2025
    Author

    Here is an updated configuration set based on PBench auto-generation

    coordinator=true
discovery-server.enabled=true
discovery.uri=http://coordinator:8080
http-server.http.port=8080
node-scheduler.include-coordinator=false
presto.version=${PRESTO_BUILD_VERSION}

http-server.log.max-size=67108864B
http-server.log.max-history=2
http-server.threads.max=1000

http-server.authentication.type=PASSWORD
http-server.authentication.allow-forwarded-https=true

log.max-history=30
log.max-size=104857600B
memory.heap-headroom-per-node=22GB

node-scheduler.max-pending-splits-per-task=2000
node-scheduler.max-splits-per-node=2000

# Optimizer flags
optimizer.joins-not-null-inference-strategy=USE_FUNCTION_METADATA
optimizer.default-filter-factor-enabled=true
optimizer.exploit-constraints=true
optimizer.in-predicates-as-inner-joins-enabled=true
optimizer.partial-aggregation-strategy=automatic
optimizer.prefer-partial-aggregation=true
optimizer.default-join-selectivity-coefficient=0.1
optimizer.infer-inequality-predicates=true
optimizer.handle-complex-equi-joins=true
optimizer.generate-domain-filters=true
# https://github.com/prestodb/presto/issues/22161#issuecomment-1994128619
join-max-broadcast-table-size=1250MB

query.client.timeout=30m
query.execution-policy=phased
query.low-memory-killer.policy=total-reservation-on-blocked-nodes
query.max-execution-time=30m
query.max-history=1000
query.max-total-memory-per-node=87GB
query.max-total-memory=696GB
query.max-memory-per-node=78GB
query.max-memory=624GB
# query.max-stage-count Needed for Nielsen Q5
query.max-stage-count=1300
query.min-expire-age=120.00m
query.min-schedule-split-batch-size=2000
query.stage-count-warning-threshold=150
# query.max-length Needed for Nielsen Q5 and Q9
query.max-length=2000000

scale-writers=true
scheduler.http-client.max-requests-queued-per-destination=4096
shutdown.grace-period=30s
sink.max-buffer-size=1GB

experimental.enable-dynamic-filtering=false
experimental.max-revocable-memory-per-node=50GB
experimental.max-spill-per-node=50GB
experimental.optimized-repartitioning=true
experimental.pushdown-dereference-enabled=true
experimental.pushdown-subfields-enabled=true
experimental.query-max-spill-per-node=50GB
experimental.reserved-pool-enabled=false
experimental.spiller-max-used-space-threshold=0.7
experimental.spiller-spill-path=/home/centos/

jmx.rmiregistry.port=10100
jmx.rmiserver.port=10101

# Configs only for the coordinator of a Prestissmo cluster
# Without this, you might see errors like "Scalar function name not registered: presto.default.$operator$hash_code, called with arguments: (BIGINT)."
optimizer.optimize-hash-generation=false
# Without this, decimal columns are treated as complex type while it should be scalar type. You will see errors like "scalar type has no children"
# Another error we may observe is "Chunked http transferring encoding is not supported." https://github.ibm.com/lakehouse/velox/issues/74
use-alternative-function-signatures=true
inline-sql-functions=false
offset-clause-enabled=true
# We set this to true when comparing with DB2 https://github.ibm.com/lakehouse/velox/issues/10
# parse-decimal-literals-as-double=true
regex-library=RE2J
# For VeloxRuntimeError: NULL_IF not supported in specialForm https://github.ibm.com/lakehouse/velox/issues/59
native-execution-enabled=true
# Enable sidecar by default
coordinator-sidecar-enabled=true
presto.default-namespace=native.default
plugin.dir=/opt/presto-server/native-plugin/

  7. misiugodfrey commented on Oct 23, 2025

    @misiugodfrey
    misiugodfrey
    on Oct 23, 2025
    Contributor

    This should now all be handled via pbench in the generate_presto_config.sh script.

  8. misiugodfrey
    closed this as completedon Oct 23, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions