[Presto] Presto parameter configuration optimization

Preface

Former company presto The service has not been very stable , With resource group , Configuration optimization , Code secondary development .presto The service has basically stabilized . This record presto What optimization has been done to service parameter configuration .

The production environment here JVM Are configured as 40G.

config.properties

1. Disable reservation pool

Reserved Pool : When you have one worker Of General Pool Out of memory ,reserved pool Will work . This is the time coordinator The cluster will select the query that consumes the most memory , And assign the query to all worker Of reserved pool.

Reserved Pool The value is equal to the query.max-total-memory-per-node Size . query.max-total-memory-per-node Is a single node User memory and system memory Maximum occupancy .
and Reserved Pool It is reserved by the system at startup , Moving together will occupy , So now there's a question , In reality , We seldom use Reserved Pool, And it also takes up a lot of memory resources .
So we can configure parameters

experimental.reserved-pool-enabled=false

2. query.max-memory-per-node

In a single worker The largest one that can be used above user memory value （ Default JVM max memory * 0.1） To adjust to JVM max memory * 0.25

query.max-memory-per-node=10GB

3. query.max-total-memory-per-node
   Single Query In a single Worker The maximum allowed on the user memory + system memory To adjust to JVM max memory * 0.4

query.max-total-memory-per-node=16GB

4. memory.heap-headroom-per-node
   This memory is mainly the memory allocation of third-party libraries , Unable to track statistics . （ Default JVM max memory * 0.3） To adjust to JVM max memory * 0.2

memory.heap-headroom-per-node=8GB

5. query.max-memory
   The maximum memory that a single query can use instantly on all task scheduling nodes （ Peak memory for a single query ） To adjust to <= query.max-total-memory-per-node * workers * 0.8

query.max-memory=192GB

6. query.low-memory-killer.policy
   When presto Occurrence cluster OOM Memory protection strategy when , To configure total-reservation yes kill Lose the task that occupies the most memory .

query.low-memory-killer.policy=total-reservation

7. More configuration

# Extend the waiting time 
exchange.http-client.request-timeout=10s
# From the other Presto The number of threads that the node obtains data . Higher values can improve the performance of large clusters or clusters with high concurrency ( The default value is ：25)
exchange.client-threads=50
exchange.http-client.idle-timeout = 10s
# Parallel operator （ For example, join and aggregate ） Default local concurrency . Lower values are better for clusters that run many queries at the same time （ The default value is ：16, Must be 2 The power of ）
task.concurrency=2
task.max-worker-threads=60
# You can create for processing HTTP The maximum number of threads responding . On a cluster with a large number of concurrent queries or on a cluster with hundreds or thousands of workers , It can be raised （ The default value is :100）
task.http-response-threads=200
join-distribution-type=AUTOMATIC
node-scheduler.max-splits-per-node=200
query.max-stage-count=400

# Automatically kill Run longer than 20 Minutes of sql（ Replace script kill Mission , Give Way presto  Automatic management ）：
query.max-run-time=1200s
# Solve the new version  remote too large  Report errors  
exchange.http-client.max-content-length=128MB
node-manager.http-client.max-content-length=64MB

jvm.config

-Xmx40G
-Xms40G
-XX:-UseBiasedLocking
-XX:+UseG1GC
-XX:+ExplicitGCInvokesConcurrent
-XX:+HeapDumpOnOutOfMemoryError
-XX:+UseGCOverheadLimit
-XX:OnOutOfMemoryError=kill -9 %p
-DHADOOP_USER_NAME=hive
-Duser.timezone=Asia/Shanghai
-Djdk.attach.allowAttachSelf=true
-Djava.security.krb5.conf=/etc/krb5.conf
-XX:G1ReservePercent=15
-XX:InitiatingHeapOccupancyPercent=40
-XX:ConcGCThreads=8