Azure HDInsight Interactive Query Cluster (Hive LLAP) sizing guide

This document describes the sizing of the HDInsight Interactive Query Cluster (Hive LLAP cluster) for a typical workload to achieve reasonable performance. Note that the recommendations provided in this document are generic guidelines and specific workloads may need  specific tuning.

Azure Default VM Types for HDInsight Interactive Query Cluster(LLAP)

Node Type Instance Size
Head D13 v2 8 vcpus, 56 GB RAM, 400 GB SSD
Worker D14 v2 16 vcpus, 112 GB RAM, 800 GB SSD
ZooKeeper A4 v2 4 vcpus, 8-GB RAM, 40 GB SSD

Note: All recommended configurations values are based on D14 v2 type worker node

Configuration:

Configuration Key Recommended value Description
yarn.nodemanager.resource.memory-mb 102400 (MB) Total memory given, in MB, for all YARN containers on a node
yarn.scheduler.maximum-allocation-mb 102400 (MB) The maximum allocation for every container request at the RM, in MBs. Memory requests higher than this value won't take effect
yarn.scheduler.maximum-allocation-vcores 12 The maximum number of CPU cores for every container request at the Resource Manager. Requests higher than this value won't take effect.
yarn.nodemanager.resource.cpu-vcores 12 Number of CPU cores per NodeManager that can be allocated for containers.
yarn.scheduler.capacity.root.llap.capacity 85 (%) YARN capacity allocation for llap queue
tez.am.resource.memory.mb 4096 (MB) The amount of memory in MB to be used by the tez AppMaster
hive.server2.tez.sessions.per.default.queue <number_of_worker_nodes> The number of sessions for each queue named in the hive.server2.tez.default.queues. This number corresponds to number of query coordinators(Tez AMs)
hive.tez.container.size 4096 (MB) Specified Tez container size in MB
hive.llap.daemon.num.executors 19 Number of executors per LLAP daemon
hive.llap.io.threadpool.size 19 Thread pool size for executors
hive.llap.daemon.yarn.container.mb 81920 (MB) Total memory in MB used by individual LLAP daemons (Memory per daemon)
hive.llap.io.memory.size 242688 (MB) Cache size in MB per LLAP daemon provided SSD cache is enabled
hive.auto.convert.join.noconditionaltask.size 2048 (MB) memory size in MB to do Map Join

LLAP Architecture/Components:

`LLAP Architecture/Components`

LLAP Daemon size estimations:

1. Determining total YARN memory allocation for all containers on a node

Configuration: yarn.nodemanager.resource.memory-mb

This value indicates a maximum sum of memory in MB that can be used by the YARN containers on each node. The value specified should be lesser than the total amount of physical memory on that node.
Total memory for all YARN containers on a node = (Total physical memory – memory for OS + Other services)
Set this value to ~90% of the available RAM size.
For D14 v2, the recommended value is 102400 MB

2. Determining maximum amount of memory per YARN container request

Configuration: yarn.scheduler.maximum-allocation-mb

This value indicates the maximum allocation for every container request at the Resource Manager, in MB. Memory requests higher than the specified value won't take effect. The Resource Manager can give memory to containers in increments of yarn.scheduler.minimum-allocation-mb and can't exceed the size specified by yarn.scheduler.maximum-allocation-mb. The value specified shouldn't be more than the total given memory for all containers on the node specified by yarn.nodemanager.resource.memory-mb.
For D14 v2 worker nodes, the recommended value is 102400 MB

3. Determining maximum amount of vcores per YARN container request

Configuration: yarn.scheduler.maximum-allocation-vcores

This value indicates the maximum number of virtual CPU cores for every container request at the Resource Manager. Requesting a higher number of vcores than this value won't take effect. It's a global property of the YARN scheduler. For LLAP daemon container, this value can be set to 75% of total available vcores. The remaining 25% should be reserved for NodeManager, DataNode, and other services running on the worker nodes.
There are 16 vcores on D14 v2 VMs and 75% of total 16 vcores can be used by LLAP daemon container.
For D14 v2, the recommended value is 12.

4. Number of concurrent queries

Configuration: hive.server2.tez.sessions.per.default.queue

This configuration value determines the number of Tez sessions that can be launched in parallel. These Tez sessions will be launched for each of the queues specified by "hive.server2.tez.default.queues". It corresponds to the number of Tez AMs (Query Coordinators). It's recommended to be the same as the number of worker nodes. The number of Tez AMs can be higher than the number of LLAP daemon nodes. The Tez AM's primary responsibility is to coordinate the query execution and assign query plan fragments to corresponding LLAP daemons for execution. Keep this value as multiple of a number of LLAP daemon nodes to achieve higher throughput.

Default HDInsight cluster has four LLAP daemons running on four worker nodes, so the recommended value is 4.

Ambari UI slider for Hive config variable hive.server2.tez.sessions.per.default.queue:

`LLAP maximum concurrent queries`

5. Tez Container and Tez Application Master size

Configuration: tez.am.resource.memory.mb, hive.tez.container.size

tez.am.resource.memory.mb - defines the Tez Application Master size.
The recommended value is 4096 MB.

hive.tez.container.size - defines the amount of memory given for Tez container. This value must be set between the YARN minimum container size(yarn.scheduler.minimum-allocation-mb) and the YARN maximum container size(yarn.scheduler.maximum-allocation-mb). The LLAP daemon executors use this value for limiting memory usage per executor.
The recommended value is 4096 MB.

6. LLAP Queue capacity allocation

Configuration: yarn.scheduler.capacity.root.llap.capacity

This value indicates a percentage of capacity given to llap queue. The capacity allocations may have different values for different workloads depending on how the YARN queues are configured. If your workload is read-only operations, then setting it as high as 90% of the capacity should work. However, if your workload is mix of update/delete/merge operations using managed tables, it's recommended to give 85% of the capacity for llap queue. The remaining 15% capacity can be used by other tasks such as compaction etc. to allocate containers from default queue. That way tasks in default queue won't deprive of YARN resources.

For D14v2 worker nodes, the recommended value for llap queue is 85.
(For readonly workloads, it can be increased up to 90 as suitable.)

7. LLAP daemon container size

Configuration: hive.llap.daemon.yarn.container.mb

LLAP daemon is run as a YARN container on each worker node. The total memory size for LLAP daemon container depends on following factors,

  • Configurations of YARN container size (yarn.scheduler.minimum-allocation-mb, yarn.scheduler.maximum-allocation-mb, yarn.nodemanager.resource.memory-mb)
  • Number of Tez AMs on a node
  • Total memory configured for all containers on a node and LLAP queue capacity

Memory needed by Tez Application Masters(Tez AM) can be calculated as follows.
Tez AM acts as a query coordinator and the number of Tez AMs should be configured based on a number of concurrent queries to be served. Theoretically, we can consider one Tez AM per worker node. However, its possible that you may see more than one Tez AM on a worker node. For calculation purpose, we assume uniform distribution of Tez AMs across all LLAP daemon nodes/worker nodes. It's recommended to have 4 GB of memory per Tez AM.

Number of Tez Ams = value specified by Hive config hive.server2.tez.sessions.per.default.queue.
Number of LLAP daemon nodes = specified by env variable num_llap_nodes_for_llap_daemons in Ambari UI.
Tez AM container size = value specified by Tez config tez.am.resource.memory.mb.

Tez AM memory per node = ( ceil ( Number of Tez AMs / Number of LLAP daemon nodes ) x Tez AM container size )
For D14 v2, the default configuration has four Tez AMs and four LLAP daemon nodes.
Tez AM memory per node = (ceil(4/4) x 4 GB) = 4 GB

Total Memory available for LLAP queue per worker node can be calculated as follows:
This value depends on the total amount of memory available for all YARN containers on a node(yarn.nodemanager.resource.memory-mb) and the percentage of capacity configured for llap queue(yarn.scheduler.capacity.root.llap.capacity).
Total memory for LLAP queue on worker node = Total memory available for all YARN containers on a node x Percentage of capacity for llap queue.
For D14 v2, this value is (100 GB x 0.85) = 85 GB.

The LLAP daemon container size is calculated as follows;

LLAP daemon container size = (Total memory for LLAP queue on a workernode) – (Tez AM memory per node) - (Service Master container size)
There is only one Service Master (Application Master for LLAP service) on the cluster spawned on one of the worker nodes. For calculation purpose, we consider one service master per worker node.
For D14 v2 worker node, HDI 4.0 - the recommended value is (85 GB - 4 GB - 1 GB)) = 80 GB
(For HDI 3.6, recommended value is 79 GB because you should reserve additional ~2 GB for slider AM.)

8. Determining number of executors per LLAP daemon

Configuration: hive.llap.daemon.num.executors, hive.llap.io.threadpool.size

hive.llap.daemon.num.executors:
This configuration controls the number of executors that can execute tasks in parallel per LLAP daemon. This value depends on the number of vcores, the amount of memory used per executor, and the amount of total memory available for LLAP daemon container. The number of executors can be oversubscribed to 120% of available vcores per worker node. However, it should be adjusted if it doesn't meet the memory requirements based on memory needed per executor and the LLAP daemon container size.

Each executor is equivalent to a Tez container and can consume 4GB(Tez container size) of memory. All executors in LLAP daemon share the same heap memory. With the assumption that not all executors run memory intensive operations at the same time, you can consider 75% of Tez container size(4 GB) per executor. This way you can increase the number of executors by giving each executor less memory (e.g. 3 GB) for increased parallelism. However, it is recommended to tune this setting for your target workload.

There are 16 vcores on D14 v2 VMs. For D14 v2, the recommended value for num of executors is (16 vcores x 120%) ~= 19 on each worker node considering 3GB per executor.

hive.llap.io.threadpool.size:
This value specifies the thread pool size for executors. Since executors are fixed as specified, it will be same as number of executors per LLAP daemon.
For D14 v2, the recommended value is 19.

9. Determining LLAP daemon cache size

Configuration: hive.llap.io.memory.size

LLAP daemon container memory consists of following components;

  • Head room
  • Heap memory used by executors (Xmx)
  • In-memory cache per daemon (its off-heap memory size, not applicable when SSD cache is enabled)
  • In-memory cache metadata size (applicable only when SSD cache is enabled)

Headroom size: This size indicates a portion of off-heap memory used for Java VM overhead (metaspace, threads stack, gc data structures, etc.). Generally, this overhead is about 6% of the heap size (Xmx). To be on the safer side, this value can be calculated as 6% of total LLAP daemon memory size.
For D14 v2, the recommended value is ceil(80 GB x 0.06) ~= 4 GB.

Heap size(Xmx): It is amount of heap memory available for all executors. Total Heap size = Number of executors x 3 GB
For D14 v2, this value is 19 x 3 GB = 57 GB

Ambari environment variable for LLAP heap size:

`LLAP heap size`

When SSD cache is disabled, the in-memory cache is amount of memory that is left after taking out Headroom size and Heap size from the LLAP daemon container size.

Cache size calculation differs when SSD cache is enabled. Setting hive.llap.io.allocator.mmap = true will enable SSD caching. When SSD cache is enabled, some portion of the memory will be used to store metadata for the SSD cache. The metadata is stored in memory and it's expected to be ~8% of SSD cache size.
SSD Cache in-memory metadata size = LLAP daemon container size - (Head room + Heap size)
For D14 v2, with HDI 4.0, SSD cache in-memory metadata size = 80 GB - (4 GB + 57 GB) = 19 GB
For D14 v2, with HDI 3.6, SSD cache in-memory metadata size = 79 GB - (4 GB + 57 GB) = 18 GB

Given the size of available memory for storing SSD cache metadata, we can calculate the size of SSD cache that can be supported.
Size of in-memory metadata for SSD cache = LLAP daemon container size - (Head room + Heap size) = 19 GB
Size of SSD cache = size of in-memory metadata for SSD cache(19 GB) / 0.08 (8 percent)

For D14 v2 and HDI 4.0, the recommended SSD cache size = 19 GB / 0.08 ~= 237 GB
For D14 v2 and HDI 3.6, the recommended SSD cache size = 18 GB / 0.08 ~= 225 GB

10. Adjusting Map Join memory

Configuration: hive.auto.convert.join.noconditionaltask.size

Make sure you have hive.auto.convert.join.noconditionaltask enabled for this parameter to take effect. This configuration determine the threshold for MapJoin selection by Hive optimizer that considers oversubscription of memory from other executors to have more room for in-memory hash tables to allow more map join conversions. Considering 3GB per executor, this size can be oversubscribed to 3GB, but some heap memory may also be used for sort buffers, shuffle buffers, etc. by the other operations.
So for D14 v2, with 3 GB memory per executor, it's recommended to set this value to 2048 MB.

(Note: This value may need adjustments that are suitable for your workload. Setting this value too low may not use autoconvert feature. And setting it too high may result into out of memory exceptions or GC pauses that can result into adverse performance.)

11. Number of LLAP daemons

Ambari environment variables: num_llap_nodes, num_llap_nodes_for_llap_daemons

num_llap_nodes - specifies number of nodes used by Hive LLAP service, this includes nodes running LLAP daemon, LLAP Service Master, and Tez Application Master(Tez AM).

`Number of Nodes for LLAP service`

num_llap_nodes_for_llap_daemons - specified number of nodes used only for LLAP daemons. LLAP daemon container sizes are set to max fit node, so it will result in one llap daemon on each node.

`Number of Nodes for LLAP daemons`

It's recommended to keep both values same as number of worker nodes in Interactive Query cluster.

Considerations for Workload Management

If you want to enable workload management for LLAP, make sure you reserve enough capacity for workload management to function as expected. The workload management requires configuration of a custom YARN queue, which is in addition to llap queue. Make sure you divide total cluster resource capacity between llap queue and workload management queue in accordance to your workload requirements. Workload management spawns Tez Application Masters(Tez AMs) when a resource plan is activated. Please note:

  • Tez AMs spawned by activating a resource plan consume resources from the workload management queue as specified by hive.server2.tez.interactive.queue.
  • The number of Tez AMs would depend on the value of QUERY_PARALLELISM specified in the resource plan.
  • Once the workload management is active, Tez AMs in llap queue will not used. Only Tez AMs from workload management queue are used for query coordination. Tez AMs in the llap queue are used when workload management is disabled.

For example: Total cluster capacity = 100 GB memory, divided between LLAP, Workload Management, and Default queues as follows:

  • llap queue capacity = 70 GB
  • Workload management queue capacity = 20 GB
  • Default queue capacity = 10 GB

With 20 GB in workload management queue capacity, a resource plan can specify QUERY_PARALLELISM value as five, which means workload management can launch five Tez AMs with 4 GB container size each. If QUERY_PARALLELISM is higher than the capacity, you may see some Tez AMs stop responding in ACCEPTED state. The Hiveserver2 Interactive cannot submit query fragments to the Tez AMs that are not in RUNNING state.

Next Steps

If setting these values didn't resolve your issue, visit one of the following...