Scheduler Configuration

The Aurora scheduler can take a variety of configuration options through command-line arguments. Examples are available under examples/scheduler/. For a list of available Aurora flags and their documentation, see Scheduler Configuration Reference.

A Note on Configuration

Like Mesos, Aurora uses command-line flags for runtime configuration. As such the Aurora “configuration file” is typically a scheduler.sh shell script of the form.

#!/bin/bash
AURORA_HOME=/usr/local/aurora-scheduler

# Flags controlling the JVM.
JAVA_OPTS=(
  -Xmx2g
  -Xms2g
  # GC tuning, etc.
)

# Flags controlling the scheduler.
AURORA_FLAGS=(
  # Port for client RPCs and the web UI
  -http_port=8081
  # Log configuration, etc.
)

# Environment variables controlling libmesos
export JAVA_HOME=...
export GLOG_v=1
export LIBPROCESS_PORT=8083
export LIBPROCESS_IP=192.168.33.7

JAVA_OPTS="${JAVA_OPTS[*]}" exec "$AURORA_HOME/bin/aurora-scheduler" "${AURORA_FLAGS[@]}"

That way Aurora’s current flags are visible in ps and in the /vars admin endpoint.

JVM Configuration

JVM settings are dependent on your environment and cluster size. They might require custom tuning. As a starting point, we recommend:

  • Ensure the initial (-Xms) and maximum (-Xmx) heap size are idential to prevent heap resizing at runtime.
  • Either -XX:+UseConcMarkSweepGC or -XX:+UseG1GC -XX:+UseStringDeduplication are sane defaults for the garbage collector.
  • -Djava.net.preferIPv4Stack=true makes sense in most cases as well.

Network Configuration

By default, Aurora binds to all interfaces and auto-discovers its hostname. To reduce ambiguity it helps to hardcode them though:

-http_port=8081
-ip=192.168.33.7
-hostname="aurora1.us-east1.example.org"

Two environment variables control the ip and port for the communication with the Mesos master and for the replicated log used by Aurora:

export LIBPROCESS_PORT=8083
export LIBPROCESS_IP=192.168.33.7

It is important that those can be reached from all Mesos master and Aurora scheduler instances.

Replicated Log Configuration

Aurora schedulers use ZooKeeper to discover log replicas and elect a leader. Only one scheduler is leader at a given time - the other schedulers follow log writes and prepare to take over as leader but do not communicate with the Mesos master. Either 3 or 5 schedulers are recommended in a production deployment depending on failure tolerance and they must have persistent storage.

Below is a summary of scheduler storage configuration flags that either don’t have default values or require attention before deploying in a production environment.

-native_log_quorum_size

Defines the Mesos replicated log quorum size. In a cluster with N schedulers, the flag -native_log_quorum_size should be set to floor(N/2) + 1. So in a cluster with 1 scheduler it should be set to 1, in a cluster with 3 it should be set to 2, and in a cluster of 5 it should be set to 3.

Number of schedulers (N) -native_log_quorum_size setting (floor(N/2) + 1)
1 1
3 2
5 3
7 4

Incorrectly setting this flag will cause data corruption to occur!

-native_log_file_path

Location of the Mesos replicated log files. For optimal and consistent performance, consider allocating a dedicated disk (preferably SSD) for the replicated log. Ensure that this disk is not used by anything else (e.g. no process logging) and in particular that it is a real disk and not just a partition.

Even when a dedicated disk is used, switching from CFQ to deadline I/O scheduler of Linux kernel can furthermore help with storage performance in Aurora (see this ticket for details).

-native_log_zk_group_path

ZooKeeper path used for Mesos replicated log quorum discovery.

See code for other available Mesos replicated log configuration options and default values.

Changing the Quorum Size

Special care needs to be taken when changing the size of the Aurora scheduler quorum. Since Aurora uses a Mesos replicated log, similar steps need to be followed as when changing the Mesos quorum size.

As a preparation, increase -native_log_quorum_size on each existing scheduler and restart them. When updating from 3 to 5 schedulers, the quorum size would grow from 2 to 3.

When starting the new schedulers, use the -native_log_quorum_size set to the new value. Failing to first increase the quorum size on running schedulers can in some cases result in corruption or truncating of the replicated log used by Aurora. In that case, see the documentation on recovering from backup.

Backup Configuration

Configuration options for the Aurora scheduler backup manager.

  • -backup_interval: The interval on which the scheduler writes local storage backups. The default is every hour.
  • -backup_dir: Directory to write backups to. As stated above, this should not be co-located on the same disk as the replicated log.
  • -max_saved_backups: Maximum number of backups to retain before deleting the oldest backup(s).

Resource Isolation

For proper CPU, memory, and disk isolation as mentioned in our enduser documentation, we recommend to add the following isolators to the --isolation flag of the Mesos agent:

  • cgroups/cpu
  • cgroups/mem
  • disk/du

In addition, we recommend to set the following agent flags:

  • --cgroups_limit_swap to enable memory limits on both memory and swap instead of just memory. Alternatively, you could disable swap on your agent hosts.
  • --cgroups_enable_cfs to enable hard limits on CPU resources via the CFS bandwidth limiting feature.
  • --enforce_container_disk_quota to enable disk quota enforcement for containers.

To enable the optional GPU support in Mesos, please see the GPU related flags in the Mesos configuration. To enable the corresponding feature in Aurora, you have to start the scheduler with the flag

-allow_gpu_resource=true

If you want to use revocable resources, first follow the Mesos oversubscription documentation and then set set this Aurora scheduler flag to allow receiving revocable Mesos offers:

-receive_revocable_resources=true

Both CPUs and RAM are supported as revocable resources. The former is enabled by the default, the latter needs to be enabled via:

-enable_revocable_ram=true

Unless you want to use the default tier configuration, you will also have to specify a file path:

-tier_config=path/to/tiers/config.json

Multi-Framework Setup

Aurora holds onto Mesos offers in order to provide efficient scheduling and preemption. This is problematic in multi-framework environments as Aurora might starve other frameworks.

With a downside of increased scheduling latency, Aurora can be configured to be more cooperative:

  • Lowering -min_offer_hold_time (e.g. to 1mins) can ensure unused offers are returned back to Mesos more frequently.
  • Increasing -offer_filter_duration (e.g to 30secs) will instruct Mesos not to re-offer rejected resources for the given duration.

Setting a minimum amount of resources for each Mesos role can furthermore help to ensure no framework is starved entirely.

Containers

Both the Mesos and Docker containerizers require configuration of the Mesos agent.

Mesos Containerizer

The minimal agent configuration requires to enable Docker and Appc image support for the Mesos containerizer:

--containerizers=mesos
--image_providers=appc,docker
--isolation=filesystem/linux,docker/runtime  # as an addition to your other isolators

Further details can be found in the corresponding Mesos documentation.

Docker Containerizer

The Docker containerizer requires the Docker engine is installed on each agent host. In addition, it must be enabled on the Mesos agents by launching them with the option:

--containerizers=mesos,docker

If you would like to run a container with a read-only filesystem, it may also be necessary to use the scheduler flag -thermos_home_in_sandbox in order to set HOME to the sandbox before the executor runs. This will make sure that the executor/runner PEX extractions happens inside of the sandbox instead of the container filesystem root.

If you would like to supply your own parameters to docker run when launching jobs in docker containers, you may use the following flags:

-allow_docker_parameters
-default_docker_parameters

-allow_docker_parameters controls whether or not users may pass their own configuration parameters through the job configuration files. If set to false (the default), the scheduler will reject jobs with custom parameters. NOTE: this setting should be used with caution as it allows any job owner to specify any parameters they wish, including those that may introduce security concerns (privileged=true, for example).

-default_docker_parameters allows a cluster operator to specify a universal set of parameters that should be used for every container that does not have parameters explicitly configured at the job level. The argument accepts a multimap format:

-default_docker_parameters="read-only=true,tmpfs=/tmp,tmpfs=/run"

Common Options

The following Aurora options work for both containerizers.

A scheduler flag, -global_container_mounts allows mounting paths from the host (i.e the agent machine) into all containers on that host. The format is a comma separated list of hostpath:containerpath[:mode] tuples. For example -global_container_mounts=/opt/secret_keys_dir:/mnt/secret_keys_dir:ro mounts /opt/secret_keys_dir from the agents into all launched containers. Valid modes are ro and rw.

Thermos Process Logs

Log destination

By default, Thermos will write process stdout/stderr to log files in the sandbox. Process object configuration allows specifying alternate log file destinations like streamed stdout/stderr or suppression of all log output. Default behavior can be configured for the entire cluster with the following flag (through the -thermos_executor_flags argument to the Aurora scheduler):

--runner-logger-destination=both

both configuration will send logs to files and stream to parent stdout/stderr outputs.

See Configuration Reference for all destination options.

Log rotation

By default, Thermos will not rotate the stdout/stderr logs from child processes and they will grow without bound. An individual user may change this behavior via configuration on the Process object, but it may also be desirable to change the default configuration for the entire cluster. In order to enable rotation by default, the following flags can be applied to Thermos (through the -thermos_executor_flags argument to the Aurora scheduler):

--runner-logger-mode=rotate
--runner-rotate-log-size-mb=100
--runner-rotate-log-backups=10

In the above example, each instance of the Thermos runner will rotate stderr/stdout logs once they reach 100 MiB in size and keep a maximum of 10 backups. If a user has provided a custom setting for their process, it will override these default settings.

Thermos Executor Wrapper

If you need to do computation before starting the Thermos executor (for example, setting a different --announcer-hostname parameter for every executor), then the Thermos executor should be invoked inside a wrapper script. In such a case, the aurora scheduler should be started with -thermos_executor_path pointing to the wrapper script and -thermos_executor_resources set to a comma separated string of all the resources that should be copied into the sandbox (including the original Thermos executor). Ensure the wrapper script does not access resources outside of the sandbox, as when the script is run from within a Docker container those resources may not exist.

For example, to wrap the executor inside a simple wrapper, the scheduler will be started like this -thermos_executor_path=/path/to/wrapper.sh -thermos_executor_resources=/usr/share/aurora/bin/thermos_executor.pex

Custom Executors

The scheduler can be configured to utilize a custom executor by specifying the -custom_executor_config flag. The flag must be set to the path of a valid executor configuration file.

For more information on this feature please see the custom executors documentation.

A note on increasing executor overhead

Increasing executor overhead on an existing cluster, whether it be for custom executors or for Thermos, will result in degraded preemption performance until all task which began life with the previous executor configuration with less overhead are preempted/restarted.

Controlling MTTA via Update Affinity

When there is high resource contention in your cluster you may experience noticably elevated job update times, as well as high task churn across the cluster. This is due to Aurora’s first-fit scheduling algorithm. To alleviate this, you can enable update affinity where the Scheduler will make a best-effort attempt to reuse the same agent for the updated task (so long as the resources for the job are not being increased).

To enable this in the Scheduler, you can set the following options:

-enable_update_affinity=true
-update_affinity_reservation_hold_time=3mins

You will need to tune the hold time to match the behavior you see in your cluster. If you have extremely high update throughput, you might have to extend it as processing updates could easily add significant delays between scheduling attempts. You may also have to tune scheduling parameters to achieve the throughput you need in your cluster. Some relevant settings (with defaults) are:

-max_schedule_attempts_per_sec=40
-initial_schedule_penalty=1secs
-max_schedule_penalty=1mins
-scheduling_max_batch_size=3
-max_tasks_per_schedule_attempt=5

There are metrics exposed by the Scheduler which can provide guidance on where the bottleneck is. Example metrics to look at:

- schedule_attempts_blocks (if this number is greater than 0, then task throughput is hitting
                            limits controlled by --max_scheduler_attempts_per_sec)
- scheduled_task_penalty_* (metrics around scheduling penalties for tasks, if the numbers here are high
                            then you could have high contention for resources)

Most likely you’ll run into limits with the number of update instances that can be processed per minute before you run into any other limits. So if your total work done per minute starts to exceed 2k instances, you may need to extend the updateaffinityreservationholdtime.

Cluster Maintenance

Aurora performs maintenance related task drains. One of the scheduler options that can control how often the scheduler polls for maintenance work can be controlled via,

-host_maintenance_polling_interval=1min

Enforcing SLA limitations

Since tasks can specify their own SLAPolicy, the cluster needs to limit these SLA requirements. Too aggressive a requirement can permanently block any type of maintenance work (ex: OS/Kernel/Security upgrades) on a host and hold it hostage.

An operator can control the limits for SLA requirements via these scheduler configuration options:

-max_sla_duration_secs=2hrs
-min_required_instances_for_sla_check=20

Note: These limits only apply for CountSlaPolicy and PercentageSlaPolicy.

Limiting Coordinator SLA

With CoordinatorSlaPolicy the SLA calculation is off-loaded to an external HTTP service. Some relevant scheduler configuration options are,

-sla_coordinator_timeout=1min
-max_parallel_coordinated_maintenance=10

Since handing off the SLA calculation to an external service can potentially block maintenance on hosts for an indefinite amount of time (either due to a mis-configured coordinator or due to a valid degraded service). In those situations the following metrics will be helpful to identify the offending tasks.

sla_coordinator_user_errors_*     (counter tracking number of times the coordinator for the task
                                   returned a bad response.)
sla_coordinator_errors_*          (counter tracking number of times the scheduler was not able
                                   to communicate with the coordinator of the task.)
sla_coordinator_lock_starvation_* (counter tracking number of times the scheduler was not able to
                                   get the lock for the coordinator of the task.)