Skip to content

Configuring all_or_nothing_batch launches

Jump to bottom
chenwany edited this page Sep 27, 2022 · 5 revisions

Only applicable for pcluster>=v2.10.0 && scheduler == slurm.

Overview

Starting with AWS ParallelCluster version 2.10, you have additional control over instance launch behavior. By default, in versions <= 2.9, instances in a batch follow a soft launch (best-effort strategy), with at least some of the batch launching when some of the requested instances fail. However, with version 2.10, you can opt to configure instances in a batch to have a firm launch using an all-or-nothing launch behavior. As such, the entire batch fails if any one of the instances in the batch cannot be provisioned.

How to enable this feature

To enable the all-or-nothing launch behavior for dynamic nodes for firm launches, set all_or_nothing_batch = True in the ResumeProgram config at /etc/parallelcluster/slurm_plugin/parallelcluster_slurm_resume.conf.

To further configure instance launch behavior with respect to the specific requirements of your application, you can set the max_batch_size parameter. This parameter specifies the upper limit to the number of instances in your launch batch. The next section provides a more in-depth explanation for this parameter as well as several related considerations.

Note The default settings are all_or_nothing_batch = False and max_batch_size = 500. These settings ensure that instances in a batch have a soft launch with a smaller number of RunInstances calls.

The following example is a ResumeProgram config with the firm (all-or-nothing) launch behavior enabled:

$ cat /etc/parallelcluster/slurm_plugin/parallelcluster_slurm_resume.conf
# DO NOT MODIFY
[slurm_resume]
cluster_name = demo
region = us-east-2
proxy = NONE
dynamodb_table = parallelcluster-demo
hosted_zone = ZXXXXXXXXXXXXXXX
dns_domain = demo.pcluster
use_private_hostname = false
master_private_ip = 172.x.x.x
master_hostname = ip-172-x-x-x.us-east-2.compute.internal
# Options added:
all_or_nothing_batch = True      
max_batch_size = 100

Understanding the launch-batch and ResumeProgram workflow

With its improved Slurm integration added to AWS ParallelCluster in version 2.9, dynamic nodes in ParallelCluster use the Slurm cloud bursting plugin and rely on the Slurm integrated power saving mechanism (for more details you can consult our Slurm user guide).

The workflow to launch instances for dynamic nodes is as follows:

  1. A job is submitted to Slurm.
  2. Slurm allocates the job to multiple nodes, putting any node that is in the POWER_SAVED state into the POWER_UP state.
  3. Slurm calls ResumeProgram at a specified time internal to request nodes that need to be launched as part of the POWER_UP workflow. No job information is passed into ResumeProgram. Slurm can call ResumeProgram with nodes from multiple jobs but doesn’t pass node-job correspondence information into ResumeProgram.
  4. Slum calls ResumeProgram with a list of nodes to launch. The nodes are parsed from the input following the same order in which they are passed.
  5. For each of the instance types in a queue or partition, AWS ParallelCluster groups the nodes into batches based on their max_batch_size parameter configuration and uses the RunInstances call to launch these batches. Each call is referred to as a launch-batch.
  6. The RunInstances call also launches the nodes along with each of their corresponding instances. Any node that can’t be launched is placed in the POWER_DOWN state later reset to the POWER_SAVED state.
  7. When the launched instances finish the configuration process, slurmd is started on compute instance. At this time, the job can also begin to use the corresponding slurm node.

The all_or_nothing_batch parameter modifies how RunInstances calls behave. By default, RunInstances calls follow a soft launch strategy in which some but not necessarily all instances in a launch-batch are launched by the call. If all_or_nothing_batch = True is set, each RunInstances call follows a firm launch strategy. For this, all instances in a launch-batch must be launched. If the capacity for the entire launch-batch isn’t reached, then the RunInstances call fails and no instance is launched.

ResumeProgram workflow examples

  • Let's assume max_batch_size = 4 and Slurm calls ResumeProgram with the following list of nodes queue1-dy-c5xlarge-[1-5], queue1-dy-t2micro-2, queue2-dy-c5xlarge[1-3]. This results in the following:
    • For every instance type in every queue and partition, nodes are grouped in to launch-batches according to the max_batch_size = 4 configuration. Therefore, we need to have the following batches: queue1-dy-c5xlarge-[1-4], queue1-dy-c5xlarge-5, queue1-dy-t2micro-2, queue2-dy-c5xlarge-[1-3]. One RunInstances call is made for every launch-batch.
    • If all_or_nothing_batch = False is set, and RunInstances call for queue2-dy-c5xlarge-[1-3] can’t launch at full capacity, and can only launch one c5.xlarge instance, then queue2-dy-c5xlarge-1 is configured with the successfully launched instance, and queue2-dy-c5xlarge-[2-3] is marked down due to an unsuccessful launch for other instances.
    • If all_or_nothing_batch = True is set, and the RunInstances call for queue2-dy-c5xlarge-[1-3] can’t launch at full capacity, the entire RunInstances call fails and no instance is launched. queue2-dy-c5xlarge-[1-3] is marked down due to an unsuccessful launch of the instances.

Limits

The following are the limits on this feature:

  • Requests are executed on the basis of Slurm requests rather a job basis. Slurm groups multiple job requests that are submitted within close succession. This can sometimes result in failed an instance launch for groups of jobs even when the capacity is available to satisfy some (though not all) of the jobs.
    • Example: Three jobs are submitted in close succession. Job1 is allocated to queue1-dy-c5xlarge-[1-4], Job 2 is allocated to queue1-dy-c5xlarge-[5-6], and Job 3 is allocated to queue1-dy-c5xlarge-[7-8]. For this, max_batch_size = 10, all_or_nothing_batch = True are set. Slurm calls ResumeProgram to POWER_UP queue1-dy-c5xlarge-[1-8]. Because the max_batch_size = 10 is set, queue1-dy-c5xlarge-[1-8] is grouped into the same launch-batch. However, because of this, if the entire capacity can’t be launched, then queue1-dy-c5xlarge-[1-8] fails, causing all three jobs to fail.
    • This limit can be mitigated by spacing out job requests so each call to ResumeProgram only contains the nodes that are required for one job. We recommend waiting at least one minute between submitting two jobs, so that each call to ResumeProgram only contains nodes from one job each time.
  • Firm all-or-nothing scaling is achieved on a per-instance type basis. You can’t use multiple instance types for all-or-nothing scaling. If you want to run a job that requires multiple instance types (for example, the P3dn and C5n instances types), all-or-nothing scaling guarantee isn’t plausible.
    • *Example: *Job1 is allocated to queue1-dy-c5xlarge-[1-4],queue1-dy-t2micro-1, max_batch_size = 10, and all_or_nothing_batch = True is set. The launch-batches are queue1-dy-c5xlarge-[1-4] and queue1-dy-t2micro-1. These launch-batches are launched by two separate RunInstances calls.
  • All-or-nothing isn’t guaranteed if a node set allocated to a job isn’t contiguous.
    • Slurm most often calls ResumeProgram with an ordered node list, and ResumeProgram groups nodes into launch-batches based on the order they are passed in. As a result, launch-batches aren’t correlated correctly with node used by each job.
    • Example: Two jobs are submitted in close succession. If Job1 is allocated to queue1-dy-c5xlarge-[1, 3-4], Job 2 is allocated to queue1-dy-c5xlarge-[2, 5-6], and max_batch_size=3 is set. Slurm calls ResumeProgram with queue1-dy-c5xlarge-[1-6], and launch-batches are queue1-dy-c5xlarge-[1-3], and queue1-dy-c5xlarge-[4-6], which don’t correspond to the two jobs directly
    • Using --contiguous when submitting your jobs might help.
  • The maximum number of nodes that can be launched in a firm launch is limited to maximum number of instance-launches possible with a RunInstances call. This number is restricted by the limitations of the RunInstances API. By default, the API can allow up to 500 instance to be launched in a single RunInstances call (max_batch_size = 500). If you want to increase this limit, contact AWS customer support.
  • This functionality is only available as a per-cluster option, all queues would be subject to the same all_or_nothing_batch and max_batch_size settings.
  • If an instance fails during boot (for example an EC2 instance or a node fails), then the specific node that fails is automatically terminated. Then, Slurm reassigns the job to a new node or new set of nodes. In most cases, Slurm reuses the previously successful nodes and requests a replacement for the failed instance or instances. When this occurs, all-or-nothing scaling canot be guaranteed because the nodes running the job are from different launch-batches.
  • This feature is only available for dynamic nodes.
    • The ResumeProgram workflow is only applicable for dynamic nodes
    • Static nodes are managed separately by clustermgtd on a soft launch, best-effort basis.

Example use case for the all-or-nothing batches

With all-or-nothing batches, you can opt to configure a launch-batch to be launched with a soft launch or a firm launch. Note that all-or-nothing scaling can’t be guaranteed on a per-job basis because launch-batches don’t directly correspond to all the nodes required by a job.

The following example illustrates how all-or-nothing batches functions. Homogeneous jobs are submitted in a way that matches each launch batch with all the nodes required by a job. In practice, a majority of job submissions are heterogeneous in terms of instance type and the number of nodes required.

We recommend that, when you enable all-or-nothing batches, you set the batch size to the largest number of single-instance-type instances that is required for a job. This is so that your largest launches can all be completed in a all-or-nothing launch behavior.

Consider the following conditions when setting your batch size.

  • Large max_batch_size: Large launches can follow an all-or-nothing strategy. However, smaller launches for small jobs might be grouped together, and might collectively fail if there isn’t enough capacity for the entire group.
  • Small max_batch_size: Small launches can be separated into independent launch batches, but large launches might be separated into multiple launch batches, and an all-or-nothing launch strategy can’t be adapted use large launches.

Example Scenario

  • Consider the following scenario. The following details the initial state of a cluster. There is no job in the queues. Consider the scheduled2 queue. In this queue, there are 5,000 nodes. It is in the POWER_SAVING state.
$ sinfo
PARTITION   AVAIL  TIMELIMIT  NODES  STATE NODELIST 
scheduled1*    up   infinite     20  idle~ scheduled1-dy-m54xlarge-[1-20] 
scheduled2     up   infinite   5000  idle~ scheduled2-dy-c5xlarge-[1-5000] 
$ squeue
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
  • Submit jobs that require 50 nodes for each job and use all-or-nothing behavior for launch-batches.
  • Add the options all_or_nothing_batch = True to the ResumeProgram config located at /etc/parallelcluster/slurm_plugin/parallelcluster_slurm_resume.conf. This turns on the all-or-nothing launch behavior for each launch-batch
  • The default setting is max_batch_size = 500. Don't change this setting, but understand that, because the queue only contains one instance type, all the nodes in a queue are most likely grouped into one launch batch. Then, because of this, if the entire batch can’t be launched, this might also cause multiple small jobs to fail.
  • Because the largest number of single-instance-type instances that can be required for a job is 50 for this example, each launch batch can only correspond to 1 job only. For this, set max_batch_size = 50 to ResumeProgram config located at /etc/parallelcluster/slurm_plugin/parallelcluster_slurm_resume.conf. Note that ResumeProgram config will have other parameters already in place. Don’t modify those parameters.
$ cat /etc/parallelcluster/slurm_plugin/parallelcluster_slurm_resume.conf
# DO NOT MODIFY
[slurm_resume]
cluster_name = scheduled
region = us-east-2
proxy = NONE
dynamodb_table = parallelcluster-scheduled
hosted_zone = ZXXXXXXXXXXXXXXX
dns_domain = scheduled.pcluster
use_private_hostname = false
master_private_ip = 172.x.x.x
master_hostname = ip-172-x-x-x.us-east-2.compute.internal
# Option added:
all_or_nothing_batch = True
max_batch_size = 50
  • Submit 20 “sleep 300” array jobs, each requiring 50 nodes, or 1000 nodes in total. Failed jobs in most cases are re-queued, but to clearly show which jobs failed because of unsuccessful launch, specify the ——no-requeue option so that jobs that failed are removed from the queue.
# -a 1-20: submit 20 array jobs
# -N 50: each job requires 50 nodes
# -p scheduled2: submit job to scheduled2 queue/partition
# --exclusive: allocate each node to a job completely, so no 2 jobs will share a node
# --no-requeue: no automatic requeue in case of failure
# --contiguous: allocated nodes must form a contiguous set
$ sbatch --wrap "sleep 300" -a 1-20 -N 50 -p scheduled2 --exclusive --no-requeue --contiguous
  • After the job is submitted, 1000 nodes total are assigned to jobs by Slurm. Because the nodes were initially in the POWER_SAVING state, Slurm places the nodes in the POWER_UP state and calls ResumeProgram to launch instances for the nodes.
# 1000 nodes placed into `POWER_UP` state
$ sinfo
PARTITION   AVAIL  TIMELIMIT  NODES  STATE NODELIST 
scheduled1*    up   infinite     20  idle~ scheduled1-dy-m54xlarge-[1-20] 
scheduled2     up   infinite   1000 alloc# scheduled2-dy-c5xlarge-[1-1000] 
scheduled2     up   infinite   4000  idle~ scheduled2-dy-c5xlarge-[1001-5000]

  • 1000 nodes are launched in batches of up to 50 by ResumeProgram. With all_or_nothing_batch = True set, each launch-batch of 50 nodes, the number of instances launched is also 50 withstanding that this capacity can be reached. Otherwise, the number of instances launched is lower if the capacity is lower. If there is no capacity, all instances for each launch-batch and the entire batch fail.

  • Because of capacity limitations, you can only launch 300 instances (6 batches of 50) for 300 nodes. The remaining 700 nodes (14 batches failed to launch) are reset and put back into POWER_DOWN.

# Only able to launch 300 nodes
# Failed 700 nodes placed into down automatically by `ResumeProgram`
$ sinfo
PARTITION   AVAIL  TIMELIMIT  NODES  STATE NODELIST 
scheduled1*    up   infinite     20  idle~ scheduled1-dy-m54xlarge-[1-20] 
scheduled2     up   infinite    300 alloc# scheduled2-dy-c5xlarge-[1-300] 
scheduled2     up   infinite    700  down# scheduled2-dy-c5xlarge-[301-1000] 
scheduled2     up   infinite   4000  idle~ scheduled2-dy-c5xlarge-[1001-5000]
  • Because of how the job was setup and because of the max_batch_size setting, the node set is allocated to jobs according to the launch-batches. Six jobs run successfully because there are six successful launch batch. The other jobs failed because their launch batch also failed.
# 6 job able to run, the rest failed and removed from queue because no-requeue
$ squeue
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON) 
             872_1 scheduled     wrap   ubuntu CF       1:25     50 scheduled2-dy-c5xlarge-[1-50] 
             872_2 scheduled     wrap   ubuntu CF       1:25     50 scheduled2-dy-c5xlarge-[51-100] 
             872_3 scheduled     wrap   ubuntu CF       1:25     50 scheduled2-dy-c5xlarge-[101-150] 
             872_4 scheduled     wrap   ubuntu CF       1:25     50 scheduled2-dy-c5xlarge-[151-200] 
             872_5 scheduled     wrap   ubuntu CF       1:25     50 scheduled2-dy-c5xlarge-[201-250] 
             872_6 scheduled     wrap   ubuntu CF       1:25     50 scheduled2-dy-c5xlarge-[251-300]
# After a bit of time, jobs are up and running, 700 failed to launch nodes in POWER_DOWN
$ squeue
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON) 
             872_5 scheduled     wrap   ubuntu  R       0:01     50 scheduled2-dy-c5xlarge-[201-250] 
             872_1 scheduled     wrap   ubuntu  R       1:01     50 scheduled2-dy-c5xlarge-[1-50] 
             872_2 scheduled     wrap   ubuntu  R       1:01     50 scheduled2-dy-c5xlarge-[51-100] 
             872_3 scheduled     wrap   ubuntu  R       1:01     50 scheduled2-dy-c5xlarge-[101-150] 
             872_4 scheduled     wrap   ubuntu  R       1:01     50 scheduled2-dy-c5xlarge-[151-200] 
             872_6 scheduled     wrap   ubuntu  R       1:01     50 scheduled2-dy-c5xlarge-[251-300] 
$ sinfo
PARTITION   AVAIL  TIMELIMIT  NODES  STATE NODELIST 
scheduled1*    up   infinite     20  idle~ scheduled1-dy-m54xlarge-[1-20] 
scheduled2     up   infinite    700  idle% scheduled2-dy-c5xlarge-[301-1000] 
scheduled2     up   infinite   4000  idle~ scheduled2-dy-c5xlarge-[1001-5000] 
scheduled2     up   infinite    300  alloc scheduled2-dy-c5xlarge-[1-300]

Summary

You can enable the all-or-nothing scaling behavior at launch-batch level by setting all_or_nothing_batch = True in the ResumeProgram config at /etc/parallelcluster/slurm_plugin/parallelcluster_slurm_resume.conf.

We recommend that, if you set all_or_nothing_batch = True, you should set max_batch_size to the largest number of single-instance-type instances that can be used for a job. As such, so your largest launches are made in all-or-nothing fashion.

You should consider the following conditions when setting your max_batch_size:

  • Large max_batch_size: Large launches can be made in all-or-nothing fashion. However, smaller launches for small jobs might be grouped together, and might fail collective if capacity for the entire group cannot be satisfied
  • Small max_batch_size: Small launches can be separated into even smaller, separate launch batches, but large launches might be separated into multiple launch batches, and all-or-nothing behavior will not apply to large launches.

ResumeProgram logs are located in /var/log/parallelcluster/slurm_resume.log, and might be helpful to debug and understand launch behavior. As a general rule, having control over all_or_nothing_batch and max_batch_size options in the ResumeProgram workflow can allow you to have greater control over how dynamic nodes are launched.

Clone this wiki locally