You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: articles/cyclecloud/openpbs.md
+23-45Lines changed: 23 additions & 45 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -2,16 +2,16 @@
2
2
title: OpenPBS Integration
3
3
description: OpenPBS scheduler configuration in Azure CycleCloud.
4
4
author: adriankjohnson
5
-
ms.date: 07/29/2021
5
+
ms.date: 06/11/2025
6
6
ms.author: adjohnso
7
7
---
8
8
9
9
# OpenPBS
10
10
11
-
[//]: #(Need to link to the scheduler README on Github)
11
+
[//]: #(Need to link to the scheduler README on GitHub)
12
12
13
13
::: moniker range="=cyclecloud-7"
14
-
[OpenPBS](http://openpbs.org/) can easily be enabled on a CycleCloud cluster by modifying the "run_list" in the configuration section of your cluster definition. The two basic components of a PBS Professional cluster are the 'master' node which provides a shared filesystem on which the PBS Professional software runs, and the 'execute' nodes which are the hosts that mount the shared filesystem and execute the jobs submitted. For example, a simple cluster template snippet may look like:
14
+
[OpenPBS](http://openpbs.org/) can easily be enabled on a CycleCloud cluster by modifying the "run_list", in the configuration section of your cluster definition. A PBS Professional (PBS Pro) cluster has two main parts: the 'master' node, which runs the software on a shared filesystem, and the 'execute' nodes, which mount that filesystem and run the submitted jobs. For example, a simple cluster template snippet may look like:
15
15
16
16
```ini
17
17
[cluster my-pbspro]
@@ -31,7 +31,7 @@ ms.author: adjohnso
31
31
run_list = role[pbspro_execute_role]
32
32
```
33
33
34
-
Importing and starting a cluster with definition in CycleCloud will yield a single 'master' node. Execute nodes can be added to the cluster via the `cyclecloud add_node` command. For example, to add 10 more execute nodes:
34
+
Importing and starting a cluster with definition in CycleCloud yields a single 'master' node. Execute nodes can be added to the cluster via the `cyclecloud add_node` command. For example, to add 10 more execute nodes:
Cyclecloud maintains two resources to expand the dynamic provisioning capability. These resources are *nodearray* and *machinetype*.
43
43
44
-
If you submit a job and specify a nodearray resource by `qsub -l nodearray=highmem -- /bin/hostname`
45
-
then CycleCloud will add nodes to the nodearray named 'highmem'. If there is no such nodearray then the job will remain idle.
44
+
If you submit a job and specify a nodearray resource by `qsub -l nodearray=highmem -- /bin/hostname` then, CycleCloud adds nodes to the nodearray named 'highmem'. If no such nodearray exists, the job stays idle.
46
45
47
-
Similarly if a machinetype resource is specified which a job submission, e.g. `qsub -l machinetype:Standard_L32s_v2 my-job.sh`, then CycleCloud autoscales the 'Standard_L32s_v2' in the 'execute' (default) nodearray. If that machine type is not available in the 'execute' node array then the job will remain idle.
46
+
Similarly, if a machinetype resource is specified which a job submission, for example, `qsub -l machinetype:Standard_L32s_v2 my-job.sh`, then CycleCloud autoscales the 'Standard_L32s_v2' in the 'execute' (default) nodearray. If that machine type isn’t available in the 'execute' node array, the job stays idle.
Which autoscales only if the 'Standard_HB60rs' machines are specified in the 'hpc' node array.
54
54
55
-
which will autoscale only if the 'Standard_HB60rs' machines are specified an the 'hpc' node array.
55
+
## Adding extra queues assigned to nodearrays
56
56
57
-
## Adding additional queues assigned to nodearrays
58
-
59
-
On clusters with multiple nodearrays, it's common to create separate queues to automatically route jobs to the appropriate VM type. In this example, we'll assume the following "gpu" nodearray has been defined in your cluster template:
57
+
On clusters with multiple nodearrays, it's common to create separate queues to automatically route jobs to the appropriate VM type. In this example, we assume the following "gpu" nodearray is defined in your cluster template:
60
58
61
59
```bash
62
60
[[nodearray gpu]]
@@ -82,29 +80,25 @@ After importing the cluster template and starting the cluster, the following com
82
80
```
83
81
84
82
> [!NOTE]
85
-
> The above queue definition will pack all VMs in the queue into a single VM scale set to support MPI jobs. To define the queue for serial jobs and allow multiple VM Scalesets, set `ungrouped = true` for both `resources_default` and `default_chunk`. You can also set `resources_default.place = pack` if you want the scheduler to pack jobs onto VMs instead of round-robin allocation of jobs. For more information on PBS job packing, see the official [PBS Professional OSS documentation](https://www.altair.com/pbs-works-documentation/).
83
+
> As shown in the example, queue definition packs all VMs in the queue into a single VM scale set to support MPI jobs. To define the queue for serial jobs and allow multiple VM Scalesets, set `ungrouped = true` for both `resources_default` and `default_chunk`. You can also set `resources_default.place = pack` if you want the scheduler to pack jobs onto VMs instead of round-robin allocation of jobs. For more information on PBS job packing, see the official [PBS Professional OSS documentation](https://www.altair.com/pbs-works-documentation/).
86
84
87
85
## PBS Professional Configuration Reference
88
86
89
-
The following are the PBS Professional specific configuration options you can toggle to customize functionality:
87
+
The following are the PBS Professional(PBS Pro) specific configuration options you can toggle to customize functionality:
90
88
91
89
| PBS Pro Options | Description |
92
90
| --------------- | ----------- |
93
91
| pbspro.slots | The number of slots for a given node to report to PBS Pro. The number of slots is the number of concurrent jobs a node can execute, this value defaults to the number of CPUs on a given machine. You can override this value in cases where you don't run jobs based on CPU but on memory, GPUs, etc. |
94
-
| pbspro.slot_type | The name of type of 'slot' a node provides. The default is 'execute'. When a job is tagged with the hard resource `slot_type=<type>`, that job will*only*run on a machine of the same slot type. This allows you to create different software and hardware configurations per node and ensure an appropriate job is always scheduled on the correct type of node. |
95
-
| pbspro.version | Default: '18.1.3-0'. This is the PBS Professional version to install and run. This is currently the default and *only* option. In the future additional versions of the PBS Professional software may be supported. |
92
+
| pbspro.slot_type | The name of type of 'slot' a node provides. The default is 'execute'. When a job is tagged with the hard resource `slot_type=<type>`, that job runs*only* on the machine of the same slot type. It allows you to create a different software and hardware configurations per node and ensure an appropriate job is always scheduled on the correct type of node. |
93
+
| pbspro.version | Default: '18.1.3-0'. This is currently the default version and *only* option to install and run. This is currently the default version and *only* option. In the future more versions of the PBS Pro software may be supported. |
96
94
97
95
::: moniker-end
98
96
99
97
::: moniker range=">=cyclecloud-8"
100
98
101
99
## Connect PBS with CycleCloud
102
100
103
-
CycleCloud manages [OpenPBS](http://openpbs.org/) clusters through an installable agent called
104
-
[`azpbs`](https://github.com/Azure/cyclecloud-pbspro). This agent connect to
105
-
CycleCloud to read cluster and VM configurations and also integrates with OpenPBS
106
-
to effectively process the job and host information. All `azpbs` configurations
107
-
are found in the `autoscale.json` file, normally `/opt/cycle/pbspro/autoscale.json`.
101
+
CycleCloud manages [OpenPBS](http://openpbs.org/) clusters through an installable agent called [`azpbs`](https://github.com/Azure/cyclecloud-pbspro). This agent connects to CycleCloud to read cluster and VM configurations and also integrates with OpenPBS to effectively process the job and host information. All `azpbs` configurations are found in the `autoscale.json` file, normally `/opt/cycle/pbspro/autoscale.json`.
108
102
109
103
```
110
104
"password": "260D39rWX13X",
@@ -118,9 +112,7 @@ are found in the `autoscale.json` file, normally `/opt/cycle/pbspro/autoscale.js
118
112
119
113
### Important Files
120
114
121
-
The `azpbs` agent parses the PBS configuration each time it's called - jobs, queues, resources.
122
-
Information is provided in the stderr and stdout of the command as well as to a log file, both
123
-
at configurable levels. All PBS management commands (`qcmd`) with arguments are logged to file as well.
115
+
The `azpbs` agent parses the PBS configuration each time it's called - jobs, queues, resources. Information is provided in the stderr and stdout of the command and to a log file, both at configurable levels. All PBS management commands (`qcmd`) with arguments are logged to file as well.
124
116
125
117
All these files can be found in the _/opt/cycle/pbspro/_ directory where the agent is installed.
126
118
@@ -134,10 +126,7 @@ All these files can be found in the _/opt/cycle/pbspro/_ directory where the age
134
126
135
127
136
128
### Defining OpenPBS Resources
137
-
This project allows for a generally association of OpenPBS resources with Azure
138
-
VM resources via the cyclecloud-pbspro (azpbs) project. This resource relationship
139
-
defined in `autoscale.json`.
140
-
129
+
This project allows general association of OpenPBS resources with Azure VM resources via the cyclecloud-pbspro (azpbs) project. This resource relationship defined in `autoscale.json`.
141
130
The default resources defined with the cluster template we ship with are
142
131
143
132
```json
@@ -175,13 +164,9 @@ The default resources defined with the cluster template we ship with are
175
164
}
176
165
```
177
166
178
-
The OpenPBS resource named `mem` is equated to a node attribute named `node.memory`,
179
-
which is the total memory of any virtual machine. This configuration allows `azpbs`
180
-
to process a resource request such as `-l mem=4gb` by comparing the value of the
181
-
job resource requirements to node resources.
167
+
The OpenPBS resource named `mem` is equated to a node attribute named `node.memory`, which is the total memory of any virtual machine. This configuration allows `azpbs` to process a resource request such as `-l mem=4gb` by comparing the value of the job resource requirements to node resources.
182
168
183
-
Note that disk is currently hardcoded to `size::20g`.
184
-
Here is an example of handling VM Size specific disk size
169
+
Currently, disk size is hardcoded to `size::20g`. Here's an example of handling VM Size specific disk size
185
170
```json
186
171
{
187
172
"select": {"node.vm_size": "Standard_F2"},
@@ -197,21 +182,14 @@ Here is an example of handling VM Size specific disk size
197
182
198
183
### Autoscale and Scalesets
199
184
200
-
CycleCloud treats spanning and serial jobs differently in OpenPBS clusters.
201
-
Spanning jobs will land on nodes that are part of the same placement group. The
202
-
placement group has a particular platform meaning (VirtualMachineScaleSet with
203
-
SinglePlacementGroup=true) and CC will managed a named placement group for each
204
-
spanned node set. Use the PBS resource `group_id` for this placement group name.
185
+
CycleCloud treats spanning and serial jobs differently in OpenPBS clusters. Spanning jobs land on nodes that are part of the same placement group. The placement group has a particular platform meaning VirtualMachineScaleSet with SinglePlacementGroup=true) and CycleCloud manages a named placement group for each spanned node set. Use the PBS resource `group_id` for this placement group name.
205
186
206
-
The `hpc` queue appends
207
-
the equivalent of `-l place=scatter:group=group_id` by using native queue defaults.
187
+
The `hpc` queue appends the equivalent of `-l place=scatter:group=group_id` by using native queue defaults.
208
188
209
189
210
190
### Installing the CycleCloud OpenPBS Agent `azpbs`
211
191
212
-
The OpenPBS CycleCloud cluster will manage the installation and configuration of
213
-
the agent on the server node. The preparation includes setting PBS resources,
214
-
queues, and hooks. A scripted install can be done outside of CycleCloud as well.
192
+
The OpenPBS CycleCloud cluster manages the installation and configuration of the agent on the server node. The preparation includes setting PBS resources, queues, and hooks. A scripted install can be done outside of CycleCloud as well.
215
193
216
194
```bash
217
195
# Prerequisite: python3, 3.6 or newer, must be installed and in the PATH
0 commit comments