[FEATURE] Restore memory based partitions #261
Assignees
Comments
cartalla
changed the title
|
Will you be adding slurm constraints (features), so I would use would use If I wanted a 32G machine and spot? Something like that? Or are you thinking of modifying the pcluster files so the memory is the physical memory of the machine so I can use --mem? |
|
Upon further investigation, ParallelCluster doesn’t allow me to add custom Features to the node configuration. So, I’ll be restoring the partitions that previously existed so you can specify that to get, for example, an od-4-gb machine or a sp-8-gb machine. I also already have spot and on-demand partitions that allow you to specify the purchase option you want to use.
From: Guntram Wolski ***@***.***>
Reply-To: aws-samples/aws-eda-slurm-cluster ***@***.***>
Date: Wednesday, October 16, 2024 at 1:05 PM
To: aws-samples/aws-eda-slurm-cluster ***@***.***>
Cc: "Carter, Allan" ***@***.***>, Assign ***@***.***>
Subject: Re: [aws-samples/aws-eda-slurm-cluster] [FEATURE] Restore memory based partitions (Issue #261)
Will you be adding slurm constraints (features), so I would use would use
--constraint=MEM32,SPOT
If I wanted a 32G machine and spot? Something like that?
Or are you thinking of modifying the pcluster files so the memory is the physical memory of the machine so I can use --mem?
—
Reply to this email directly, view it on GitHub<#261 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AFY4I5EVYJGX4NEM5F56BM3Z32TD7AVCNFSM6AAAAABQB4KX7CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIMJXGU2DGMRTGQ>.
You are receiving this because you were assigned.Message ID: ***@***.***>
|
|
That’s what I wanted to do, but ParallelCluster doesn’t allow me to customize the node’s features.
See this PR for what I’ve done. It will restore the queues that you previously had.
#263
From: Guntram Wolski ***@***.***>
Reply-To: aws-samples/aws-eda-slurm-cluster ***@***.***>
Date: Wednesday, October 16, 2024 at 1:05 PM
To: aws-samples/aws-eda-slurm-cluster ***@***.***>
Cc: "Carter, Allan" ***@***.***>, Assign ***@***.***>
Subject: Re: [aws-samples/aws-eda-slurm-cluster] [FEATURE] Restore memory based partitions (Issue #261)
Will you be adding slurm constraints (features), so I would use would use
--constraint=MEM32,SPOT
If I wanted a 32G machine and spot? Something like that?
Or are you thinking of modifying the pcluster files so the memory is the physical memory of the machine so I can use --mem?
—
Reply to this email directly, view it on GitHub<#261 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AFY4I5EVYJGX4NEM5F56BM3Z32TD7AVCNFSM6AAAAABQB4KX7CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDIMJXGU2DGMRTGQ>.
You are receiving this because you were assigned.Message ID: ***@***.***>
|
cartalla
added a commit
that referenced
this issue
Oct 16, 2024
Default Exclude only works correctly with default includes because it excludes instance types to keep the total instance types down. If user specifies any includes, then the default EDA excludes may exclude instance types that they are trying to include. Only use the default EDA includes and excludes if no includes or excludes are configured. Clean up the defaults to be less opinionated. The EDA configuration is opinionated and meant to be a starting point. But the defaults are wide open and will fail because they select too many instance types. So, the errors will give guidance on winnowing down the instance types. Resolves #262 Restore memory based partitions. Related to #235. Create partitions that include the purchase option (sp|od) and the amount of instance memory. This maintains backward compantibility for those using partitions to select the purchase option and amount of total memory. Resolves #261
cartalla
added a commit
that referenced
this issue
Oct 16, 2024
Default Exclude only works correctly with default includes because it excludes instance types to keep the total instance types down. If user specifies any includes, then the default EDA excludes may exclude instance types that they are trying to include. Only use the default EDA includes and excludes if no includes or excludes are configured. Clean up the defaults to be less opinionated. The EDA configuration is opinionated and meant to be a starting point. But the defaults are wide open and will fail because they select too many instance types. So, the errors will give guidance on winnowing down the instance types. Resolves #262 Restore memory based partitions. Related to #235. Create partitions that include the purchase option (sp|od) and the amount of instance memory. This maintains backward compantibility for those using partitions to select the purchase option and amount of total memory. Resolves #261 Add UseOnDemand configuration option that defaults to true so that you could create a cluster with only spot instances. Currently you always had to configure on-demand instances and only spot was optional. Make both purchase options optional, but require at least one.
cartalla
added a commit
that referenced
this issue
Oct 16, 2024
Default Exclude only works correctly with default includes because it excludes instance types to keep the total instance types down. If user specifies any includes, then the default EDA excludes may exclude instance types that they are trying to include. Only use the default EDA includes and excludes if no includes or excludes are configured. Clean up the defaults to be less opinionated. The EDA configuration is opinionated and meant to be a starting point. But the defaults are wide open and will fail because they select too many instance types. So, the errors will give guidance on winnowing down the instance types. Resolves #262 Restore memory based partitions. Related to #235. Create partitions that include the purchase option (sp|od) and the amount of instance memory. This maintains backward compantibility for those using partitions to select the purchase option and amount of total memory. Resolves #261 Add UseOnDemand configuration option that defaults to true so that you could create a cluster with only spot instances. Currently you always had to configure on-demand instances and only spot was optional. Make both purchase options optional, but require at least one.
cartalla
added a commit
that referenced
this issue
Oct 16, 2024
Default Exclude only works correctly with default includes because it excludes instance types to keep the total instance types down. If user specifies any includes, then the default EDA excludes may exclude instance types that they are trying to include. Only use the default EDA includes and excludes if no includes or excludes are configured. Clean up the defaults to be less opinionated. The EDA configuration is opinionated and meant to be a starting point. But the defaults are wide open and will fail because they select too many instance types. So, the errors will give guidance on winnowing down the instance types. Resolves #262 Restore memory based partitions. Related to #235. Create partitions that include the purchase option (sp|od) and the amount of instance memory. This maintains backward compantibility for those using partitions to select the purchase option and amount of total memory. Resolves #261 Add UseOnDemand configuration option that defaults to true so that you could create a cluster with only spot instances. Currently you always had to configure on-demand instances and only spot was optional. Make both purchase options optional, but require at least one.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Is your feature request related to a problem? Please describe.
PR #235 changed the way that AWS ParallelCluster (APC) partitions and compute resources (CRs) were created and named.
Previously, a CR could contain multiple instance types with the same number of cores and amount of memory.
The CR name was (od|sp)-\d+-gb-[\d+]-cores.
CRs with the same amount of memory were grouped into a single APC queue.
The queue name was (od|sp)-\d+-gb.
The new partition naming convention is (od|sp)--. For example, od-c7i-4xl.
Users prefer to have memory-based partitions that use total memory, not available memory and then constrain the number of required cores.
Form older version, the partitions would look like:
So a user could submit a job like the following:
Note that the current version does support specifying memory and cores.
Proposal
Configure the memory and purchase model partitions.
Add constraints for on-demand and spot to give a non-partition-based way of selecting the purchase model.
Also add constraint on nodes with the amount of total memory, not available memory.
The text was updated successfully, but these errors were encountered: