Support for extended memory on custom machine type

[vincent-octo]

Hi,

I have a workload that requires VMs of 1 CPU and 20 GB of RAM.

When using dsub, if I use --provider google-v2 --min-cores 1 --min-ram 20 then it will spawn a custom machine with 4 vCPUs and 20 GB of RAM. Ideally I would like to have it spawn a custom machine with 1 vCPU and 20 GB of RAM using extended memory, as it reduces the cost a bit.

Is there support for extended memory in dsub? If so, how to set it properly on the command line?

[mbookman]

Hi @vincent-octo !

I'm not sure if this is supported by the Pipelines API. If it is, this would be done by specifying a custom machine type as documented here:

https://cloud.google.com/genomics/reference/rest/Shared.Types/Metadata#VirtualMachine

The machine type of the virtual machine to create. Must be the short name of a standard machine type (such as "n1-standard-1") or a custom machine type (such as "custom-1-4096", where "1" indicates the number of vCPUs and "4096" indicates the memory in MB).

Please try using the --machine-type flag with a custom value. If this does not work, then we will need to pass a feature request on to the Cloud Health team.

[vincent-octo]

Hi @mbookman and thanks for your reply.

I didn't find a mention of extended memory in the documentation you provided. However, the GCP pricing calculator names this instance type n1-custom-1-20480-extended.

I tried it with dsub but it results in the following error:

dsub --provider google-v2 ... --machine-type n1-custom-1-20480-extended --command 'echo "Hello "' --wait

Job: dsub-xxx--xxx--190830-135045-31 ... googleapiclient.errors.HttpError: <HttpError 400 when requesting https://genomics.googleapis.com/v2alpha1/pipelines:run?alt=json returned "Error: validating pipeline: invalid machine type: unknown machine type">

Like you said, the API doesn't seem to support extended memory. So I will use custom VMs withtout extended memory (the price difference is not that big). Feel free to close this issue.

[mbookman]

Thank-you for testing it out. I have filed a feature request to have support in the Pipelines API. Will leave this issue open for tracking.

[mbookman]

Hi @vincent-octo .

The Pipelines API team has added initial support for custom machine types. For N1 machine types, you can use the notation:

--machine-type custom-CORES-MEMORY-ext

Support for the N2 machine types has not been added yet.

[vincent-octo]

Hi @mbookman ,

The custom VM with extended memory seems to be working from my initial test.

Just a note: the memory should be put in MiB and be a multiple of 256MiB. Otherwise it will result in the following error:

FAILURE   Execution failed: creating instance: inserting instance: Invalid value for field 'resource.machineType': 'zones/europe-west1-c/machineTypes/custom-1-20-ext'.
Memory should be a multiple of 256MiB, while 20MiB is requested. Memory size for 1 vCPU instance should be between 1024MiB and 6656MiB, while 20MiB is requested

Thanks for your help.