Merge pull request #289 from etiennedub/sharding_doc

Sharding GPU support

Author: cmd-ntrf

aws/infrastructure.tf

@@ -193,10 +193,11 @@ locals {
       prefix      = values.prefix
       tags        = values.tags
       specs = {
-        cpus = data.aws_ec2_instance_type.instance_type[values.prefix].default_vcpus
-        ram  = data.aws_ec2_instance_type.instance_type[values.prefix].memory_size
-        gpus = try(one(data.aws_ec2_instance_type.instance_type[values.prefix].gpus).count, 0)
-        mig  = lookup(values, "mig", null)
+        cpus   = data.aws_ec2_instance_type.instance_type[values.prefix].default_vcpus
+        ram    = data.aws_ec2_instance_type.instance_type[values.prefix].memory_size
+        gpus   = try(one(data.aws_ec2_instance_type.instance_type[values.prefix].gpus).count, 0)
+        mig    = lookup(values, "mig", null)
+        shard  = lookup(values, "shard", null)
       }
       volumes = contains(keys(module.design.volume_per_instance), x) ? {
         for pv_key, pv_values in var.volumes:

azure/infrastructure.tf

@@ -156,10 +156,11 @@ locals {
       prefix    = values.prefix
       tags      = values.tags
       specs = {
-        cpus = local.vmsizes[values.type].vcpus
-        ram  = local.vmsizes[values.type].ram
-        gpus = local.vmsizes[values.type].gpus
-        mig  = lookup(values, "mig", null)
+        cpus   = local.vmsizes[values.type].vcpus
+        ram    = local.vmsizes[values.type].ram
+        gpus   = local.vmsizes[values.type].gpus
+        mig    = lookup(values, "mig", null)
+        shard  = lookup(values, "shard", null)
       }
       volumes = contains(keys(module.design.volume_per_instance), x) ? {
         for pv_key, pv_values in var.volumes:

docs/README.md

@@ -540,6 +540,7 @@ the operating system and service software
     ```
     This is only functional with [MIG supported GPUs](https://docs.nvidia.com/datacenter/tesla/mig-user-guide/index.html#supported-gpus),
     and with x86-64 processors (see [NVIDIA/mig-parted issue #30](https://github.com/NVIDIA/mig-parted/issues/30)).
+6. `shard`: total number of [Sharding](https://slurm.schedmd.com/gres.html#Sharding) on the node. Sharding allows sharing the same GPU on multiple jobs. The total number of shards is evenly distributed across all GPUs on the node.
 
 For some cloud providers, it possible to define additional attributes.
 The following sections present the available attributes per provider.

gcp/infrastructure.tf

@@ -166,10 +166,11 @@ locals {
       prefix    = values.prefix
       tags      = values.tags
       specs = {
-        cpus = data.external.machine_type[values["prefix"]].result["vcpus"]
-        ram  = data.external.machine_type[values["prefix"]].result["ram"]
-        gpus = try(data.external.machine_type[values["prefix"]].result["gpus"], lookup(values, "gpu_count", 0))
-        mig  = lookup(values, "mig", null)
+        cpus   = data.external.machine_type[values["prefix"]].result["vcpus"]
+        ram    = data.external.machine_type[values["prefix"]].result["ram"]
+        gpus   = try(data.external.machine_type[values["prefix"]].result["gpus"], lookup(values, "gpu_count", 0))
+        mig    = lookup(values, "mig", null)
+        shard  = lookup(values, "shard", null)
       }
       volumes = contains(keys(module.design.volume_per_instance), x) ? {
         for pv_key, pv_values in var.volumes:

openstack/infrastructure.tf

@@ -121,13 +121,14 @@ locals {
       prefix    = values.prefix
       tags      = values.tags
       specs = {
-        cpus = data.openstack_compute_flavor_v2.flavors[values.prefix].vcpus
-        ram  = data.openstack_compute_flavor_v2.flavors[values.prefix].ram
-        gpus = sum([
+        cpus   = data.openstack_compute_flavor_v2.flavors[values.prefix].vcpus
+        ram    = data.openstack_compute_flavor_v2.flavors[values.prefix].ram
+        gpus   = sum([
           parseint(lookup(data.openstack_compute_flavor_v2.flavors[values.prefix].extra_specs, "resources:VGPU", "0"), 10),
           parseint(split(":", lookup(data.openstack_compute_flavor_v2.flavors[values.prefix].extra_specs, "pci_passthrough:alias", "gpu:0"))[1], 10)
         ])
-        mig  = lookup(values, "mig", null)
+        mig    = lookup(values, "mig", null)
+        shard  = lookup(values, "shard", null)
       }
       volumes = contains(keys(module.design.volume_per_instance), x) ? {
         for pv_key, pv_values in var.volumes: