No-DNS config working + MPI over RDMA

Author: Matt Pryor

.github/workflows/publish-benchmark-images.yaml

@@ -15,6 +15,7 @@ jobs:
     strategy:
       matrix:
         include:
+          - component: discovery
           - component: fio
           - component: iperf
           - component: mpi-benchmarks

README.md

@@ -46,9 +46,6 @@ metadata:
 spec:
   # Indicates whether to use the host network or the pod network
   hostNetwork: true
-  # Indicates whether the client should access the server pod directly (false)
-  # or via a Kubernetes service (true)
-  serverService: false
   # The number of parallel streams to use
   streams: 8
   # The duration of the test
@@ -121,11 +118,9 @@ spec:
     # Permutations are calculated as a cross-product of the specified names and values
     product:
       hostNetwork: [true, false]
-      serverService: [true, false]
       streams: [1, 2, 4, 8, 16, 32, 64]
     # A list of explicit permutations to include
     explicit:
       - hostNetwork: true
-        serverService: false
         streams: 128
 ```

chart/templates/clusterrole.yaml

@@ -24,6 +24,7 @@ rules:
       - mpipingpongs.perftest.stackhpc.com
       - openfoams.perftest.stackhpc.com
       - rdmabandwidths.perftest.stackhpc.com
+      - rdmalatencies.perftest.stackhpc.com
     verbs:
       - update
       - patch
@@ -64,6 +65,14 @@ rules:
       - jobs
     verbs:
       - "*"
+  - apiGroups:
+      - "rbac.authorization.k8s.io"
+    resources:
+      - serviceaccounts
+      - roles
+      - rolebindings
+    verbs:
+      - "*"
   - apiGroups:
       - ""
     resources:
@@ -76,6 +85,7 @@ rules:
   - apiGroups:
       - ""
     resources:
+      - endpoints
       - pods
       - pods/log
     verbs:

examples/accelerated/iperf-set.yaml

@@ -0,0 +1,19 @@
+apiVersion: perftest.stackhpc.com/v1alpha1
+kind: BenchmarkSet
+metadata:
+  name: iperf-accelerated
+spec:
+  template:
+    apiVersion: perftest.stackhpc.com/v1alpha1
+    kind: IPerf
+    spec:
+      imagePullPolicy: Always
+      hostNetwork: false
+      networkName: mlnx-net
+      resources:
+        limits:
+          mellanox.com/mlnxnet: 1
+      duration: 30
+  permutations:
+    product:
+      streams: [1, 2, 4, 8, 16, 32, 64, 128]

examples/rdma-lat-set.yaml renamed to examples/accelerated/mpi-pingpong-set.yaml

@@ -1,13 +1,18 @@
 apiVersion: perftest.stackhpc.com/v1alpha1
 kind: BenchmarkSet
 metadata:
-  name: rdma-latency
+  name: mpi-pingpong-accelerated
 spec:
   template:
     apiVersion: perftest.stackhpc.com/v1alpha1
-    kind: RDMALatency
+    kind: MPIPingPong
     spec:
       imagePullPolicy: Always
+      hostNetwork: false
+      networkName: mlnx-net
+      resources:
+        limits:
+          mellanox.com/mlnxnet: 1
   permutations:
     product:
-      hostNetwork: [true]
+      transport: [TCP, RDMA]

examples/openfoam-set.yaml renamed to examples/accelerated/openfoam-set.yaml

@@ -1,21 +1,21 @@
 apiVersion: perftest.stackhpc.com/v1alpha1
 kind: BenchmarkSet
 metadata:
-  name: openfoam
+  name: openfoam-accelerated
 spec:
   template:
     apiVersion: perftest.stackhpc.com/v1alpha1
     kind: OpenFOAM
     spec:
       imagePullPolicy: Always
-      numProcs: 48
+      hostNetwork: false
+      networkName: mlnx-net
+      resources:
+        limits:
+          mellanox.com/mlnxnet: 1
       problemSize: S
+      numProcs: 48
   permutations:
-    # Produce a matrix of jobs with hostNetwork=false
     product:
-      hostNetwork: [false]
+      transport: [TCP, RDMA]
       numNodes: [2, 4, 8, 12, 16, 24, 48]
-    # Produce an explcit job for hostNetwork=true
-    explicit:
-      - hostNetwork: true
-        numNodes: 2

examples/accelerated/rdma-bw-set.yaml

@@ -0,0 +1,22 @@
+apiVersion: perftest.stackhpc.com/v1alpha1
+kind: BenchmarkSet
+metadata:
+  name: rdma-bandwidth-accelerated
+spec:
+  template:
+    apiVersion: perftest.stackhpc.com/v1alpha1
+    kind: RDMABandwidth
+    spec:
+      imagePullPolicy: Always
+      hostNetwork: false
+      networkName: mlnx-net
+      resources:
+        limits:
+          mellanox.com/mlnxnet: 1
+      qps: 512
+      iterations: 25
+      extraArgs:
+        - --tclass=96
+  permutations:
+    product:
+      mode: [read, write]

examples/accelerated/rdma-lat-set.yaml

@@ -0,0 +1,21 @@
+apiVersion: perftest.stackhpc.com/v1alpha1
+kind: BenchmarkSet
+metadata:
+  name: rdma-latency-accelerated
+spec:
+  template:
+    apiVersion: perftest.stackhpc.com/v1alpha1
+    kind: RDMALatency
+    spec:
+      imagePullPolicy: Always
+      hostNetwork: false
+      networkName: mlnx-net
+      resources:
+        limits:
+          mellanox.com/mlnxnet: 1
+      iterations: 25
+      extraArgs:
+        - --tclass=96
+  permutations:
+    product:
+      mode: [read, write]

examples/iperf-set.yaml renamed to examples/cni/iperf-set.yaml

@@ -1,16 +1,15 @@
 apiVersion: perftest.stackhpc.com/v1alpha1
 kind: BenchmarkSet
 metadata:
-  name: iperf
+  name: iperf-cni
 spec:
   template:
     apiVersion: perftest.stackhpc.com/v1alpha1
     kind: IPerf
     spec:
       imagePullPolicy: Always
+      hostNetwork: false
       duration: 30
   permutations:
     product:
-      hostNetwork: [true, false]
-      serverService: [true, false]
-      streams: [1, 2, 4, 8, 16, 32, 64]
+      streams: [1, 2, 4, 8, 16, 32, 64, 128]

examples/cni/mpi-pingpong.yaml

@@ -0,0 +1,8 @@
+apiVersion: perftest.stackhpc.com/v1alpha1
+kind: MPIPingPong
+metadata:
+  name: mpi-pingpong-cni
+spec:
+  imagePullPolicy: Always
+  hostNetwork: false
+  transport: TCP