Merge branch 'release/0.8' into xiongyf/add-release-note-0.8

.github/workflows/build-image.yml

@@ -24,9 +24,9 @@ jobs:
     strategy:
       matrix:
         include:
-        - name: cuda11.8
-          dockerfile: cuda11.8
-          tags: superbench/main:cuda11.8
+        - name: cuda12.1
+          dockerfile: cuda12.1
+          tags: superbench/main:cuda12.1
         - name: cuda11.1.1
           dockerfile: cuda11.1.1
           tags: superbench/main:cuda11.1.1,superbench/superbench:latest

dockerfile/cuda11.8.dockerfile → dockerfile/cuda12.1.dockerfile

@@ -1,18 +1,18 @@
-FROM nvcr.io/nvidia/pytorch:22.12-py3
+FROM nvcr.io/nvidia/pytorch:23.03-py3
 
 # OS:
 #   - Ubuntu: 20.04
 #   - OpenMPI: 4.1.5a1
 #   - Docker Client: 20.10.8
 # NVIDIA:
-#   - CUDA: 11.8.0
-#   - cuDNN: 8.7.0.84
-#   - NCCL: v2.15.5-1
+#   - CUDA: 12.1.0
+#   - cuDNN: 8.8.1.3
+#   - NCCL: v2.17.1-1
 # Mellanox:
-#   - OFED: 5.2-2.2.3.0
-#   - HPC-X: v2.8.3
+#   - OFED: 5.2-2.2.3.0 # TODO
+#   - HPC-X: v2.14
 # Intel:
-#   - mlc: v3.9a
+#   - mlc: v3.10
 
 LABEL maintainer="SuperBench"
 
@@ -71,37 +71,27 @@ RUN mkdir -p /root/.ssh && \
 # Install OFED
 ENV OFED_VERSION=5.2-2.2.3.0
 RUN cd /tmp && \
-    wget -q http://content.mellanox.com/ofed/MLNX_OFED-${OFED_VERSION}/MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu20.04-x86_64.tgz && \
+    wget -q https://content.mellanox.com/ofed/MLNX_OFED-${OFED_VERSION}/MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu20.04-x86_64.tgz && \
     tar xzf MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu20.04-x86_64.tgz && \
     MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu20.04-x86_64/mlnxofedinstall --user-space-only --without-fw-update --force --all && \
     rm -rf /tmp/MLNX_OFED_LINUX-${OFED_VERSION}*
 
 # Install HPC-X
+ENV HPCX_VERSION=v2.14
 RUN cd /opt && \
     rm -rf hpcx && \
-    wget -q https://azhpcstor.blob.core.windows.net/azhpc-images-store/hpcx-v2.8.3-gcc-MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu20.04-x86_64.tbz && \
-    tar xf hpcx-v2.8.3-gcc-MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu20.04-x86_64.tbz && \
-    ln -s hpcx-v2.8.3-gcc-MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu20.04-x86_64 hpcx && \
-    rm hpcx-v2.8.3-gcc-MLNX_OFED_LINUX-${OFED_VERSION}-ubuntu20.04-x86_64.tbz
+    wget -q https://content.mellanox.com/hpc/hpc-x/${HPCX_VERSION}/hpcx-${HPCX_VERSION}-gcc-MLNX_OFED_LINUX-5-ubuntu20.04-cuda12-gdrcopy2-nccl2.17-x86_64.tbz -O hpcx.tbz && \
+    tar xf hpcx.tbz && \
+    mv hpcx-${HPCX_VERSION}-gcc-MLNX_OFED_LINUX-5-ubuntu20.04-cuda12-gdrcopy2-nccl2.17-x86_64 hpcx && \
+    rm hpcx.tbz
 
 # Install Intel MLC
 RUN cd /tmp && \
-    wget -q https://downloadmirror.intel.com/736634/mlc_v3.9a.tgz -O mlc.tgz && \
+    wget -q https://downloadmirror.intel.com/763324/mlc_v3.10.tgz -O mlc.tgz && \
     tar xzf mlc.tgz Linux/mlc && \
     cp ./Linux/mlc /usr/local/bin/ && \
     rm -rf ./Linux mlc.tgz
 
-ENV PATH="${PATH}" \
-    LD_LIBRARY_PATH="/usr/local/lib:${LD_LIBRARY_PATH}" \
-    SB_HOME=/opt/superbench \
-    SB_MICRO_PATH=/opt/superbench \
-    ANSIBLE_DEPRECATION_WARNINGS=FALSE \
-    ANSIBLE_COLLECTIONS_PATH=/usr/share/ansible/collections
-
-RUN echo PATH="$PATH" > /etc/environment && \
-    echo LD_LIBRARY_PATH="$LD_LIBRARY_PATH" >> /etc/environment && \
-    echo SB_MICRO_PATH="$SB_MICRO_PATH" >> /etc/environment
-
 # Install AOCC compiler
 RUN cd /tmp && \
     wget https://download.amd.com/developer/eula/aocc-compiler/aocc-compiler-4.0.0_1_amd64.deb && \
@@ -115,6 +105,18 @@ RUN cd /tmp && \
     mv amd-blis /opt/AMD && \
     rm -rf aocl-blis-linux-aocc-4.0.tar.gz
 
+
+ENV PATH="${PATH}" \
+    LD_LIBRARY_PATH="/usr/local/lib:${LD_LIBRARY_PATH}" \
+    SB_HOME=/opt/superbench \
+    SB_MICRO_PATH=/opt/superbench \
+    ANSIBLE_DEPRECATION_WARNINGS=FALSE \
+    ANSIBLE_COLLECTIONS_PATH=/usr/share/ansible/collections
+
+RUN echo PATH="$PATH" > /etc/environment && \
+    echo LD_LIBRARY_PATH="$LD_LIBRARY_PATH" >> /etc/environment && \
+    echo SB_MICRO_PATH="$SB_MICRO_PATH" >> /etc/environment
+
 # Add config files
 ADD dockerfile/etc /opt/microsoft/
 

docs/user-tutorial/benchmarks/micro-benchmarks.md

@@ -180,11 +180,11 @@ Performed by [High-Performance Linpack Benchmark for Distributed-Memory Computer
 
 #### Metrics
 
-| Name                | Unit               | Description                                                                |
-|---------------------|--------------------|----------------------------------------------------------------------------|
-| cpu-hpl/tests_pass  |                    | HPL completed running and correctness test has passed (1: pass, 0: fail).  |
-| cpu-hpl/throughput  | bandwidth (GFlops) | Compute bandwidth.                                                         |
-| cpu-hpl/time        | time (s)           | Time elapsed during HPL run.                                               |
+| Name               | Unit               | Description                                                               |
+|--------------------|--------------------|---------------------------------------------------------------------------|
+| cpu-hpl/tests_pass |                    | HPL completed running and correctness test has passed (1: pass, 0: fail). |
+| cpu-hpl/throughput | bandwidth (GFlops) | Compute bandwidth.                                                        |
+| cpu-hpl/time       | time (s)           | Time elapsed during HPL run.                                              |
 
 ### `cpu-stream`
 
@@ -216,13 +216,13 @@ performed by [Intel MLC Tool](https://www.intel.com/content/www/us/en/developer/
 
 | Name                                                                    | Unit             | Description                                                         |
 |-------------------------------------------------------------------------|------------------|---------------------------------------------------------------------|
-| cpu-memory-bw-latency/mem\_bandwidth\_matrix\_numa\_[0-9]+\_[0-9]+\_bw  | bandwidth (GB/s) | Former NUMA to latter NUMA memory bandwidth.                        |
-| cpu-memory-bw-latency/mem\_bandwidth\_matrix\_numa\_[0-9]+\_[0-9]+\_lat | time (us)        | Former NUMA to latter NUMA memory latency.                          |
-| cpu-memory-bw-latency/mem\_max\_bandwidth\_all\_reads\_bw               | bandwidth (GB/s) | Whole-CPU maximum memory bandwidth, full read.                      |
-| cpu-memory-bw-latency/mem\_max\_bandwidth\_3_1\_reads-writes\_bw        | bandwidth (GB/s) | Whole-CPU maximum memory bandwidth, read : write = 3 : 1.           |
-| cpu-memory-bw-latency/mem\_max\_bandwidth\_2_1\_reads-writes\_bw        | bandwidth (GB/s) | Whole-CPU maximum memory bandwidth, read : write = 2 : 1.           |
-| cpu-memory-bw-latency/mem\_max\_bandwidth\_1_1\_reads-writes\_bw        | bandwidth (GB/s) | Whole-CPU maximum memory bandwidth, read : write = 1 : 1.           |
-| cpu-memory-bw-latency/mem\_max\_bandwidth\_stream-triad\_like\_bw       | bandwidth (GB/s) | Whole-CPU maximum memory bandwidth, with stream-triad like pattern. |
+| cpu-memory-bw-latency/mem\_bandwidth\_matrix\_numa\_[0-9]+\_[0-9]+\_bw  | bandwidth (MB/s) | Former NUMA to latter NUMA memory bandwidth.                        |
+| cpu-memory-bw-latency/mem\_bandwidth\_matrix\_numa\_[0-9]+\_[0-9]+\_lat | time (ns)        | Former NUMA to latter NUMA memory latency.                          |
+| cpu-memory-bw-latency/mem\_max\_bandwidth\_all\_reads\_bw               | bandwidth (MB/s) | Whole-CPU maximum memory bandwidth, full read.                      |
+| cpu-memory-bw-latency/mem\_max\_bandwidth\_3_1\_reads-writes\_bw        | bandwidth (MB/s) | Whole-CPU maximum memory bandwidth, read : write = 3 : 1.           |
+| cpu-memory-bw-latency/mem\_max\_bandwidth\_2_1\_reads-writes\_bw        | bandwidth (MB/s) | Whole-CPU maximum memory bandwidth, read : write = 2 : 1.           |
+| cpu-memory-bw-latency/mem\_max\_bandwidth\_1_1\_reads-writes\_bw        | bandwidth (MB/s) | Whole-CPU maximum memory bandwidth, read : write = 1 : 1.           |
+| cpu-memory-bw-latency/mem\_max\_bandwidth\_stream-triad\_like\_bw       | bandwidth (MB/s) | Whole-CPU maximum memory bandwidth, with stream-triad like pattern. |
 
 ### `mem-bw`
 

superbench/benchmarks/model_benchmarks/model_base.py

@@ -78,6 +78,13 @@ def add_parser_arguments(self):
             required=False,
             help='The number of batch size.',
         )
+        self._parser.add_argument(
+            '--num_workers',
+            type=int,
+            default=8,
+            required=False,
+            help='Number of subprocesses to use for data loading.',
+        )
         self._parser.add_argument(
             '--precision',
             type=Precision,

superbench/benchmarks/model_benchmarks/pytorch_base.py

@@ -181,7 +181,7 @@ def _init_dataloader(self):
             dataset=self._dataset,
             batch_size=self._args.batch_size,
             shuffle=False,
-            num_workers=8,
+            num_workers=self._args.num_workers,
             sampler=train_sampler,
             drop_last=True,
             pin_memory=self._args.pin_memory

superbench/common/utils/gen_traffic_pattern_config.py

@@ -182,15 +182,14 @@ def gen_traffic_pattern_host_groups(host_list, pattern, mpi_pattern_path, benchm
         logger.error('Unsupported traffic pattern: {}'.format(pattern.type))
     host_groups = __convert_config_to_host_group(config, host_list)
     # write traffic pattern host groups to specified path
-    if pattern.mpi_pattern:
-        with open(mpi_pattern_path, 'a') as f:
-            f.write('benchmark_name: {} pattern_type: {}'.format(benchmark_name, pattern.type) + '\n')
-            for host_group in host_groups:
-                row = []
-                for host_list in host_group:
-                    group = ','.join(host_list)
-                    row.append(group)
-                group = ';'.join(row)
-                f.write(group + '\n')
-            f.write('\n')
+    with open(mpi_pattern_path, 'a') as f:
+        f.write('benchmark_name: {} pattern_type: {}'.format(benchmark_name, pattern.type) + '\n')
+        for host_group in host_groups:
+            row = []
+            for host_list in host_group:
+                group = ','.join(host_list)
+                row.append(group)
+            group = ';'.join(row)
+            f.write(group + '\n')
+        f.write('\n')
     return host_groups

tests/benchmarks/model_benchmarks/test_model_base.py

@@ -167,6 +167,7 @@ def test_arguments_related_interfaces():
   --no_gpu              Disable GPU training.
   --num_steps int       The number of test step.
   --num_warmup int      The number of warmup step.
+  --num_workers int     Number of subprocesses to use for data loading.
   --pin_memory          Enable option to pin memory in data loader.
   --precision Precision [Precision ...]
                         Model precision. E.g. fp8_hybrid fp8_e4m3 fp8_e5m2
@@ -206,6 +207,7 @@ def test_preprocess():
   --no_gpu              Disable GPU training.
   --num_steps int       The number of test step.
   --num_warmup int      The number of warmup step.
+  --num_workers int     Number of subprocesses to use for data loading.
   --pin_memory          Enable option to pin memory in data loader.
   --precision Precision [Precision ...]
                         Model precision. E.g. fp8_hybrid fp8_e4m3 fp8_e5m2