[gpu] performance and functionality improvements (#1265)

* [gpu] performance and functionality improvements

* Capturing disk usage statistics to reduce excessive disk space
* created exit handler to clean up environment on completion or failure
* created prepare function to prepare for the installation
* when sufficient memory is available, configure a ramdisk
* reduce noise by turning off -x in utility functions
* added descriptive comments before the obscurely coded
  compare_versions_lte and compare_versions_lt functions
* removed some intermediate driver versions
* added cuda url for 12.6
* execute_with_retries now logs on failure, captures runtime and
  cleans before installing on debian
* saving OS installation and NV .run files and their temp files to ramdisk
* piping source .xz file directly xz instead of saving to disk first
* new utility function "is_debuntu" checks for the frequently used
  conditon of whether the running OS is either debian or ubuntu
* added support for specifying an http proxy (thank you प्रकाश)
* moving load of kernel module to later in the code and exercising
  modprobe of all modules to avoid regression
* fixed problem with attempting to fetch from incorrect vault
  directory when rocky kernel package is not found in primary repo
* using correct cran-r signing key for ubuntu18
* corrected file check condition for /etc/apt/trusted.gpg

* do not update all packages on rocky ; move preparation to prepare function

* increasing memory to make use of ramdisk

* using something a little smaller

* create mount_ramdisk function and call it ; fix up the version comparison functions ; create ge and le comparisons for OSs

* iterating better, caching results of system calls ; renamed to repair_old_backports

* comparing correct version numbers

* rocky uses a tmpfs on /tmp in the base image

* tested on rocky and ubuntu

* tested harder on rocky

* cuda 11 no longer available for debian 12

* cuda v11 no longer supported on debian12

* corrected use of ubuntu regex for rocky version

* re-enabling spark job tests

* correct a couple of edge cases

* added instructions for manually running tests

* open a monitor session by default

* cleaning up cuda and cudnn url generation

* condition better

* cleaned up generation of NVIDIA_CUDA_URL

* updated versions and GPU accelerators in the documentation

* ensure this test to be skipped based on cuda version rather than dataproc version alone

* fix for /usr/local/cuda-12.4/bin/nvcc: No such file or directory

* correcting path to run-bazel-tests.sh

* runing variable definition

* cleaned up skip conditions

* order of operations

* works with 2.0-rocky8

* remove redundant conditional check

* supported version limits are tightened up a bit ; clean up rocky vault install code

* corrected syntax errors

* failure to run dnf here should not fail the entire installer

* order matters here

* 2.2-ubuntu22 works with cuda 11, other 2.2 do not

* 2.2-ubuntu22 works with cuda 11, other 2.2 do not

* fixes ubuntu22 kernel version mismatch error

* disabling rocky9 builds due to out of date base dataproc image

* cuda 2.0 not supported in debian12

* some 2.0-rocky8 single instance tests fail

* intended to use <= and not >=

* simplify gpu resource script

* setting default discoveryScript ; testing pyspark in its own function

* remove spark: prefix from property names

* comment out quite a few tests

* new version numbers

* fixed a syntax error with documentation

* musn't forget the commas

* half as many tasks with twice as much cpu and gpu each

* pause before first ssh ; correct variable name

gpu/Dockerfile

@@ -0,0 +1,40 @@
+# This Dockerfile builds the container from which manual tests are run
+# This process needs to be executed manually from a git clone
+#
+# See manual-test-runner.sh for instructions
+
+FROM gcr.io/cloud-builders/gcloud
+
+RUN useradd -m -d /home/ia-tests -s /bin/bash ia-tests
+
+# Installed here are packages on which the tests depend
+RUN apt-get -qq update \
+  && apt-get -y -qq install \
+     apt-transport-https apt-utils \
+     ca-certificates libmime-base64-perl gnupg \
+     curl jq less screen > /dev/null 2>&1  && apt-get clean
+
+# Install bazel signing key, repo and package
+ENV bazel_kr_path=/usr/share/keyrings/bazel-release.pub.gpg
+ENV bazel_repo_data="http://storage.googleapis.com/bazel-apt stable jdk1.8"
+
+RUN /usr/bin/curl -s https://bazel.build/bazel-release.pub.gpg \
+      | gpg --dearmor -o "${bazel_kr_path}" \
+    && echo "deb [arch=amd64 signed-by=${bazel_kr_path}] ${bazel_repo_data}" \
+      | dd of=/etc/apt/sources.list.d/bazel.list status=none \
+    && apt-get update -qq
+
+RUN apt-get autoremove -y -qq && \
+    apt-get install -y -qq default-jdk python3-setuptools bazel > /dev/null 2>&1 && \
+    apt-get clean
+
+# Install here any utilities you find useful when troubleshooting
+RUN apt-get -y -qq install emacs-nox vim uuid-runtime > /dev/null 2>&1 && apt-get clean
+
+WORKDIR /init-actions
+
+USER ia-tests
+COPY --chown=ia-tests:ia-tests . ${WORKDIR}
+
+ENTRYPOINT ["/bin/bash"]
+#CMD ["/bin/bash"]

gpu/README.md

@@ -14,12 +14,14 @@ for CUDA, the nvidia kernel driver, cuDNN, and NCCL.
 Specifying a supported value for the `cuda-version` metadata variable
 will select the following values for Driver, CuDNN and NCCL.  At the
 time of writing, the default value for cuda-version, if unspecified is
-12.4.  In addition to 12.4, we have also tested with 11.8.
+12.4.  In addition to 12.4, we have also tested with 11.8, 12.0 and 12.6.
 
-CUDA | Full Version | Driver    | CuDNN     | NCCL    | Supported OSs
+CUDA | Full Version | Driver    | CuDNN     | NCCL    | Tested Dataproc Image Versions
 -----| ------------ | --------- | --------- | ------- | -------------------
-11.8 | 11.8.0       | 525.147   | 8.6.0.163 | 2.15.5  | All
-12.4 | 12.4.1       | 550.90.07 | 9.1.0.70  | 2.21.5  | ALL
+11.8 | 11.8.0       | 560.35.03 | 8.6.0.163 | 2.15.5  | 2.0, 2.1, 2.2-ubuntu22
+12.0 | 12.0.0       | 550.90.07 | 8.8.1.3,  | 2.16.5  | 2.0, 2.1, 2.2-rocky9, 2.2-ubuntu22
+12.4 | 12.4.1       | 550.90.07 | 9.1.0.70  | 2.23.4  | 2.1-ubuntu20, 2.1-rocky8, 2.2
+12.6 | 12.6.2       | 560.35.03 | 9.5.1.17  | 2.23.4  | 2.1-ubuntu20, 2.1-rocky8, 2.2
 
 All variants in the preceeding table have been manually tested to work
 with the installer.  Supported OSs at the time of writing are:
@@ -28,7 +30,6 @@ with the installer.  Supported OSs at the time of writing are:
 * Ubuntu 18.04, 20.04, and 22.04 LTS
 * Rocky 8 and 9
 
-
 ## Using this initialization action
 
 **:warning: NOTICE:** See
@@ -47,16 +48,15 @@ attached GPU adapters.
     CLUSTER_NAME=<cluster_name>
     gcloud dataproc clusters create ${CLUSTER_NAME} \
         --region ${REGION} \
-        --master-accelerator type=nvidia-tesla-v100 \
-        --worker-accelerator type=nvidia-tesla-v100,count=4 \
+        --master-accelerator type=nvidia-tesla-t4 \
+        --worker-accelerator type=nvidia-tesla-t4,count=4 \
         --initialization-actions gs://goog-dataproc-initialization-actions-${REGION}/gpu/install_gpu_driver.sh
     ```
 
 1.  Use the `gcloud` command to create a new cluster with NVIDIA GPU drivers
     and CUDA installed by initialization action as well as the GPU
     monitoring service. The monitoring service is supported on Dataproc 2.0+ Debian
-    and Ubuntu images. Please create a Github issue if support is needed for other
-    Dataproc images.
+    and Ubuntu images.
 
     *Prerequisite:* Create GPU metrics in
     [Cloud Monitoring](https://cloud.google.com/monitoring/docs/) using Google
@@ -90,8 +90,8 @@ attached GPU adapters.
     CLUSTER_NAME=<cluster_name>
     gcloud dataproc clusters create ${CLUSTER_NAME} \
         --region ${REGION} \
-        --master-accelerator type=nvidia-tesla-v100 \
-        --worker-accelerator type=nvidia-tesla-v100,count=4 \
+        --master-accelerator type=nvidia-tesla-t4 \
+        --worker-accelerator type=nvidia-tesla-t4,count=4 \
         --initialization-actions gs://goog-dataproc-initialization-actions-${REGION}/gpu/install_gpu_driver.sh \
         --metadata install-gpu-agent=true \
         --scopes https://www.googleapis.com/auth/monitoring.write
@@ -136,12 +136,12 @@ attached GPU adapters.
 #### GPU Scheduling in YARN:
 
 YARN is the default Resource Manager for Dataproc. To use GPU scheduling feature
-in Spark, it requires YARN version >= 2.10 or >=3.1.1. If intended to use Spark
+in Spark, it requires YARN version >= 2.10 or >= 3.1.1. If intended to use Spark
 with Deep Learning use case, it recommended to use YARN >= 3.1.3 to get support
-for [nvidia-docker version 2](https://github.com/NVIDIA/nvidia-docker).
+for [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit).
 
 In current Dataproc set up, we enable GPU resource isolation by initialization
-script without NVIDIA Docker, you can find more information at
+script with NVIDIA container toolkit.  You can find more information at
 [NVIDIA Spark RAPIDS getting started guide](https://nvidia.github.io/spark-rapids/).
 
 #### cuDNN

gpu/bazel.screenrc

@@ -0,0 +1,11 @@
+screen -L -t 2.0-debian10 1 sh -c '/bin/bash -x gpu/run-bazel-tests.sh 2.0-debian10 ; exec /bin/bash'
+#screen -L -t 2.0-rocky8   2 sh -c '/bin/bash -x gpu/run-bazel-tests.sh 2.0-rocky8   ; exec /bin/bash'
+#screen -L -t 2.0-ubuntu18 3 sh -c '/bin/bash -x gpu/run-bazel-tests.sh 2.0-ubuntu18 ; exec /bin/bash'
+
+#screen -L -t 2.1-debian11 4 sh -c '/bin/bash -x gpu/run-bazel-tests.sh 2.1-debian11 ; exec /bin/bash'
+#screen -L -t 2.1-rocky8   5 sh -c '/bin/bash -x gpu/run-bazel-tests.sh 2.1-rocky8   ; exec /bin/bash'
+#screen -L -t 2.1-ubuntu20 6 sh -c '/bin/bash -x gpu/run-bazel-tests.sh 2.1-ubuntu20 ; exec /bin/bash'
+
+#screen -L -t 2.2-debian12 7 sh -c '/bin/bash -x gpu/run-bazel-tests.sh 2.2-debian12 ; exec /bin/bash'
+#screen -L -t 2.2-rocky9   8 sh -c '/bin/bash -x gpu/run-bazel-tests.sh 2.2-rocky9   ; exec /bin/bash'
+#screen -L -t 2.2-ubuntu22 9 sh -c '/bin/bash -x gpu/run-bazel-tests.sh 2.2-ubuntu22 ; exec /bin/bash'

gpu/env.json.sample

@@ -0,0 +1,7 @@
+{
+  "PROJECT_ID":"example-yyyy-nn",
+  "PURPOSE":"cuda-pre-init",
+  "BUCKET":"my-bucket-name",
+  "IMAGE_VERSION":"2.2-debian12",
+  "ZONE":"us-west4-ñ"
+}