This quick start guide provides the necessary instructions to set up the minimum infrastructure and configuration needed to deploy the Morpheus Developer Kit and includes example workflows leveraging the deployment.
This quick start guide consists of the following steps:
Set up of the NVIDIA Cloud Native Core Stack
Set up Morpheus AI Engine
Set up Morpheus SDK Client
Models for MLflow Triton Plugin Deployments
Set up Morpheus MLflow Triton Plugin
Deploy models to Triton inference server
Create Kafka topics
Run example workloads
Note: This guide requires access to the NGC Public Catalog.
Morpheus makes it easy to build and scale cybersecurity applications that harness adaptive pipelines supporting a wider range of model complexity than previously feasible. Morpheus makes it possible to analyze up to 100% of your data in real-time, for more accurate detection and faster remediation of threats as they occur. Morpheus also provides the ability to leverage AI to adjust to threats and compensate on the fly, at line rate.
NVIDIA Morpheus enables organizations to attack the issue of cybersecurity head on. Rather than continuously chasing the cybersecurity problem, Morpheus provides the ability to propel you ahead of the breach and address the cybersecurity issue. With the world in a “discover and respond” state, where companies are finding breaches much too late, in a way that is way behind the curve, NVIDIA’s Morpheus cybersecurity AI framework enables any organization to warp to the present and begin to defend itself in real time.
The Morpheus Developer Kit allows developers to quickly and easily set up example pipelines to run inference on different sample models provided by NVIDIA and experiment with the features and capabilities available within the Morpheus framework to address their cybersecurity and information security use cases.
Features
Built on RAPIDS™
Built on the RAPIDS™ libraries, deep learning frameworks, and NVIDIA Triton™ Inference Server, Morpheus simplifies the analysis of logs and telemetry to help detect and mitigate security threats.
Massive Performance and Scale
Enables AI inference and real-time monitoring of every server and packet across the entire network.
Rapid Development and Deployment
Integrates AI frameworks and tools that make it easier for developers to build cybersecurity solutions. Organizations that lack AI expertise can still leverage AI for cybersecurity because Morpheus leverages tools for every stage of the AI workflow, from data preparation to training, inference, and deploying at scale.
Real-time Telemetry
The Morpheus native graph streaming engine can receive rich, real-time network telemetry from every NVIDIA BlueField DPU-accelerated server or NVIDIA AppShield in the data center without impacting performance. Integrating the framework into a third-party cybersecurity offering brings the world’s best AI computing to communication networks.
AI Cybersecurity Capabilities
Deploy your own models using common deep learning frameworks. Or get a jump-start in building applications to identify leaked sensitive information, detect malware or fraud, do network mapping, flag user behavior changes, and identify errors via logs by using one of NVIDIA’s pre-trained and tested models.
Prerequisites
Refer to Appendix A for Cloud (AWS) or On-Prem (Ubuntu)
Registration in the NGC Public Catalog
Continue with the setup steps below once the host system is installed, configured, and satisfies all prerequisites.
Set up NGC API Key and Install NGC Registry CLI
First, you will need to set up your NGC API Key to access all the Morpheus components, using the linked instructions from the NGC Registry CLI User Guide.
Once you’ve created your API key, create an environment variable containing your API key for use by the commands used further in this document:
$ export API_KEY="<NGC_API_KEY>"
Next, install and configure the NGC Registry CLI on your system using the linked instructions from the NGC Registry CLI User Guide.
Create Namespace for Morpheus
Next, create a namespace and an environment variable for the namespace to organize the Kubernetes cluster deployed via the Cloud Native Core Stack and logically separate Morpheus related deployments from other projects using the following command:
$ export NAMESPACE="<YOUR_NAMESPACE>"
$ kubectl create namespace ${NAMESPACE}
Install Morpheus AI Engine
The Morpheus AI Engine consists of the following components:
Triton Inference Server [ ai-engine ] from NVIDIA for processing inference requests.
Kafka Broker [ broker ] to consume and publish messages.
Zookeeper [ zookeeper ] to maintain coordination between the Kafka Brokers.
Follow the below steps to install Morpheus AI Engine:
$ helm fetch https://helm.ngc.nvidia.com/nvidia/morpheus/charts/morpheus-ai-engine-22.09.tgz --username='$oauthtoken' --password=$API_KEY --untar
$ helm install --set ngc.apiKey="$API_KEY" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-ai-engine
After the installation, you can verify that the Kubernetes pods are running successfully using the following command:
$ kubectl -n $NAMESPACE get all
Output:
pod/ai-engine-65f59ddcf7-mdmdt 1/1 Running 0 54s
pod/broker-76f7c64dc9-6rldp 1/1 Running 1 54s
pod/zookeeper-87f9f4dd-znjnb 1/1 Running 0 54s
NAME TYPE CLUSTER-IP EXTERNAL-IP PORT(S) AGE
service/ai-engine ClusterIP 10.109.56.231 <none> 8000/TCP,8001/TCP,8002/TCP 54s
service/broker ClusterIP 10.101.103.250 <none> 9092/TCP 54s
service/zookeeper ClusterIP 10.110.55.141 <none> 2181/TCP 54s
NAME READY UP-TO-DATE AVAILABLE AGE
deployment.apps/ai-engine 1/1 1 1 54s
deployment.apps/broker 1/1 1 1 54s
deployment.apps/zookeeper 1/1 1 1 54s
NAME DESIRED CURRENT READY AGE
replicaset.apps/ai-engine-65f59ddcf7 1 1 1 54s
replicaset.apps/broker-76f7c64dc9 1 1 1 54s
replicaset.apps/zookeeper-87f9f4dd 1 1 1 54s
Install Morpheus SDK Client
Run the following command to pull the Morpheus SDK Client chart on to your instance:
$ helm fetch https://helm.ngc.nvidia.com/nvidia/morpheus/charts/morpheus-sdk-client-22.09.tgz --username='$oauthtoken' --password=$API_KEY --untar
Morpheus SDK Client in Sleep Mode
Install the Morpheus SDK client pod in sleep mode to copy its sample datasets and models from the container to a shared location that other pods can access. If no sdk.args is supplied, the default value /bin/sleep infinity from the chart is used in the following command.
$ helm install --set ngc.apiKey="$API_KEY" \
--namespace $NAMESPACE \
helper \
morpheus-sdk-client
Check the status of the pod to make sure it’s up and running.
$ kubectl -n $NAMESPACE get all | grep sdk-cli-helper
Output:
pod/sdk-cli-helper 1/1 Running 0 41s
Models for MLflow Plugin Deployment
Connect to the sdk-cli-helper container and copy the models to /common, which is mapped to /opt/morpheus/common on the host and where MLflow will have access to model files.
$ kubectl -n $NAMESPACE exec sdk-cli-helper -- cp -RL /workspace/models /common
Install Morpheus MLflow Triton Plugin
The Morpheus MLflow Triton Plugin is used to deploy, update, and remove models from the Morpheus AI Engine. The MLflow server UI can be accessed using NodePort 30500. Follow the below steps to install the Morpheus MLflow Triton Plugin:
$ helm fetch https://helm.ngc.nvidia.com/nvidia/morpheus/charts/morpheus-mlflow-22.09.tgz --username='$oauthtoken' --password=$API_KEY --untar
$ helm install --set ngc.apiKey="$API_KEY" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-mlflow
Note: If the default port is already allocated, helm throws below error. Choose an alternative by adjusting the dashboardPort value in the morpheus-mlflow/values.yaml file, remove the previous release and reinstall it.
Error: Service "mlflow" is invalid: spec.ports[0].nodePort: Invalid value: 30500: provided port is already allocated
After the installation, you can verify that the MLflow pod is running successfully using the following command:
$ kubectl -n $NAMESPACE get all | grep pod/mlflow
Output:
pod/mlflow-6d98 1/1 Running 0 39s
Model Deployment
Attach to the MLfLow pod to publish models to the MLflow server and then deploy it onto Morpheus AI Engine:
$ kubectl -n $NAMESPACE exec -it deploy/mlflow -- bash
(mlflow) root@mlflow-6d98:/mlflow#
Important: When (mlflow) is present, commands are directly within the container.
First let’s have a look at the syntax of the commands we will be using to communicate with the MLflow Triton plugin before we start deploying models. Publish models to MLflow server looks like:
(mlflow) root@mlflow-6d98:/mlflow# python publish_model_to_mlflow.py \
--model_name <REF_MODEL_NAME> \
--model_directory <MODEL_DIR_PATH> \
--flavor <MODEL_FLAVOR>
Deploy models to Morpheus AI Engine:
(mlflow) root@mlflow-6d98:/mlflow# mlflow deployments create -t triton \
--flavor <MODEL_FLAVOR> \
--name <REF_MODEL_NAME> \
-m models:/<REF_MODEL_NAME>/1 \
-C "version=<VERSION_NUMBER>"
Update deployed models in Morpheus AI Engine:
(mlflow) root@mlflow-6d98:/mlflow# mlflow deployments update -t triton \
--flavor <MODEL_FLAVOR> \
--name <REF_MODEL_NAME>/<EXISTING_VERSION_NUMBER> \
-m models:/<REF_MODEL_NAME>/<DESIRED_VERSION_NUMBER>
Delete deployed models from Morpheus AI Engine:
(mlflow) root@mlflow-6d98:/mlflow# mlflow deployments delete -t triton \
--name <REF_MODEL_NAME>/<VERSION_NUMBER>
Now that we’ve figured out how to deploy models let’s move on to the next step. Now it’s time to deploy the relevant models, which have already been copied to /opt/morpheus/common/models which are bound to /common/models within the MLflow pod.
(mlflow) root@mlflow-6d98:/mlflow# ls -lrt /common/models
Output:
drwxr-xr-x 3 ubuntu ubuntu 4096 Apr 13 23:47 sid-minibert-onnx
drwxr-xr-x 2 root root 4096 Apr 21 17:09 abp-models
drwxr-xr-x 4 root root 4096 Apr 21 17:09 datasets
drwxr-xr-x 4 root root 4096 Apr 21 17:09 fraud-detection-models
drwxr-xr-x 2 root root 4096 Apr 21 17:09 dfp-models
drwxr-xr-x 3 root root 4096 Apr 21 17:10 mlflow
drwxr-xr-x 2 root root 4096 Apr 21 17:10 log-parsing-models
drwxr-xr-x 2 root root 4096 Apr 21 17:10 phishing-models
drwxr-xr-x 2 root root 4096 Apr 21 17:10 sid-models
drwxr-xr-x 8 root root 4096 Apr 21 17:10 training-tuning-scripts
drwxr-xr-x 7 root root 4096 Apr 21 17:10 validation-inference-scripts
drwxr-xr-x 7 root root 4096 Apr 21 17:10 triton-model-repo
-rw-r--r-- 1 root root 4213 Apr 21 17:10 README.md
-rw-r--r-- 1 root root 4862 Apr 21 17:10 model_cards.csv
-rw-r--r-- 1 root root 1367 Apr 21 17:10 model-information.csv
Publish and deploy sid-minibert-onnx model:
(mlflow) root@mlflow-6d98:/mlflow# python publish_model_to_mlflow.py \
--model_name sid-minibert-onnx \
--model_directory /common/models/triton-model-repo/sid-minibert-onnx \
--flavor triton
(mlflow) root@mlflow-6d98:/mlflow# mlflow deployments create -t triton \
--flavor triton \
--name sid-minibert-onnx \
-m models:/sid-minibert-onnx/1 \
-C "version=1"
Publish and deploy phishing-bert-onnx model:
(mlflow) root@mlflow-6d98:/mlflow# python publish_model_to_mlflow.py \
--model_name phishing-bert-onnx \
--model_directory /common/models/triton-model-repo/phishing-bert-onnx \
--flavor triton
(mlflow) root@mlflow-6d98:/mlflow# mlflow deployments create -t triton \
--flavor triton \
--name phishing-bert-onnx \
-m models:/phishing-bert-onnx/1 \
-C "version=1"
Publish and deploy abp-nvsmi-xgb model:
(mlflow) root@mlflow-6d98:/mlflow# python publish_model_to_mlflow.py \
--model_name abp-nvsmi-xgb \
--model_directory /common/models/triton-model-repo/abp-nvsmi-xgb \
--flavor triton
(mlflow) root@mlflow-6d98:/mlflow# mlflow deployments create -t triton \
--flavor triton \
--name abp-nvsmi-xgb \
-m models:/abp-nvsmi-xgb/1 \
-C "version=1"
Exit from the container
(mlflow) root@mlflow-6d98:/mlflow# exit
Verify Model Deployment
Run the following command to verify that the models were successfully deployed on the AI Engine:
$ kubectl -n $NAMESPACE logs deploy/ai-engine
Output:
I1202 14:09:03.098085 1 api.cu:79] TRITONBACKEND_ModelInitialize: abp-nvsmi-xgb (version 1)
I1202 14:09:03.101910 1 api.cu:123] TRITONBACKEND_ModelInstanceInitialize: abp-nvsmi-xgb_0 (GPU device 0)
I1202 14:09:03.543719 1 model_instance_state.cu:101] Using GPU for predicting with model 'abp-nvsmi-xgb_0'
I1202 14:09:03.563425 1 api.cu:123] TRITONBACKEND_ModelInstanceInitialize: abp-nvsmi-xgb_0 (GPU device 1)
I1202 14:09:03.980621 1 model_instance_state.cu:101] Using GPU for predicting with model 'abp-nvsmi-xgb_0'
I1202 14:09:03.981678 1 model_repository_manager.cc:1183] successfully loaded 'abp-nvsmi-xgb' version 1
Create Kafka Topics
We will need to create Kafka topics for input and output data to run some of the pipeline examples.
Check if any Kafka topics exist already. If any exist, you can either delete the previous topics or re-use them.
$ kubectl -n $NAMESPACE exec deploy/broker -c broker -- kafka-topics.sh --list --zookeeper zookeeper:2181
Run the following command twice, once to create an input topic, and again to create an output topic, making sure that the input topic and output topic have different names:
$ kubectl -n $NAMESPACE exec deploy/broker -c broker -- kafka-topics.sh \
--create \
--bootstrap-server broker:9092 \
--replication-factor 1 \
--partitions 3 \
--topic <YOUR_KAFKA_TOPIC>
This section describes example workflows to run on Morpheus. Four sample pipelines are provided.
AutoEncoder pipeline performing Digital Fingerprinting (DFP).
NLP pipeline performing Phishing Detection (PD).
NLP pipeline performing Sensitive Information Detection (SID).
FIL pipeline performing Anomalous Behavior Profiling (ABP).
Multiple command options are given for each pipeline, with varying data input/output methods, ranging from local files to Kafka Topics.
We recommend only deploying one pipeline at a time. To remove previously deployed pipelines, run the following command:
$ helm delete -n $NAMESPACE <YOUR_RELEASE_NAME>
To publish messages to a Kafka topic, we need to copy datasets to locations where they can be accessed from the host.
kubectl -n $NAMESPACE exec sdk-cli-helper -- cp -R /workspace/examples/data /common
Refer to the Using Morpheus SDK Client to Run Pipelines section of the Appendix for more information regarding the commands.
Note: Before running the example pipelines, ensure that the criteria below are met:
Ensure that models specific to the pipeline are deployed.
Input and Output Kafka topics have been created.
Recommended to create an output directory under
/opt/morpheus/common/datawhich is bound to/common/data(pod/container) for storing inference or validation results.Replace <YOUR_OUTPUT_DIR> with your directory name.
Replace <YOUR_INPUT_KAFKA_TOPIC> with your input Kafka topic name.
Replace <YOUR_OUTPUT_KAFKA_TOPIC> with your output Kafka topic name.
Replace <YOUR_RELEASE_NAME> with the name you want.
For reference, the Morpheus SDK Client install pipeline command template is provided. Let’s take a closer look at this when running example workflows, but for now, let’s proceed to the next step.
$ helm install --set ngc.apiKey="$API_KEY" \
--set sdk.args="<REPLACE_RUN_PIPELINE_COMMAND_HERE>" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-sdk-client
Run AutoEncoder Digital Fingerprinting Pipeline
The following AutoEncoder pipeline example shows how to train and validate the AutoEncoder model and write the inference results to a specified location. Digital fingerprinting has also been referred to as HAMMAH (Human as Machine <> Machine as Human). These use cases are currently implemented to detect user behavior changes that indicate a change from a human to a machine or a machine to a human, thus leaving a “digital fingerprint”. The model is an ensemble of an autoencoder and fast fourier transform reconstruction.
Inference and training based on a userid (user123). The model is trained once and inference is conducted on the supplied input entries in the example pipeline below. The --train_data_glob parameter must be removed for continuous training.
$ helm install --set ngc.apiKey="$API_KEY" \
--set sdk.args="morpheus --log_level=DEBUG run \
--num_threads=2 \
--edge_buffer_size=4 \
--pipeline_batch_size=1024 \
--model_max_batch_size=1024 \
--use_cpp=False \
pipeline-ae \
--columns_file=data/columns_ae_cloudtrail.txt \
--userid_filter=user123 \
--feature_scaler=standard \
--userid_column_name=userIdentitysessionContextsessionIssueruserName \
from-cloudtrail --input_glob=/common/models/datasets/validation-data/dfp-cloudtrail-*-input.csv \
--max_files=200 \
train-ae --train_data_glob=/common/models/datasets/training-data/dfp-cloudtrail-*.csv \
--source_stage_class=morpheus.stages.input.cloud_trail_source_stage.CloudTrailSourceStage \
--seed 42 \
preprocess \
inf-pytorch \
add-scores \
timeseries --resolution=1m --zscore_threshold=8.0 --hot_start \
monitor --description 'Inference Rate' --smoothing=0.001 --unit inf \
serialize \
to-file --filename=/common/data/<YOUR_OUTPUT_DIR>/cloudtrail-dfp-detections.csv --overwrite" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-sdk-client
For more information on the Digital Fingerprint use cases, please refer to the starter example and a more production-ready example that can be found in the examples source directory.
Run NLP Phishing Detection Pipeline
The following Phishing Detection pipeline examples use a pre-trained NLP model to analyze emails (body) and determine phishing or benign. Here is the sample data as shown below is used to pass as an input to the pipeline.
{"data":"Abedin Huma <AbedinH@state.gov>Wednesday July 15 2009 1:44 PMRe: ArtWill be off campus at meetingBut you should definitely come I think they have found some good things."}
{"data":"See NIMills Cheryl D <MillsCD@state.gov>Saturday December 112010 1:36 PMFw: S is calling Leahy today - thx for all the help; advise if a diff no for him today"}
{"data":"Here is Draft"}
{"data":"Ok"}
Pipeline example to read data from a file, run inference using a phishing-bert-onnx model, and write inference results to the specified output file:
$ helm install --set ngc.apiKey="$API_KEY" \
--set sdk.args="morpheus --log_level=DEBUG run \
--num_threads=2 \
--edge_buffer_size=4 \
--pipeline_batch_size=1024 \
--model_max_batch_size=32 \
--use_cpp=True \
pipeline-nlp \
--model_seq_length=128 \
--labels_file=data/labels_phishing.txt \
from-file --filename=./examples/data/email.jsonlines \
monitor --description 'FromFile Rate' --smoothing=0.001 \
deserialize \
preprocess --vocab_hash_file=data/bert-base-uncased-hash.txt --truncation=True --do_lower_case=True --add_special_tokens=False \
monitor --description 'Preprocess Rate' \
inf-triton --model_name=phishing-bert-onnx --server_url=ai-engine:8000 --force_convert_inputs=True \
monitor --description 'Inference Rate' --smoothing=0.001 --unit inf \
add-class --label=pred --threshold=0.7 \
serialize \
to-file --filename=/common/data/<YOUR_OUTPUT_DIR>/phishing-bert-onnx-output.jsonlines --overwrite" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-sdk-client
When the pipeline runs successfully, an output file phishing-bert-onnx-output.jsonlines will appear in the output directory.
Pipeline example to read messages from an input Kafka topic, run inference using a phishing-bert-onnx model, and write the results of the inference to an output Kafka topic:
$ helm install --set ngc.apiKey="$API_KEY" \
--set sdk.args="morpheus --log_level=DEBUG run \
--num_threads=2 \
--edge_buffer_size=4 \
--pipeline_batch_size=1024 \
--model_max_batch_size=32 \
--use_cpp=True \
pipeline-nlp \
--model_seq_length=128 \
--labels_file=data/labels_phishing.txt \
from-kafka --input_topic <YOUR_INPUT_KAFKA_TOPIC> --bootstrap_servers broker:9092 \
monitor --description 'FromKafka Rate' --smoothing=0.001 \
deserialize \
preprocess --vocab_hash_file=data/bert-base-uncased-hash.txt --truncation=True --do_lower_case=True --add_special_tokens=False \
monitor --description 'Preprocess Rate' \
inf-triton --force_convert_inputs=True --model_name=phishing-bert-onnx --server_url=ai-engine:8000 \
monitor --description='Inference Rate' --smoothing=0.001 --unit inf \
add-class --label=pred --threshold=0.7 \
serialize --exclude '^ts_' \
to-kafka --output_topic <YOUR_OUTPUT_KAFKA_TOPIC> --bootstrap_servers broker:9092" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-sdk-client
Make sure you create input and output Kafka topics before you start the pipeline. After the pipeline has been started, load the individual corresponding data files from the downloaded sample into the selected input topic using the command below:
$ kubectl -n $NAMESPACE exec -it deploy/broker -c broker -- kafka-console-producer.sh \
--broker-list broker:9092 \
--topic <YOUR_INPUT_KAFKA_TOPIC> < \
<YOUR_INPUT_DATA_FILE_PATH_EXAMPLE: /opt/morpheus/common/data/email.jsonlines>
Note: This should be used for development purposes only via this developer kit. Loading from the file into Kafka should not be used in production deployments of Morpheus.
Run NLP Sensitive Information Detection Pipeline
The following Sensitive Information Detection pipeline examples use a pre-trained NLP model to ingest and analyze PCAP (packet capture network traffic) input sample data, like the example below, to inspect IP traffic across data center networks.
{"timestamp": 1616380971990, "host_ip": "10.188.40.56", "data_len": "309", "data": "POST /simpledatagen/ HTTP/1.1\r\nHost: echo.gtc1.netqdev.cumulusnetworks.com\r\nUser-Agent: python-requests/2.22.0\r\nAccept-Encoding: gzip, deflate\r\nAccept: */*\r\nConnection: keep-alive\r\nContent-Length: 73\r\nContent-Type: application/json\r\n\r\n", "src_mac": "04:3f:72:bf:af:74", "dest_mac": "b4:a9:fc:3c:46:f8", "protocol": "6", "src_ip": "10.20.16.248", "dest_ip": "10.244.0.59", "src_port": "50410", "dest_port": "80", "flags": "24", "is_pii": false}
{"timestamp": 1616380971991, "host_ip": "10.188.40.56", "data_len": "139", "data": "\"{\\\"markerEmail\\\": \\\"FuRLFaAZ identify benefit BneiMvCZ join 92694759\\\"}\"", "src_mac": "04:3f:72:bf:af:74", "dest_mac": "b4:a9:fc:3c:46:f8", "protocol": "6", "src_ip": "10.244.0.1", "dest_ip": "10.244.0.25", "src_port": "50410", "dest_port": "80", "flags": "24", "is_pii": false}
Pipeline example to read data from a file, run inference using a sid-minibert-onnx model, and write inference results to the specified output file:
$ helm install --set ngc.apiKey="$API_KEY" \
--set sdk.args="morpheus --log_level=DEBUG run \
--num_threads=3 \
--edge_buffer_size=4 \
--use_cpp=True \
--pipeline_batch_size=1024 \
--model_max_batch_size=32 \
pipeline-nlp \
--model_seq_length=256 \
from-file --filename=./examples/data/pcap_dump.jsonlines \
monitor --description 'FromFile Rate' --smoothing=0.001 \
deserialize \
preprocess --vocab_hash_file=data/bert-base-uncased-hash.txt --truncation=True --do_lower_case=True --add_special_tokens=False \
monitor --description='Preprocessing rate' \
inf-triton --force_convert_inputs=True --model_name=sid-minibert-onnx --server_url=ai-engine:8000 \
monitor --description='Inference rate' --smoothing=0.001 --unit inf \
add-class \
serialize --exclude '^ts_' \
to-file --filename=/common/data/<YOUR_OUTPUT_DIR>/sid-minibert-onnx-output.jsonlines --overwrite" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-sdk-client
When the pipeline runs successfully, an output file sid-minibert-onnx-output.jsonlines will appear in the output directory.
Pipeline example to read messages from an input Kafka topic, run inference using a sid-minibert-onnx model, and write the results of the inference to an output Kafka topic:
$ helm install --set ngc.apiKey="$API_KEY" \
--set sdk.args="morpheus --log_level=DEBUG run \
--num_threads=3 \
--edge_buffer_size=4 \
--use_cpp=True \
--pipeline_batch_size=1024 \
--model_max_batch_size=32 \
pipeline-nlp \
--model_seq_length=256 \
from-kafka --input_topic <YOUR_INPUT_KAFKA_TOPIC> --bootstrap_servers broker:9092 \
monitor --description 'FromKafka Rate' --smoothing=0.001 \
deserialize \
preprocess --vocab_hash_file=data/bert-base-uncased-hash.txt --truncation=True --do_lower_case=True --add_special_tokens=False \
monitor --description='Preprocessing Rate' \
inf-triton --force_convert_inputs=True --model_name=sid-minibert-onnx --server_url=ai-engine:8000 \
monitor --description='Inference Rate' --smoothing=0.001 --unit inf \
add-class \
serialize --exclude '^ts_' \
to-kafka --output_topic <YOUR_OUTPUT_KAFKA_TOPIC> --bootstrap_servers broker:9092" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-sdk-client
Make sure you create input and output Kafka topics before you start the pipeline. After the pipeline has been started, load the individual corresponding data files from the downloaded sample into the selected input topic using the command below:
$ kubectl -n $NAMESPACE exec -it deploy/broker -c broker -- kafka-console-producer.sh \
--broker-list broker:9092 \
--topic <YOUR_INPUT_KAFKA_TOPIC> < \
<YOUR_INPUT_DATA_FILE_PATH_EXAMPLE: ${HOME}/examples/data/pcap_dump.jsonlines>
Note: This should be used for development purposes only via this developer kit. Loading from the file into Kafka should not be used in production deployments of Morpheus.
Run FIL Anomalous Behavior Profiling Pipeline
The following Anomalous Behavior Profiling pipeline examples use a pre-trained FIL model to ingest and analyze Nvidia System Management Interface (nvidia-smi) logs, like the example below, as input sample data to identify crypto mining activity on GPU devices.
{"nvidia_smi_log.gpu.pci.tx_util": "0 KB/s", "nvidia_smi_log.gpu.pci.rx_util": "0 KB/s", "nvidia_smi_log.gpu.fb_memory_usage.used": "3980 MiB", "nvidia_smi_log.gpu.fb_memory_usage.free": "12180 MiB", "nvidia_smi_log.gpu.bar1_memory_usage.total": "16384 MiB", "nvidia_smi_log.gpu.bar1_memory_usage.used": "11 MiB", "nvidia_smi_log.gpu.bar1_memory_usage.free": "16373 MiB", "nvidia_smi_log.gpu.utilization.gpu_util": "0 %", "nvidia_smi_log.gpu.utilization.memory_util": "0 %", "nvidia_smi_log.gpu.temperature.gpu_temp": "61 C", "nvidia_smi_log.gpu.temperature.gpu_temp_max_threshold": "90 C", "nvidia_smi_log.gpu.temperature.gpu_temp_slow_threshold": "87 C", "nvidia_smi_log.gpu.temperature.gpu_temp_max_gpu_threshold": "83 C", "nvidia_smi_log.gpu.temperature.memory_temp": "57 C", "nvidia_smi_log.gpu.temperature.gpu_temp_max_mem_threshold": "85 C", "nvidia_smi_log.gpu.power_readings.power_draw": "61.77 W", "nvidia_smi_log.gpu.clocks.graphics_clock": "1530 MHz", "nvidia_smi_log.gpu.clocks.sm_clock": "1530 MHz", "nvidia_smi_log.gpu.clocks.mem_clock": "877 MHz", "nvidia_smi_log.gpu.clocks.video_clock": "1372 MHz", "nvidia_smi_log.gpu.applications_clocks.graphics_clock": "1312 MHz", "nvidia_smi_log.gpu.applications_clocks.mem_clock": "877 MHz", "nvidia_smi_log.gpu.default_applications_clocks.graphics_clock": "1312 MHz", "nvidia_smi_log.gpu.default_applications_clocks.mem_clock": "877 MHz", "nvidia_smi_log.gpu.max_clocks.graphics_clock": "1530 MHz", "nvidia_smi_log.gpu.max_clocks.sm_clock": "1530 MHz", "nvidia_smi_log.gpu.max_clocks.mem_clock": "877 MHz", "nvidia_smi_log.gpu.max_clocks.video_clock": "1372 MHz", "nvidia_smi_log.gpu.max_customer_boost_clocks.graphics_clock": "1530 MHz", "nvidia_smi_log.gpu.processes.process_info.0.process_name": "python", "nvidia_smi_log.gpu.processes.process_info.1.process_name": "tritonserver", "hostname": "ip-10-100-8-98", "timestamp": 1615542360.9566503}
Pipeline example to read data from a file, run inference using an abp-nvsmi-xgb model, and write inference results to the specified output file.
$ helm install --set ngc.apiKey="$API_KEY" \
--set sdk.args="morpheus --log_level=DEBUG run \
--num_threads=3 \
--edge_buffer_size=4 \
--pipeline_batch_size=1024 \
--model_max_batch_size=64 \
--use_cpp=True \
pipeline-fil \
from-file --filename=./examples/data/nvsmi.jsonlines \
monitor --description 'FromFile Rate' --smoothing=0.001 \
deserialize \
preprocess \
monitor --description='Preprocessing Rate' \
inf-triton --model_name=abp-nvsmi-xgb --server_url=ai-engine:8000 --force_convert_inputs=True \
monitor --description='Inference Rate' --smoothing=0.001 --unit inf \
add-class \
serialize --exclude '^nvidia_smi_log' --exclude '^ts_' \
to-file --filename=/common/data/<YOUR_OUTPUT_DIR>/abp-nvsmi-xgb-output.jsonlines --overwrite" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-sdk-client
Pipeline example to read messages from an input Kafka topic, run inference using an abp-nvsmi-xgb model, and write the results of the inference to an output Kafka topic:
$ helm install --set ngc.apiKey="$API_KEY" \
--set sdk.args="morpheus --log_level=DEBUG run \
--num_threads=3 \
--pipeline_batch_size=1024 \
--model_max_batch_size=64 \
--use_cpp=True \
pipeline-fil \
from-kafka --input_topic <YOUR_INPUT_KAFKA_TOPIC> --bootstrap_servers broker:9092 \
monitor --description 'FromKafka Rate' --smoothing=0.001 \
deserialize \
preprocess \
monitor --description='Preprocessing Rate' \
inf-triton --model_name=abp-nvsmi-xgb --server_url=ai-engine:8000 --force_convert_inputs=True \
monitor --description='Inference Rate' --smoothing=0.001 --unit inf \
add-class \
serialize --exclude '^nvidia_smi_log' \ --exclude '^ts_' \
to-kafka --output_topic <YOUR_OUTPUT_KAFKA_TOPIC> --bootstrap_servers broker:9092" \
--namespace $NAMESPACE \
<YOUR_RELEASE_NAME> \
morpheus-sdk-client
Make sure you create input and output Kafka topics before you start the pipeline. After the pipeline has been started, load the individual corresponding data files from the downloaded sample into the selected input topic using the command below:
$ kubectl -n $NAMESPACE exec -it deploy/broker -c broker -- kafka-console-producer.sh \
--broker-list broker:9092 \
--topic <YOUR_INPUT_KAFKA_TOPIC> < \
<YOUR_INPUT_DATA_FILE_PATH_EXAMPLE: ${HOME}/examples/data/nvsmi.jsonlines>
Note: This should be used for development purposes only via this developer kit. Loading from the file into Kafka should not be used in production deployments of Morpheus.
Verify Running Pipeline
Once you’ve deployed the SDK client to run a pipeline, you can check the status of the pod using the following command:
$ kubectl -n $NAMESPACE get pods sdk-cli-<RELEASE_NAME>
NAME READY STATUS RESTARTS AGE
sdk-cli-6c9575f648-gfdd2 1/1 Running 0 3m23s
Then check that the pipeline is running successfully using the following command:
kubectl -n $NAMESPACE logs sdk-cli-<RELEASE_NAME>
Output:
Configuring Pipeline via CLI
Starting pipeline via CLI... Ctrl+C to Quit
Preprocessing rate: 7051messages [00:09, 4372.75messages/s]
Inference rate: 7051messages [00:04, 4639.40messages/s]
Prerequisites and Installation for AWS
Prerequisites
AWS account with the ability to create/modify EC2 instances
AWS EC2 G4 instance with T4 or V100 GPU, at least 64GB RAM, 8 cores CPU, and 100 GB storage.
Install Cloud Native Core Stack for AWS
On your AWS EC2 G4 instance, follow the instructions in the linked document to install NVIDIA’s Cloud Native Core Stack for AWS.
Prerequisites and Installation for Ubuntu
Prerequisites
NVIDIA-Certified System
NVIDIA Pascal GPU or newer (Compute Capability >= 6.0)
Ubuntu 20.04 LTS or newer
Installing Cloud Native Core Stack on NVIDIA Certified Systems
On your NVIDIA-Certified System, follow the instructions in the linked document to install NVIDIA’s Cloud Native Core Stack.
Kafka Topic Commands
List available Kafka topics.
$ kubectl -n $NAMESPACE exec deploy/broker -c broker -- kafka-topics.sh \
--list --zookeeper zookeeper:2181
Create a partitioned Kafka topic with a single replication factor.
$ kubectl -n $NAMESPACE exec deploy/broker -c broker -- kafka-topics.sh \
--create \
--bootstrap-server broker:9092 \
--replication-factor 1 \
--partitions 1 \
--topic <YOUR_KAFKA_TOPIC>
Load data from a file to Kafka topic:
$ kubectl -n $NAMESPACE exec -it deploy/broker -c broker -- kafka-console-producer.sh \
--broker-list broker:9092 \
--topic <YOUR_KAFKA_TOPIC> < \
<YOUR_INPUT_DATA_FILE>
Note: This should be used for development purposes only via this developer kit. Loading from the file into Kafka should not be used in production deployments of Morpheus.
Consume messages from Kafka topic:
$ kubectl -n $NAMESPACE exec deploy/broker -c broker -- kafka-console-consumer.sh \
--bootstrap-server broker:9092 \
--topic <YOUR_KAFKA_TOPIC> \
--group <YOUR_CONSUMER_GROUP_ID>
Delete Kafka topic:
$ kubectl -n $NAMESPACE exec deploy/broker -c broker -- kafka-topics.sh \
--delete --zookeeper zookeeper:2181 \
--topic <YOUR_KAFKA_TOPIC>
Using Morpheus SDK Client to Run Pipelines
The Morpheus SDK client allows you to configure several supported pipelines and provides flexibility to execute the pipeline in multithread mode.
(morpheus) root@sdk-cli:/workspace# morpheus run --help
Usage: morpheus run [OPTIONS] COMMAND [ARGS]...
Options:
--num_threads INTEGER RANGE Number of internal pipeline threads to use [default: 8; x>=1]
--pipeline_batch_size INTEGER RANGE
Internal batch size for the pipeline.
Can be much larger than the model batch size.
Also used for Kafka consumers [default: 256; x>=1]
--model_max_batch_size INTEGER RANGE
Max batch size to use for the model [default: 8; x>=1]
--edge_buffer_size INTEGER RANGE
The size of buffered channels to use between nodes in a pipeline.
Larger values reduce backpressure at the cost of memory.
Smaller values will push messages through the pipeline quicker.
Must be greater than 1 and a power of 2 (i.e. 2, 4, 8, 16, etc.) [default: 128; x>=2]
--use_cpp BOOLEAN Whether or not to use C++ node and message types or to prefer python.
Only use as a last resort if bugs are encountered [default: True]
--help Show this message and exit.
Commands:
pipeline-ae Run the inference pipeline with an AutoEncoder model
pipeline-fil Run the inference pipeline with a FIL model
pipeline-nlp Run the inference pipeline with a NLP model
pipeline-other Run a custom inference pipeline without a specific model type
Four different pipelines are currently supported: a pipeline running an NLP model, a pipeline running a FIL model, a pipeline running an AutoEncoder model, and a generic pipeline.
For details of running pipeline-other, please refer to the GNN Fraud Detection use case in the examples source directory.
The Morpheus SDK Client provides the commands below to run the NLP pipeline:
(morpheus) root@sdk-cli:/workspace# morpheus run pipeline-nlp --help
Usage: morpheus run pipeline-nlp [OPTIONS] COMMAND1 [ARGS]... [COMMAND2 [ARGS]...]...
Configure and run the pipeline. To configure the pipeline, list the stages in the order that data should flow. The output of each stage will become the input for the
next stage. For example, to read, classify and write to a file, the following stages could be used
pipeline from-file --filename=my_dataset.json deserialize preprocess inf-triton --model_name=my_model
--server_url=localhost:8001 filter --threshold=0.5 to-file --filename=classifications.json
Pipelines must follow a few rules:
1. Data must originate in a source stage. Current options are `from-file` or `from-kafka`
2. A `deserialize` stage must be placed between the source stages and the rest of the pipeline
3. Only one inference stage can be used. Zero is also fine
4. The following stages must come after an inference stage: `add-class`, `filter`, `gen-viz`
Options:
--model_seq_length INTEGER RANGE
Limits the length of the sequence returned. If tokenized string is shorter than max_length, output will be padded with 0s.
If the tokenized string is longer than max_length and do_truncate == False,
there will be multiple returned sequences containing the overflowing
token-ids. Default value is 256 [default: 256; x>=1]
--label TEXT Specify output labels.
--labels_file DATA FILE Specifies a file to read labels from in order to convert class IDs into labels.
A label file is a simple text file where each line corresponds to a label.
Ignored when --label is specified [default: data/labels_nlp.txt]
--viz_file FILE Save a visualization of the pipeline at the specified location
--help Show this message and exit.
Commands:
add-class Add detected classifications to each message.
add-scores Add probability scores to each message.
buffer (Deprecated) Buffer results.
delay (Deprecated) Delay results for a certain duration.
deserialize Deserialize source data into Dataframes.
dropna Drop null data entries from a DataFrame.
filter Filter message by a classification threshold.
from-file Load messages from a file.
from-kafka Load messages from a Kafka cluster.
gen-viz (Deprecated) Write out vizualization DataFrames.
inf-identity Perform inference for testing that performs a no-op.
inf-pytorch Perform inference with PyTorch.
inf-triton Perform inference with Triton Inference Server.
mlflow-drift Report model drift statistics to ML Flow.
monitor Display throughput numbers at a specific point in the pipeline.
preprocess Prepare NLP input DataFrames for inference.
serialize Include & exclude columns from messages.
to-file Write all messages to a file.
to-kafka Write all messages to a Kafka cluster.
trigger Buffer data until previous stage has completed.
validate Validate pipeline output for testing.
Morpheus SDK Client provides the commands below to run the FIL pipeline:
(morpheus) root@sdk-cli:/workspace# morpheus run pipeline-fil --help
Usage: morpheus run pipeline-fil [OPTIONS] COMMAND1 [ARGS]... [COMMAND2 [ARGS]...]...
Configure and run the pipeline. To configure the pipeline, list the stages in the order that data should flow. The output of each stage will become the input for the
next stage. For example, to read, classify and write to a file, the following stages could be used
pipeline from-file --filename=my_dataset.json deserialize preprocess inf-triton --model_name=my_model
--server_url=localhost:8000 filter --threshold=0.5 to-file --filename=classifications.json
Pipelines must follow a few rules:
1. Data must originate in a source stage. Current options are `from-file` or `from-kafka`
2. A `deserialize` stage must be placed between the source stages and the rest of the pipeline
3. Only one inference stage can be used. Zero is also fine
4. The following stages must come after an inference stage: `add-class`, `filter`, `gen-viz`
Options:
--model_fea_length INTEGER RANGE
Number of features trained in the model [default: 29; x>=1]
--label TEXT Specify output labels. Ignored when --labels_file is specified [default: mining]
--labels_file DATA FILE Specifies a file to read labels from in order to convert class IDs into labels. A label file is a simple text file where each line
corresponds to a label. If unspecified the value specified by the --label flag will be used.
--columns_file DATA FILE Specifies a file to read column features. [default: data/columns_fil.txt]
--viz_file FILE Save a visualization of the pipeline at the specified location
--help Show this message and exit.
Commands:
add-class Add detected classifications to each message.
add-scores Add probability scores to each message.
buffer (Deprecated) Buffer results.
delay (Deprecated) Delay results for a certain duration.
deserialize Deserialize source data into Dataframes.
dropna Drop null data entries from a DataFrame.
filter Filter message by a classification threshold.
from-appshield Source stage is used to load Appshield messages from one or more plugins into a dataframe. It normalizes nested json messages and arranges them into a
dataframe by snapshot and source(Determine which source generated the plugin messages).
from-file Load messages from a file.
from-kafka Load messages from a Kafka cluster.
inf-identity Perform inference for testing that performs a no-op.
inf-pytorch Perform inference with PyTorch.
inf-triton Perform inference with Triton Inference Server.
mlflow-drift Report model drift statistics to ML Flow.
monitor Display throughput numbers at a specific point in the pipeline.
preprocess Prepare FIL input DataFrames for inference.
serialize Include & exclude columns from messages.
to-file Write all messages to a file.
to-kafka Write all messages to a Kafka cluster.
trigger Buffer data until previous stage has completed.
validate Validate pipeline output for testing.
Morpheus SDK Client provides the commands below to run the AutoEncoder pipeline:
(morpheus) root@sdk-cli:/workspace# morpheus run pipeline-ae --help
Usage: morpheus run pipeline-ae [OPTIONS] COMMAND1 [ARGS]... [COMMAND2 [ARGS]...]...
Configure and run the pipeline. To configure the pipeline, list the stages in the order that data should flow. The output of each stage will become the input for the
next stage. For example, to read, classify and write to a file, the following stages could be used
pipeline from-file --filename=my_dataset.json deserialize preprocess inf-triton --model_name=my_model
--server_url=localhost:8000 filter --threshold=0.5 to-file --filename=classifications.json
Pipelines must follow a few rules:
1. Data must originate in a source stage. Current options are `from-file` or `from-kafka`
2. A `deserialize` stage must be placed between the source stages and the rest of the pipeline
3. Only one inference stage can be used. Zero is also fine
4. The following stages must come after an inference stage: `add-class`, `filter`, `gen-viz`
Options:
--columns_file DATA FILE [required]
--labels_file DATA FILE Specifies a file to read labels from in order to convert class IDs into labels. A label file is a simple text file where each line
corresponds to a label. If unspecified, only a single output label is created for FIL
--userid_column_name TEXT Which column to use as the User ID. [default: userIdentityaccountId; required]
--userid_filter TEXT Specifying this value will filter all incoming data to only use rows with matching User IDs. Which column is used for the User ID is
specified by `userid_column_name`
--feature_scaler [none|standard|gauss_rank]
Autoencoder feature scaler [default: standard]
--use_generic_model Whether to use a generic model when user does not have minimum number of training rows
--viz_file FILE Save a visualization of the pipeline at the specified location
--help Show this message and exit.
Commands:
add-class Add detected classifications to each message.
add-scores Add probability scores to each message.
buffer (Deprecated) Buffer results.
delay (Deprecated) Delay results for a certain duration.
filter Filter message by a classification threshold.
from-azure Source stage is used to load AWS CloudTrail messages from a file and dumping the contents into the pipeline immediately. Useful for testing performance
and accuracy of a pipeline.
from-cloudtrail Load messages from a Cloudtrail directory.
from-duo Source stage is used to load AWS CloudTrail messages from a file and dumping the contents into the pipeline immediately. Useful for testing performance
and accuracy of a pipeline.
inf-pytorch Perform inference with PyTorch.
inf-triton Perform inference with Triton Inference Server.
monitor Display throughput numbers at a specific point in the pipeline.
preprocess Prepare Autoencoder input DataFrames for inference.
serialize Include & exclude columns from messages.
timeseries Perform time series anomaly detection and add prediction.
to-file Write all messages to a file.
to-kafka Write all messages to a Kafka cluster.
train-ae Train an Autoencoder model on incoming data.
trigger Buffer data until previous stage has completed.
validate Validate pipeline output for testing.
Additional Documentation
For more information on how to use the Morpheus CLI to customize and run your own optimized AI pipelines, Refer to below documentation.
Troubleshooting
This section lists solutions to problems you might encounter with Morpheus or from it’s supporting components.
Common Problems
Models Unloaded After Reboot
When the pod is restarted, K8s will not automatically load the models. Since models are deployed to ai-engine in explicit mode using MLflow, we’d have to manually deploy them again using the Model Deployment process.
AI Engine CPU Only Mode
After a server restart, the ai-engine pod on k8s can start up before the gpu operator infrastructure is available, making it “think” there is no driver installed (i.e., CPU -only mode).
Improve Pipeline Message Processing Rate
Below settings need to be considered
Provide the workflow with the optimal number of threads (
—num threads), as having more or fewer threads can have an impact on pipeline performance.Consider adjusting
pipeline_batch_sizeandmodel_max_batch_size
Kafka Message Offset Commit Fail
Error Message
1649207839.253|COMMITFAIL|rdkafka#consumer-2| [thrd:main]: Offset commit (manual) failed for 1/1 partition(s) in join-state wait-unassign-call: Broker: Unknown member: topic[0]@112071(Broker: Unknown member)
Problem: If the standalone kafka cluster is receiving significant message throughput from the producer, this error may happen.
Solution: Reinstall the Morpheus workflow and reduce the Kafka topic’s message retention time and message producing rate.
The dropna stage
The Drop Null Attributes stage (dropna) requires the specification of a column name. This column will vary from use case (and its input data) to use case. These are the applicable columns for the pre-built pipelines provided by Morpheus.
Input |
Columns |
|---|---|
Azure DFP |
userPrincipalName |
Duo DFP |
username |
DFP Cloudtrail |
userIdentitysessionContextsessionIssueruserName |
data |
|
GNN |
index, client_node, merchant_node |
Log Parsing |
raw |
PCAP |
data |
Ransomware |
PID, Process, snapshot_id, timestamp, source |
Issue |
Description |
|---|---|
nv-morpheus/SRF#157 |
Requesting more threads than CPU can lead to an abort in SRF, so reduce the number of pipeline threads to be equal to or less than available CPU. This applies to all environments including bare metal and cloud (vCPU). |