These examples illustrate how to use Morpheus to build a binary sequence classification pipelines to perform root cause analysis on DGX kernel logs.
Like any other Linux based machine, DGX's generate a vast amount of logs. Analysts spend hours trying to identify the root causes of each failure. There could be infinitely many types of root causes of the failures. Some patterns might help to narrow it down; however, regular expressions can only help to identify previously known patterns. Moreover, this creates another manual task of maintaining a search script.
In this example, we demonstrate how using Morpheus can accelerate the analysis of the enormous amount of logs using machine learning. Another benefit of analyzing in a probabilistic way is that we can pin down previously undetected root causes. To achieve this, we will fine-tune a pre-trained BERT1 model with a classification layer using HuggingFace library.
Once the model is capable of identifying even the new root causes, it can also be deployed as a process running in the machines to predict failures before they happen.
The dataset comprises kern.log files from multiple DGX's. Each line inside has been labelled as either 0 for ordinary or 1 for root cause by a script that uses some known patterns. We will be especially interested in lines that are marked as ordinary in the test set but predicted as a root cause as they may be new types of root causes of failures.
The pipeline we will be using in this example is a simple feed-forward linear pipeline where the data from each stage flows on to the next. Morpheus pipelines can be configured via the Morpheus CLI or using the Python library. In this example we will be using the Morpheus CLI.
This example utilizes the Triton Inference Server to perform inference. The binary sequence classification neural network model is provided in the models/root-cause-models directory of the Morpheus repo.
From the Morpheus repo root directory, run the following to launch Triton and load the root-cause-binary-onnx model:
docker run --rm -ti --gpus=all -p8000:8000 -p8001:8001 -p8002:8002 -v $PWD/models:/models nvcr.io/nvidia/tritonserver:23.06-py3 tritonserver --model-repository=/models/triton-model-repo --exit-on-error=false --model-control-mode=explicit --load-model root-cause-binary-onnxWhere 23.06-py3 can be replaced with the current year and month of the Triton version to use. For example, to use May 2021, specify nvcr.io/nvidia/tritonserver:21.05-py3. Ensure that the version of TensorRT that is used in Triton matches the version of TensorRT elsewhere (refer to NGC Deep Learning Frameworks Support Matrix).
This will launch Triton and only load the model required by our example pipeline. The model has been configured with a max batch size of 32, and to use dynamic batching for increased performance.
Once Triton has loaded the model, the following should be displayed:
+----------------------------+-----+--------+
| Model | Version | Status |
+------------------------+---------+--------+
| root-cause-binary-onnx | 1 | READY |
+------------------------+---------+--------+
Note: If this is not present in the output, check the Triton log for any error messages related to loading the model.
To achieve optimized inference performance, Triton Inference Server provides an option to convert our ONNX model to TensorRT. Simply add the following to the end of your config.pbtxt:
dynamic_batching {
preferred_batch_size: [ 1, 4, 8, 16, 32 ]
max_queue_delay_microseconds: 50000
}
optimization { execution_accelerators {
gpu_execution_accelerator : [ {
name : "tensorrt"
parameters { key: "precision_mode" value: "FP16" }
parameters { key: "max_workspace_size_bytes" value: "1073741824" }
}]
}}
This will trigger the ONNX->TensorRT model conversion upon first use of the model. Therefore, expect only the first pipeline run to take several minutes to allow for this conversion. You should notice substantial inference speedups compared to ONNX in subsequent pipeline runs.
More information about TensorRT can be found here.
With the Morpheus CLI, an entire pipeline can be configured and run without writing any code. Using the morpheus run pipeline-nlp command, we can build the pipeline by specifying each stage's name and configuration right on the command line. The output of each stage will become the input for the next.
The following command line is the entire command to build and launch a root cause analysis pipeline which uses binary sequence classification. Each new line represents a new stage. The comment above each stage gives information about why the stage was added and configured this way.
From the Morpheus repo root directory, run:
export MORPHEUS_ROOT=$(pwd)morpheus --log_level=DEBUG \
`# Run a pipeline with 5 threads and a model batch size of 32 (Must match Triton config)` \
run --num_threads=8 --edge_buffer_size=4 --use_cpp=True --pipeline_batch_size=1024 --model_max_batch_size=32 \
`# Specify a NLP pipeline with 128 sequence length (Must match Triton config)` \
pipeline-nlp --model_seq_length=128 --label=not_root_cause --label=is_root_cause \
`# 1st Stage: Read from file` \
from-file --filename=${MORPHEUS_ROOT}/models/datasets/validation-data/root-cause-validation-data-input.jsonlines \
`# 2nd Stage: Deserialize from JSON strings to objects` \
deserialize \
`# 3rd Stage: Preprocessing converts the input data into BERT tokens` \
preprocess --column=log --vocab_hash_file=./data/bert-base-uncased-hash.txt --truncation=True --do_lower_case=True --add_special_tokens=False \
`# 4th Stage: Send messages to Triton for inference. Specify the binary model loaded in Setup` \
inf-triton --force_convert_inputs=True --model_name=root-cause-binary-onnx --server_url=localhost:8001 \
`# 5th Stage: Monitor stage prints throughput information to the console` \
monitor --description='Inference rate' --smoothing=0.001 --unit inf \
`# 6th Stage: Add scores from inference to the messages` \
add-scores --label=is_root_cause \
`# 7th Stage: Convert from objects back into strings` \
serialize --exclude '^ts_' \
`# 8th Stage: Write results out to CSV file` \
to-file --filename=./root-cause-binary-output.jsonlines --overwriteIf successful, the following should be displayed:
Configuring Pipeline via CLI
Loaded labels file. Current labels: [['not_root_cause', 'is_root_cause']]
Parameter, 'vocab_hash_file', with relative path, './data/bert-base-uncased-hash.txt', does not exist. Using package relative location: '/opt/conda/envs/morpheus/lib/python3.8/site-packages/morpheus/./data/bert-base-uncased-hash.txt'
Starting pipeline via CLI... Ctrl+C to Quit
Config:
{
"ae": null,
"class_labels": [
"not_root_cause",
"is_root_cause"
],
"debug": false,
"edge_buffer_size": 4,
"feature_length": 128,
"fil": null,
"log_config_file": null,
"log_level": 10,
"mode": "NLP",
"model_max_batch_size": 32,
"num_threads": 8,
"pipeline_batch_size": 1024,
"plugins": []
}
CPP Enabled: True
====Registering Pipeline====
====Building Pipeline====
====Building Segment: linear_segment_0====
====Building Segment Complete!====
====Building Pipeline Complete!====
Starting! Time: 1668537665.9479523
====Registering Pipeline Complete!====
====Starting Pipeline====
====Pipeline Started====
Added source: <from-file-0; FileSourceStage(filename=/workspace/models/datasets/validation-data/root-cause-validation-data-input.jsonlines, iterative=False, file_type=FileTypes.Auto, repeat=1, filter_null=True)>
└─> morpheus.MessageMeta
Added stage: <deserialize-1; DeserializeStage()>
└─ morpheus.MessageMeta -> morpheus.MultiMessage
Added stage: <preprocess-nlp-2; PreprocessNLPStage(vocab_hash_file=/opt/conda/envs/morpheus/lib/python3.8/site-packages/morpheus/data/bert-base-uncased-hash.txt, truncation=True, do_lower_case=True, add_special_tokens=False, stride=-1, column=log)>
└─ morpheus.MultiMessage -> morpheus.MultiInferenceNLPMessage
Added stage: <inference-3; TritonInferenceStage(model_name=root-cause-binary-onnx, server_url=localhost:8001, force_convert_inputs=True, use_shared_memory=False)>
└─ morpheus.MultiInferenceNLPMessage -> morpheus.MultiResponseMessage
Added stage: <monitor-4; MonitorStage(description=Inference rate, smoothing=0.001, unit=inf, delayed_start=False, determine_count_fn=None)>
└─ morpheus.MultiResponseMessage -> morpheus.MultiResponseMessage
Added stage: <add-scores-5; AddScoresStage(labels=('is_root_cause',), prefix=)>
└─ morpheus.MultiResponseMessage -> morpheus.MultiResponseMessage
Added stage: <serialize-6; SerializeStage(include=(), exclude=('^ts_',), fixed_columns=True)>
└─ morpheus.MultiResponseMessage -> morpheus.MessageMeta
Added stage: <to-file-7; WriteToFileStage(filename=./root-cause-binary-output.jsonlines, overwrite=True, file_type=FileTypes.Auto, include_index_col=True)>
└─ morpheus.MessageMeta -> morpheus.MessageMeta
Inference rate[Complete]: 473 inf [00:01, 340.43 inf/s]
====Pipeline Complete====The output file root-cause-binary-output.jsonlines will contain the original kernel log messages with an additional field is_root_cause. The value of the new field will be the root cause probability.