wip

Author: biswapanda

benchmarking/benchmarking.md

@@ -0,0 +1,100 @@
+# Benchmarking Harness
+
+In current state, benchmarking has a few problems:
+
+1. UX: Experimentation, debugging, and iteration is hard. 
+use case: As a user, I want to easily experiment with different configs, and get results quickly and compare them.
+
+2. Reproducibility is hard: we don't store the input configs and results.
+use case: As a user, I want to be able to reproduce my experiments and share them with others.
+
+3. benchmarking steps are tightly coupled. If a sinlge step/benchmark config fails the entire process is aborted/retried.   
+
+4. port-forwarding and benchmarking has non-deterministic latency characteristics. 
+
+## Proposed plan:
+
+1. decouple all steps and then compose them together: prep a model, deploy k8s cr, benchmark, collect data 
+
+2. capture configs for the experiment: deploy (config or a reference to deployment), benchmark, model etc
+
+3. we'd run benchmarks inside k8s cluster in k8s native approach.
+
+
+## Steps:
+Following steps are executed by the harness:
+
+Note: These steps are reusable across different tests (LLM benchmarking, Accuracy testing, Functional testing etc)
+
+Since these steps are reusable across different tests, we can swap the container used for each step.
+
+1. Initialize experiment
+
+    a. (Optional) deploy model
+
+    b. wait for the model to be ready
+
+2. Run Benchmarking test using configs and benchmark container (genai-perf, ai perf or 3rd party tool)
+
+    a. Prepare configs (matrix of params: isl/osl, concurrency, etc) 
+        pass as a config file to the harness container
+
+    b. Run test for each config
+
+3. Teardown
+
+    a. (Optional) Collect artifacts - push files to upstream storage (s3/minio)
+
+    b. Collect output results: 
+        Push benchmark metrics to a data storage layer (s3/minio/database) using a cli tool
+
+4. Analytics:
+    a. Generate charts, graphs, and tables from the benchmark metrics
+
+
+## Config
+
+Benchmarking config file:
+```yaml
+name: "blueprint-name"
+model: 
+    name: "RedHat/Llama-3.3-70B-Instruct"
+    path: "/path/to/model"
+concurrency: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+endpoint: "/v1/chat/completions"
+endpoint_type: "chat"
+benchmark:
+    isl_osl
+    - [8192, 1024]
+    - [1024, 1024]
+    - [1024, 8192]
+```
+
+## Alternatives:
+
+### Alternative 1: Benchmarking as a first class citizen in dynamo
+
+```
+kind: DynamoBenchmark
+metadata:
+  name: vllm-agg-benchmark
+spec:
+  model:
+     modelRef: llama-3-70b-instruct-v1
+  config: 
+      model: "RedHat/Llama-3.3-70B-Instruct"
+      path: "/path/to/model"
+      concurrency: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+      endpoint: "/v1/chat/completions"
+      endpoint_type: "chat"
+      benchmark:
+          isl_osl
+          - [8192, 1024]
+          - [1024, 1024]
+          - [1024, 8192]
+```
+
+### Alternative 2: Benchmarking helm chart + workflow manager
+
+Simpler to manage and deploy.
+Reuse Argo workflows for the workflow manager to orchestrate deps and workflow.