Add workloads directory with llama.cpp, vllm, pytorch, and faiss benchmarks

Migrated from schedcp/workloads. Added faiss and llama.cpp source repos
as proper submodules. Removed large binary artifacts (.venv, build dirs,
dataset files, .so files, model weights) and added .gitignore.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: LinuxDev9002

.gitmodules

@@ -10,3 +10,9 @@
 [submodule "/home/yunwei37/workspace/gpu/gpu_ext/docs/gpu-ext-paper"]
 	path = /home/yunwei37/workspace/gpu/gpu_ext/docs/gpu-ext-paper
 	url = https://github.com/yunwei37/gpu-ext-paper
+[submodule "workloads/faiss/faiss"]
+	path = workloads/faiss/faiss
+	url = https://github.com/eunomia-bpf/faiss
+[submodule "workloads/llama.cpp/llama.cpp"]
+	path = workloads/llama.cpp/llama.cpp
+	url = https://github.com/eunomia-bpf/llama.cpp

workloads/.gitignore

@@ -0,0 +1,32 @@
+# Virtual environments
+.venv/
+venv/
+env/
+
+# Build artifacts
+build/
+*.so
+*.o
+*.a
+
+# Python cache
+__pycache__/
+*.pyc
+
+# Large datasets (download via scripts)
+datasets/
+
+# Model weights
+*.gguf
+*.safetensors
+*.bin
+*.pt
+*.pth
+
+# IDE / editor
+.claude/
+.vscode/
+.idea/
+
+# Logs
+*.log

workloads/faiss/DATASET.md

@@ -0,0 +1,244 @@
+# FAISS Dataset Documentation
+
+## Overview
+
+This document describes the datasets used for FAISS benchmarking, their organization, file formats, and usage patterns.
+
+## Dataset Location
+
+All benchmark datasets are stored in:
+```
+/home/yunwei37/workspace/gpu/schedcp/workloads/faiss/faiss/benchs/bigann/
+```
+
+## SIFT1B (BigANN) Dataset
+
+### Description
+- **Name**: SIFT1B (also known as BigANN)
+- **Source**: http://corpus-texmex.irisa.fr/
+- **Purpose**: Billion-scale similarity search benchmark
+- **Vector Type**: SIFT descriptors (128-dimensional unsigned byte vectors)
+- **Total Size**: ~116 GB (with compressed archives)
+
+### Dataset Files
+
+#### Base Vectors (`bigann_base.bvecs`)
+- **Size**: 47 GB (uncompressed)
+- **Vectors**: 375,489,504 vectors (375M out of 1B total)
+- **Format**: bvecs (binary vectors)
+- **Structure**: Each vector is 128 dimensions × uint8
+- **Memory Layout**:
+  ```
+  [dim:4bytes][data:128bytes][dim:4bytes][data:128bytes]...
+  ```
+- **Usage**: Primary dataset for index building and search
+
+**Available Scales:**
+- SIFT1M: 1,000,000 vectors (~122 MB)
+- SIFT2M: 2,000,000 vectors (~244 MB)
+- SIFT5M: 5,000,000 vectors (~610 MB)
+- SIFT10M: 10,000,000 vectors (~1.22 GB)
+- SIFT20M: 20,000,000 vectors (~2.44 GB)
+- SIFT50M: 50,000,000 vectors (~6.1 GB)
+- SIFT100M: 100,000,000 vectors (~12.2 GB)
+- SIFT200M: 200,000,000 vectors (~24.4 GB)
+- SIFT375M: 375,489,504 vectors (~45.8 GB) - **Current Maximum**
+
+#### Learn Vectors (`bigann_learn.bvecs`)
+- **Size**: 13 GB
+- **Vectors**: 100,000,000 vectors
+- **Format**: bvecs
+- **Usage**: Training data for learning index quantizers and centroids
+
+#### Query Vectors (`bigann_query.bvecs`)
+- **Size**: 1.3 MB
+- **Vectors**: 10,000 vectors
+- **Format**: bvecs
+- **Usage**: Query set for evaluating search accuracy and performance
+
+### Ground Truth Files (`gnd/` directory)
+
+Located in `faiss/benchs/bigann/gnd/`, these files provide exact nearest neighbor results for accuracy evaluation.
+
+**Index Files (idx_*.ivecs)**: Nearest neighbor IDs
+- `idx_1M.ivecs` - Ground truth for SIFT1M (39 MB)
+- `idx_2M.ivecs` - Ground truth for SIFT2M (39 MB)
+- `idx_5M.ivecs` - Ground truth for SIFT5M (39 MB)
+- `idx_10M.ivecs` - Ground truth for SIFT10M (39 MB)
+- `idx_20M.ivecs` - Ground truth for SIFT20M (39 MB)
+- `idx_50M.ivecs` - Ground truth for SIFT50M (39 MB)
+- `idx_100M.ivecs` - Ground truth for SIFT100M (39 MB)
+- `idx_200M.ivecs` - Ground truth for SIFT200M (39 MB)
+- `idx_500M.ivecs` - Ground truth for SIFT500M (39 MB)
+- `idx_1000M.ivecs` - Ground truth for SIFT1000M (39 MB)
+
+**Distance Files (dis_*.fvecs)**: Distances to nearest neighbors
+- `dis_1M.fvecs` through `dis_1000M.fvecs` (39 MB each)
+
+**Ground Truth Format:**
+- Each file contains 10,000 rows (one per query)
+- Each row contains 1,000 nearest neighbor indices/distances
+- Format: `[n:4bytes][ids/dists:n*4bytes]...`
+
+## File Formats
+
+### BVECS Format (Binary Vectors, uint8)
+```
+Structure per vector:
+- Dimension (d): 4 bytes (int32)
+- Vector data: d bytes (uint8)
+
+Example for 128D:
+[128][v0][v1]...[v127][128][v0][v1]...[v127]...
+```
+
+**Reading BVECS in Python:**
+```python
+import numpy as np
+
+def mmap_bvecs(fname):
+    """Memory-map a bvecs file"""
+    x = np.memmap(fname, dtype='uint8', mode='r')
+    d = x[:4].view('int32')[0]
+    return x.reshape(-1, d + 4)[:, 4:]
+
+# Usage
+xb = mmap_bvecs('faiss/benchs/bigann/bigann_base.bvecs')
+print(f"Shape: {xb.shape}")  # (375489504, 128)
+print(f"Dtype: {xb.dtype}")  # uint8
+```
+
+### IVECS Format (Integer Vectors, int32)
+```
+Structure per vector:
+- Dimension (d): 4 bytes (int32)
+- Vector data: d * 4 bytes (int32)
+
+Used for: Ground truth indices
+```
+
+**Reading IVECS in Python:**
+```python
+import numpy as np
+
+def ivecs_read(fname):
+    """Read ivecs file"""
+    a = np.fromfile(fname, dtype='int32')
+    d = a[0]
+    return a.reshape(-1, d + 1)[:, 1:].copy()
+
+# Usage
+gt_I = ivecs_read('faiss/benchs/bigann/gnd/idx_100M.ivecs')
+print(f"Shape: {gt_I.shape}")  # (10000, 1000) - 10K queries, 1000 neighbors each
+```
+
+### FVECS Format (Float Vectors, float32)
+```
+Structure per vector:
+- Dimension (d): 4 bytes (int32)
+- Vector data: d * 4 bytes (float32)
+
+Used for: Ground truth distances, embeddings
+```
+
+## Memory-Mapped File Usage
+
+Benchmark scripts use memory-mapped files to avoid loading entire datasets into RAM:
+
+```python
+# Memory-mapped access (recommended for large files)
+xb = np.memmap('bigann_base.bvecs', dtype='uint8', mode='r')
+d = xb[:4].view('int32')[0]
+xb = xb.reshape(-1, d + 4)[:, 4:]
+
+# Access subset without loading full file
+subset = xb[:10_000_000]  # First 10M vectors
+```
+
+**Benefits:**
+- Only accessed portions loaded into RAM
+- Enables working with datasets larger than available memory
+- OS handles caching and paging automatically
+
+## Dataset Size Selection
+
+### By GPU Memory
+
+**8 GB VRAM:**
+- Direct GPU: SIFT10M (~5 GB as float32)
+- With UVM: SIFT50M (~25 GB as float32, requires 32+ GB RAM)
+
+**24 GB VRAM (RTX 4090/5090):**
+- Direct GPU: SIFT50M (~25 GB as float32)
+- With UVM: SIFT200M (~100 GB as float32, requires 128+ GB RAM)
+
+**32 GB VRAM:**
+- Direct GPU: SIFT100M (~50 GB as float32)
+- With UVM: SIFT375M (~190 GB as float32, requires 256+ GB RAM)
+
+### Dataset Trimming in Benchmarks
+
+Benchmark scripts automatically trim datasets to requested size:
+
+```python
+# From bench_gpu_1bn.py
+dbsize = 100  # Request 100M vectors
+
+# Load and trim
+xb = mmap_bvecs('faiss/benchs/bigann/bigann_base.bvecs')
+xb = xb[:dbsize * 1000 * 1000]  # Trim to 100,000,000 vectors
+
+print(f"Using {xb.shape[0]} vectors")
+```
+
+## Downloading Additional Data
+
+Currently available: **375M vectors** (47 GB)
+
+To download the full 1B dataset:
+
+```bash
+cd /home/yunwei37/workspace/gpu/schedcp/workloads/faiss/faiss/benchs/bigann
+
+# Download remaining base vectors (warning: ~120 GB total when complete)
+wget ftp://ftp.irisa.fr/local/texmex/corpus/bigann_base.bvecs
+
+# Or download compressed version
+wget ftp://ftp.irisa.fr/local/texmex/corpus/bigann_base.bvecs.gz
+gunzip bigann_base.bvecs.gz
+```
+
+**Download verification:**
+```bash
+# Check file size
+ls -lh bigann_base.bvecs
+
+# Count vectors
+python3 -c "
+import numpy as np
+x = np.memmap('bigann_base.bvecs', dtype='uint8', mode='r')
+d = x[:4].view('int32')[0]
+n = len(x) // (d + 4)
+print(f'Vectors: {n:,}')
+print(f'Dimensions: {d}')
+"
+```
+
+## Dataset Corruption and Repair
+
+If you encounter reshape errors, the file may be incomplete or corrupted:
+
+```bash
+# Fix truncated bvecs file
+python3 -c "
+import os
+size = os.path.getsize('faiss/benchs/bigann/bigann_base.bvecs')
+vector_size = 132  # 4 bytes (dim) + 128 bytes (data)
+complete_size = (size // vector_size) * vector_size
+if size != complete_size:
+    os.truncate('faiss/benchs/bigann/bigann_base.bvecs', complete_size)
+    print(f'Truncated to {complete_size:,} bytes ({(size-complete_size):+,} bytes)')
+else:
+    print('File is valid')
+"
+```