5.3 KiB
5.3 KiB
MANUAL NVMe Hybrid Test Instructions
Run these commands ON the-craw server
Since remote SSH seems to have issues, here are the exact commands to run directly on the-craw to test the NVMe hybrid system.
🎯 Goal: Test the three-tiered memory hierarchy with working large grid
📋 Prerequisites Check (Run on the-craw):
# 1. Check GPU
nvidia-smi
# 2. Check NVMe storage
lsblk | grep -i nvme
df -h | grep -i nvme
# 3. Check CUDA
nvcc --version
# 4. Check system
uname -a
free -h
🚀 Step 1: Create Test Directory
# Create directory for NVMe test
mkdir -p ~/fractal_nvme_test
cd ~/fractal_nvme_test
mkdir -p nvme_states
📦 Step 2: Get Source Files
You need these files from Beast:
probe_256.cufractal_habit_256_full.cuadd_power_limit.cu
Copy them manually or use SCP from Beast:
# FROM Beast, run:
scp probe_256.cu tiger@192.168.1.55:~/fractal_nvme_test/
scp fractal_habit_256_full.cu tiger@192.168.1.55:~/fractal_nvme_test/
scp add_power_limit.cu tiger@192.168.1.55:~/fractal_nvme_test/
🔧 Step 3: Compile on the-craw
cd ~/fractal_nvme_test
# Detect GPU architecture first
GPU_ARCH="sm_61" # Default for GTX 1050/1060
# If you have RTX card, use sm_75 for 20-series, sm_86 for 30-series
# Compile
nvcc -O3 -arch=$GPU_ARCH -o probe_256_nvme probe_256.cu -lnvml
nvcc -O3 -arch=$GPU_ARCH -o fractal_habit_256_nvme fractal_habit_256_full.cu -lnvml -lcufft
nvcc -O3 -arch=$GPU_ARCH -o set_power_limit add_power_limit.cu -lnvml
# Make executable
chmod +x probe_256_nvme fractal_habit_256_nvme set_power_limit
🔬 Step 4: Quick NVMe Test
cd ~/fractal_nvme_test
# Test 1: Check if we can write to NVMe
NVME_PATH="/mnt/nvme"
if [ ! -d "$NVME_PATH" ]; then
# Try to find NVMe
NVME_DEVICE=$(lsblk -o NAME,TYPE,MOUNTPOINT | grep 'nvme.*disk' | head -1)
if [ -n "$NVME_DEVICE" ]; then
echo "Found NVMe: $NVME_DEVICE"
# Use home directory if not mounted
NVME_PATH="~/nvme_test"
mkdir -p "$NVME_PATH"
else
echo "No NVMe found, using local directory"
NVME_PATH="./nvme_states"
fi
fi
echo "Using storage: $NVME_PATH/fractal_states"
mkdir -p "$NVME_PATH/fractal_states"
# Test write speed
echo "Testing write speed..."
time dd if=/dev/zero of="$NVME_PATH/fractal_states/test.bin" bs=1M count=100 oflag=direct
🎪 Step 5: Run the Actual Test
cd ~/fractal_nvme_test
# Monitor GPU in background (in separate terminal)
# Terminal 1:
watch -n 1 nvidia-smi
# Terminal 2: Run the test
./probe_256_nvme 2>&1 | tee nvme_test_output.log
📊 Step 6: What to Look For
Expected Output:
- First: 13 "NEW GUARDIAN" messages
- Cycles 0-599: Warmup, guardian formation
- Cycles 600-649: "INJ" in probe column (mass injection)
- Cycle 800: Probe B (shear rotation)
- Cycles 1100-1199: Probe C (VRM silence) - EXPECTED CRASH HERE
- If no crash: Continue to Probe D (1400-1499)
Critical Metrics:
- Power draw: Should be reasonable for your GPU
- Temperature: Should stay below 80°C
- Memory usage: Should stay within GPU VRAM
- Crash point: Note the exact cycle if it crashes
🛠️ Step 7: If It Works (No Crash)
If it runs past 1112 without crashing:
# Let it run longer
./probe_256_nvme 2>&1 | tee long_run.log
# Or test with power limits
sudo nvidia-smi -pl 100 # Set power limit (adjust for your GPU)
./fractal_habit_256_nvme
📝 Step 8: Report Back
Tell me:
- GPU model: (from
nvidia-smi) - NVMe status: (found/not found, path)
- Test result: (ran/crashed/errors)
- If crashed: At what cycle? Error message?
- Power/temp: What were the readings?
- Guardian count: How many formed?
⚡ Quick Test Script
Save this as quick_test.sh on the-craw:
#!/bin/bash
cd ~/fractal_nvme_test
echo "Starting NVMe hybrid test..."
echo "GPU: $(nvidia-smi --query-gpu=name --format=csv,noheader)"
echo "Time: $(date)"
echo ""
./probe_256_nvme 2>&1 | head -100
🆘 Troubleshooting:
If compilation fails:
# Check CUDA
nvcc --version
# Check libraries
ldconfig -p | grep nvml
# Try different architecture
nvcc -O3 -arch=sm_75 -o probe_256_nvme probe_256.cu -lnvml
If no NVMe found:
# Check storage
lsblk
sudo fdisk -l
# Use regular SSD/HDD for test
mkdir -p ~/fractal_states
# Update code to use this path
If permission issues:
# Check file permissions
ls -la probe_256_nvme
chmod +x probe_256_nvme
# Check write permissions
touch ~/fractal_nvme_test/test.txt
🎯 The Core Question:
Does the three-tiered memory hierarchy work with NVMe storage?
We know the grid works. We know guardians form. Now we need to test if:
- State can be saved to NVMe (crystallized memory)
- System can recover from NVMe state
- Performance is acceptable with NVMe writes
📞 Next Action:
Run the quick test on the-craw and tell me what happens:
cd ~/fractal_nvme_test
./probe_256_nvme 2>&1 | head -50
Just those 50 lines will tell us:
- If it compiles and runs
- How many guardians form
- What the initial power draw is
- If there are any immediate errors
That's all we need to start. Then we can implement the actual NVMe checkpointing based on the results.