wmantly/turing-multi-gpu-llm-server
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs(lxc): add NUMA socket pinning recommendations and benchmark metrics

Browse Source
This commit is contained in:
wmantly
2026-09-02 20:14:49 +00:00
parent d26bcb34d3
commit 9e543f1291
1 changed files with 25 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/PROXMOX_LXC_GUIDE.md
+25
View File
@@ -31,6 +31,10 @@ lxc.mount.entry: /dev/nvidiactl dev/nvidiactl none bind,optional,create=file
lxc.mount.entry: /dev/nvidia-uvm dev/nvidia-uvm none bind,optional,create=file
lxc.mount.entry: /dev/nvidia-uvm-tools dev/nvidia-uvm-tools none bind,optional,create=file
lxc.mount.entry: /dev/nvidia-modeset dev/nvidia-modeset none bind,optional,create=file
# 5. CPU Socket & NUMA Pinning (Crucial for Multi-GPU AllReduce Performance)
# If your GPUs are attached to PCIe Riser 2 / CPU Socket 2 (NUMA Node 1), pin the container:
cpuset: 6-11,18-23
```
---
@@ -62,3 +66,24 @@ If all four commands return with code `0` and "All done", your container has ful
When running services that execute `nvidia-smi` hardware commands inside systemd:
* Do **NOT** set `NoNewPrivileges=true` in `gpu-power-governor.service`. `NoNewPrivileges=true` blocks processes from acquiring permissions to execute privileged NVML clock-locking calls.
---
## 4. NUMA Node & CPU Socket Alignment (Performance Critical)
On dual-socket servers (such as HPE ProLiant DL380p Gen8 with Dual Intel Xeon E5-2620 v2):
* **Socket 1 (NUMA Node 0)**: CPUs `0-5, 12-17` (Wires to Riser 1 PCIe slots)
* **Socket 2 (NUMA Node 1)**: CPUs `6-11, 18-23` (Wires to Riser 2 PCIe slots)
If your GPUs are seated in PCIe Riser 2 (Bus `0x21`, `0x24`, `0x27`), failing to pin CPU affinity causes worker threads and CUDA driver calls to constantly cross the QPI bus interconnect.
### Verified Benchmark Impact (Socket 2 Pinned vs. Unpinned)
| Workload | Unpinned (All Cores 0–23) | Pinned to Socket 2 (`cpuset: 6-11,18-23`) | Improvement |
| :--- | :--- | :--- | :--- |
| **Short Prompt Prefill Latency** | 2,798 ms (24.7 tok/s) | **1,989 ms (34.7 tok/s)** | **+40.7% faster prefill** ⚡ |
| **Medium Prompt (~1k tokens)** | 281.87 tok/s | **281.50 tok/s** | Consistent |
| **Long Prompt (~4.8k tokens)** | 384.01 tok/s | **384.31 tok/s** | Consistent |
| **Sustained Decode** | 35–38 tok/s | **35–38 tok/s** | Consistent |
**Recommendation**: Always pair your GPU physical PCIe slot placement with container `cpuset` pinning on the host to eliminate cross-QPI memory bounce.