han.zhao/test_gpu_scripts
1
0
Fork 0
Code Issues Pull Requests 2 Actions Packages Projects Releases Wiki Activity

Add multinode NCCL diagnostic report

Browse Source
This commit is contained in:
cs 2026-05-23 15:39:15 +08:00
parent 4b17bafd53
commit 4b93fc785f
5 changed files with 284 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
configs/default.yaml
View File

@ -90,6 +90,7 @@ multinode_nccl:
net_plugin: none
nvls_enable: 1
split_data_on_qps: 1
extra_env: {}
min_peak_busbw_gbps:
allreduce: 480
alltoall: 75

60
configs/multinode_nccl_diagnostic.yaml Normal file
View File

@ -0,0 +1,60 @@
tools:
install_dir: /opt/gpu-test-tools
report:
output_dir: ./reports
format: md
multinode_nccl:
enabled: true
mode: diagnostic
hosts:
- name: nccl-gpu-1
addr: 172.72.8.12
slots: 8
- name: nccl-gpu-2
addr: 172.72.8.16
slots: 8
ssh_user: root
ssh_preflight: true
mpirun_path: /usr/mpi/gcc/openmpi-4.1.9a1/bin/mpirun
mpi_ld_preload: null
extra_ld_library_path:
- /usr/mpi/gcc/openmpi-4.1.9a1/lib
- /root/gpu-test-venv/lib/python3.10/site-packages/nvidia/nccl/lib
- /usr/local/cuda-12.4/targets/x86_64-linux/lib
nccl_tests_dir: null
tests:
- all_reduce_perf
- alltoall_perf
topologies:
- nodes: 2
gpus_per_node: 8
label: 2 nodes x 8 GPUs diagnostic
begin_size: 256M
end_size: 256M
step_factor: 2
warmup_iters: 1
iters: 3
gpus_per_rank: 1
timeout_sec: 600
debug: INFO
socket_ifname: bond0
ib_gid_index: 3
ib_sl: 5
ib_tc: 136
ib_hca: mlx5_0,mlx5_1,mlx5_6,mlx5_7
ib_timeout: 22
qps_per_connection: 4
min_nchannels: 4
net_plugin: none
nvls_enable: 1
split_data_on_qps: 1
extra_env:
NCCL_DEBUG_SUBSYS: INIT,NET
NCCL_NET_GDR_LEVEL: 5
NCCL_NET_GDR_READ: 1
NCCL_DMABUF_ENABLE: 0
min_peak_busbw_gbps:
allreduce: 480
alltoall: 75

23
modules/report.py
View File

@ -492,6 +492,29 @@ class ReportGenerator:
f"{threshold_text} | {topo.get('status', '?')} |"
)
lines.append("")
diag_rows = []
for topo in data.get("topologies", []):
net = topo.get("network") or {}
if net:
diag_rows.append((topo, net))
if diag_rows:
lines.append("| Topology | NCCL Network | GPU Direct RDMA | GDR Disabled HCAs |")
lines.append("|----------|--------------|-----------------|-------------------|")
for topo, net in diag_rows:
networks = ", ".join(net.get("networks") or []) or "unknown"
gdr = net.get("gpu_direct_rdma", "UNKNOWN")
disabled = ", ".join(net.get("gdr_disabled_hcas") or []) or "-"
lines.append(f"| {topo.get('label', '')} | {networks} | {gdr} | {disabled} |")
lines.append("")
failed_topos = [topo for topo in data.get("topologies", []) if topo.get("status") == "FAIL"]
if failed_topos:
lines.append("| Topology | Return Code | Error / Output Tail |")
lines.append("|----------|-------------|---------------------|")
for topo in failed_topos:
tail = topo.get("error") or topo.get("stderr_tail") or topo.get("stdout_tail") or ""
tail = str(tail).replace("\n", " ").replace("|", "\\|")[-240:]
lines.append(f"| {topo.get('label', '')} | {topo.get('returncode', '')} | {tail} |")
lines.append("")
lines.append(f"**Overall: {'PASS' if multinode.get('passed') else 'FAIL'}**\n")
elif multinode and multinode.get("error"):
lines.append("## Multi-node NCCL / Cross Leaf\n")

134
reports_multinode_nccl_diagnosis_20260523.md Normal file
View File

@ -0,0 +1,134 @@
# 多机多卡 NCCL 诊断报告
- 日期：2026-05-23
- 测试入口：`nccl-gpu-1` / `aikubeworker0012` / `172.72.8.12`
- 对端节点：`nccl-gpu-2` / `aikubeworker0016` / `172.72.8.16`
- 诊断配置：`configs/multinode_nccl_diagnostic.yaml`
- 原始脚本报告：`reports_multinode_nccl_diagnostic_2x8_debug_v2.md`
## 当前结论
这不是单纯 “IB 不通” 的问题。底层 CUDA RDMA perftest 可以跑到接近单端口 400Gb/s 的水平，但 NCCL 在实际 2 节点通信时把 GPU Direct RDMA 禁用了，导致 NCCL 带宽显著低于验收阈值。
同时，`nccl-gpu-2` 的 SSH 入口不稳定，会造成 `mpirun` 拉起远端 rank 失败。这个问题会直接影响 alltoall 等多机测试的稳定性，需要和 NCCL GDR 问题一起处理。
## 已完成的修正
1. 修正 `mpirun` 使用路径，避开系统 `/usr/bin/mpirun` 与 DOCA OpenMPI 动态库混用导致的崩溃。
2. 补充 `LD_LIBRARY_PATH`，确保 `mpirun`、CUDA、pip 安装的 NCCL 动态库可同时解析。
3. 将 NCCL HCA 限定到 400Gb/s 活跃端口：`mlx5_0,mlx5_1,mlx5_6,mlx5_7`。
4. 在脚本中加入 multi-node NCCL 网络诊断解析，报告会展示 `NCCL Network`、`GPU Direct RDMA`、`GDR Disabled HCAs`。
5. 增加 `multinode_nccl.extra_env`，可以在配置里快速试 NCCL 环境变量，不需要改代码。
6. 增加诊断配置 `configs/multinode_nccl_diagnostic.yaml`，固定跑 2 节点 x 8 GPU、256M、`NCCL_DEBUG=INFO` 和 `NCCL_DEBUG_SUBSYS=INIT,NET`。
## 关键证据
### 1. CUDA RDMA perftest 通过
命令类型：
```bash
CUDA_VISIBLE_DEVICES=0 ib_write_bw -d mlx5_0 -i 1 --use_cuda=0 -s 4194304 -F --report_gbits 172.72.8.16
```
结果：
| 测试 | 设备 | GPU | 平均带宽 | 结论 |
|------|------|-----|----------|------|
| `ib_write_bw --use_cuda` | `mlx5_0` | GPU0 | `387.16 Gb/s` | PASS |
解释：GPU 内存参与 RDMA 写带宽测试可以接近 400Gb/s，说明 `nvidia_peermem`/经典 GPUDirect RDMA 路径并非完全不可用。
### 2. CUDA DMA-BUF 路径不可用
命令类型：
```bash
CUDA_VISIBLE_DEVICES=0 ib_write_bw -d mlx5_0 -i 1 --use_cuda=0 --use_cuda_dmabuf -s 4194304 -F --report_gbits 172.72.8.16
```
结果：
| 测试 | 输出 | 结论 |
|------|------|------|
| `ib_write_bw --use_cuda_dmabuf` | `DMA-BUF is not supported on this GPU` | FAIL |
解释：当前环境不能走 CUDA DMA-BUF RDMA。后续 NCCL 应优先确认是否能稳定走经典 `nvidia_peermem` 路径。
### 3. NCCL 单卡跨节点仍禁用 GDR
已经尝试：
- `NCCL_NET_GDR_LEVEL=SYS`
- `NCCL_NET_GDR_LEVEL=5`
- `NCCL_NET_GDR_READ=1`
- `NCCL_DMABUF_ENABLE=0`
- `NCCL_IB_CUDA_SUPPORT=1`
- `NCCL_IB_HCA=mlx5_0`
结果仍显示：
```text
NCCL INFO Using network IB
NCCL INFO NET/IB : GPU Direct RDMA Disabled for HCA 0 'mlx5_0'
```
256M allreduce 约 `13.4 GB/s`，明显低于 400Gb/s IB 端口能力。
### 4. 脚本 2 节点 x 8 GPU 诊断结果
原始报告：`reports_multinode_nccl_diagnostic_2x8_debug_v2.md`
| Operation | Topology | Peak Bus BW | Threshold | Status | NCCL Network | GPU Direct RDMA |
|-----------|----------|-------------|-----------|--------|--------------|-----------------|
| allreduce | 2 nodes x 8 GPUs | `68.69 GB/s` | `>= 480 GB/s` | FAIL | IB | DISABLED |
| alltoall | 2 nodes x 8 GPUs | `0.00 GB/s` | `>= 75 GB/s` | FAIL | unknown | UNKNOWN |
allreduce 失败原因是带宽不达标，且报告捕获到 GDR 被 NCCL 禁用：
| GDR Disabled HCAs |
|-------------------|
| `mlx5_0, mlx5_1, mlx5_6, mlx5_7` |
alltoall 失败原因这轮不是性能本身，而是 `mpirun` 阶段受 SSH/网络发现影响失败，报告尾部显示：
```text
lack of common network interfaces and/or no route found between them
```
## 当前阻塞
### 阻塞 1：NCCL 禁用 GPU Direct RDMA
现象：
- IB 能被 NCCL 识别：`Using network IB`
- 400Gb/s HCA 被 NCCL 选中：`mlx5_0, mlx5_1, mlx5_6, mlx5_7`
- 但 NCCL 明确禁用 GDR：`GPU Direct RDMA Disabled`
- perftest 的经典 CUDA RDMA 又能跑到 `387.16 Gb/s`
判断：底层 RDMA 能力存在，但 NCCL 的 GDR 判定/注册路径没有打通。优先排查 NCCL 与 NVIDIA driver、OFED、`nvidia_peermem`、NCCL net plugin/内部 IB 后端之间的兼容性。
### 阻塞 2：`nccl-gpu-2` SSH 不稳定
现象：
- 多次出现：`kex_exchange_identification: Connection closed by remote host`
- MCP 直连 `nccl-gpu-2` 也会失败或长时间超时
- `mpirun` 依赖 SSH 拉起远端 rank，因此 SSH 抖动会让 alltoall 这类测试直接没有有效输出
判断：需要先处理 `aikubeworker0016` 的 SSHD/连接限制/MaxStartups/安全策略，否则多机测试无法稳定复现。
## 建议下一步
1. 先修 `nccl-gpu-2` SSH 稳定性：检查 `sshd_config` 的 `MaxStartups`、连接限制、安全审计组件，以及是否有过多半开 SSH 会话。
2. 对两台机器分别确认 `nvidia_peermem` 参数、OFED 版本、NVIDIA driver 版本一致性。
3. 在两台机器上测试是否需要切换 `nvidia_peermem peerdirect_support` 模式，并在变更前确认没有正在运行的业务任务。
4. 尝试安装或启用匹配当前 OFED/driver 的 NCCL net plugin；当前日志显示 `No plugin found (libnccl-net.so)`，NCCL 使用的是 internal network plugin。
5. SSH 稳定后重跑完整多机配置：2 节点 x 8 GPU，至少覆盖 `all_reduce_perf` 和 `alltoall_perf`，消息大小从 `1K` 到 `16G`。
## 当前可交付物
- `configs/multinode_nccl_diagnostic.yaml`：多机多卡诊断配置
- `reports_multinode_nccl_diagnostic_2x8_debug_v2.md`：脚本生成的原始 2x8 诊断报告
- `reports_multinode_nccl_diagnosis_20260523.md`：本中文诊断总结

66
reports_multinode_nccl_diagnostic_2x8_debug_v2.md Normal file
View File

@ -0,0 +1,66 @@
# GPU Test Report
- **Date:** 2026-05-23T07:37:41.426792
- **Host:** aikubeworker0012
## Overall Acceptance Verdict
**Result: FAIL**
Missing required evidence:
- GPU Info
- Health Check
- Memory Bandwidth
- Compute Throughput
- NVLink/NVSwitch
- NCCL
- Stress Test
- RDMA
- DCGM
- Training
## Summary
| Test | Result |
|------|--------|
| Multi-node NCCL | FAIL |
## Multi-node NCCL / Cross Leaf
Source: nccl-tests-mpirun | Mode: diagnostic
- **Hosts:** nccl-gpu-1(172.72.8.12), nccl-gpu-2(172.72.8.16)
- **Preflight:** PASS (1 warnings)
### Multi-node NCCL allreduce
| Topology | Peak Bus BW | Peak Size | Avg Bus BW | Threshold | Status |
|----------|-------------|-----------|------------|-----------|--------|
| 2 nodes x 8 GPUs diagnostic | 68.69 GB/s | 256M | 68.21 GB/s | >= 480 GB/s | FAIL |
| Topology | NCCL Network | GPU Direct RDMA | GDR Disabled HCAs |
|----------|--------------|-----------------|-------------------|
| 2 nodes x 8 GPUs diagnostic | IB | DISABLED | mlx5_0, mlx5_1, mlx5_6, mlx5_7 |
| Topology | Return Code | Error / Output Tail |
|----------|-------------|---------------------|
| 2 nodes x 8 GPUs diagnostic | 0 | aikubeworker0012:2139504:2139504 [0] NCCL INFO comm 0x55646d15f590 rank 0 nranks 16 cudaDev 0 busId 18000 - Destroy COMPLETE # Out of bounds values : 0 OK # Avg bus bandwidth : 68.2135 # # Collective test concluded: all_reduce_perf # |
### Multi-node NCCL alltoall
| Topology | Peak Bus BW | Peak Size | Avg Bus BW | Threshold | Status |
|----------|-------------|-----------|------------|-----------|--------|
| 2 nodes x 8 GPUs diagnostic | 0.00 GB/s | | 0.00 GB/s | >= 75 GB/s | FAIL |
| Topology | NCCL Network | GPU Direct RDMA | GDR Disabled HCAs |
|----------|--------------|-----------------|-------------------|
| 2 nodes x 8 GPUs diagnostic | unknown | UNKNOWN | - |
| Topology | Return Code | Error / Output Tail |
|----------|-------------|---------------------|
| 2 nodes x 8 GPUs diagnostic | 255 | lack of common network interfaces and/or no route found between them. Please check network connectivity (including firewalls and network routing requirements). -------------------------------------------------------------------------- |
**Overall: FAIL**
---
*Generated by GPU Test Suite v0.2.0*