[Performance]: Distributed Inference w/ & w/o RDMA over Infiniband (tp=8, pp=2)

[surajssd]

Description

During distributed inference I didn't see a lot of performance improvement when using Infiniband vs. not using Infiniband. I have followed the recommendations by the project here on doing distributed inference. And when doing inference with IB enabled you can see vLLM logs like these (source):

deepseek-v2-5-0:5969:5969 [0] NCCL INFO Channel 00/0 : 1[0] -> 0[0] [receive] via NET/IB/0/GDRDMA
deepseek-v2-5-0:5969:5969 [0] NCCL INFO Channel 01/0 : 1[0] -> 0[0] [receive] via NET/IB/0/GDRDMA
deepseek-v2-5-0:5969:5969 [0] NCCL INFO Channel 00/0 : 0[0] -> 1[0] [send] via NET/IB/0/GDRDMA
deepseek-v2-5-0:5969:5969 [0] NCCL INFO Channel 01/0 : 0[0] -> 1[0] [send] via NET/IB/0/GDRDMA

This blog on the vLLM website also has some benchmarks and it also does not show any significant improvements of using IB or not using IB: https://blog.vllm.ai/2024/07/23/llama31.html

Report of performance regression

Here are the performance results

Test name	GPU	# of req.	Tput (req/s)	Output Tput (tok/s)	Total Tput (tok/s)	Mean TTFT (ms)	Median TTFT (ms)	P99 TTFT (ms)	Mean TPOT (ms)	Median TPOT (ms)	P99 TPOT (ms)	Mean ITL (ms)	Median ITL (ms)	P99 ITL (ms)
no_ib_tp8_pp2_qps_01	1xno-ib-Standard_ND96asr_v4 x 2	200	0.646012	137.901	290.344	236.677	215.085	1302.77	139.814	139.073	165.73	138.785	129.591	286.317
no_ib_tp8_pp2_qps_04	1xno-ib-Standard_ND96asr_v4 x 2	200	1.17664	250.984	528.643	237.706	223.34	703.65	173.128	170.547	232.624	158.083	141.287	451.196
no_ib_tp8_pp2_qps_16	1xno-ib-Standard_ND96asr_v4 x 2	200	1.46822	312.342	658.806	246.536	242.183	397.258	235.883	170.223	758.732	160.798	143.171	723.514
no_ib_tp8_pp2_qps_inf	1xno-ib-Standard_ND96asr_v4 x 2	200	1.57831	332.795	705.238	1746.93	1643.29	2415.8	165.243	148.692	378.552	144.618	142.733	175.876
ib_tp8_pp2_qps_01	1xib-Standard_ND96asr_v4 x 2	200	0.651377	139.014	292.722	227.47	203.671	1241.54	137.99	137.408	162.498	136.986	128.114	278.476
ib_tp8_pp2_qps_04	1xib-Standard_ND96asr_v4 x 2	200	1.18058	250.23	528.818	227.527	219.97	449.453	167.116	164.781	235.652	155.036	139.355	430.007
ib_tp8_pp2_qps_16	1xib-Standard_ND96asr_v4 x 2	200	1.46607	311.019	656.974	239.256	233.958	342.987	228.171	169.589	628.591	158.357	141.627	776.336
ib_tp8_pp2_qps_inf	1xib-Standard_ND96asr_v4 x 2	200	1.57666	335.703	707.755	1838.06	1882.46	2667.65	170.38	148.324	471.125	144.155	141.769	166.42

Visualization of the above table:

Color Legend: Green is with Infiniband and red is without Infiniband.

Throughput

TTFT

TPOT

ITL

Misc discussion on performance

Grafana based benchmarking charts:

• Kubernetes / Networking
  • Cluster
    • noIB: https://snapshots.raintank.io/dashboard/snapshot/JZAMCncIPJAfMopiZyDUtkOlgmOTSzoD
    • IB: https://snapshots.raintank.io/dashboard/snapshot/iFfNXOizotQEnRbPdtzuCkeHycGATTmw
  • Namespace (Pods)
    • noIB: https://snapshots.raintank.io/dashboard/snapshot/cUUTgmTIUV4MiBzWVIoGTG8Hsrtn9aij
    • IB: https://snapshots.raintank.io/dashboard/snapshot/QLfA2as55z4EoZP9Jf4zlg6oEWUiqu09
• Ray Dashboard
  • noIB: https://snapshots.raintank.io/dashboard/snapshot/j88xd3lTYhZdBi3DEuqr49jWfgCQaVk6
  • IB: https://snapshots.raintank.io/dashboard/snapshot/TOnvEXzk5Kv1xr9oAuM3H9wHW74X6SU9
• vLLM
  • noIB: https://snapshots.raintank.io/dashboard/snapshot/VxOwdopoS8gmS3nKCNwCuPROR5vxJs33
  • IB: https://snapshots.raintank.io/dashboard/snapshot/rVBcqQHk8gRdS8P0Mtp5lDGIOBirDW7K
• NVIDIA DCGM Exporter Dashboard
  • noIB: http://localhost:3000/dashboard/snapshot/zFZrTYTQivR6ShojvMIGynegBfFZISaY
  • IB: https://snapshots.raintank.io/dashboard/snapshot/0lKqQTIwImdDnFfbSzX2WiZSiJCPI6GE
• NVIDIA DCGM Exporter
  • noIB: https://snapshots.raintank.io/dashboard/snapshot/G0QzVoliNtCSGpehqtE0EXECkjKp2BlR
  • IB: https://snapshots.raintank.io/dashboard/snapshot/VToBHzkKcXsdByLPaIXYoVSn5hr24ike

The only significant difference I see from the above Grafana charts is, network usage on nodes. I see that the network usage without IB is higher because all the data transfer is happening on ethernet. But it is not significantly higher to warrant the saturation of either of the networks either ethernet or IB + ethernet.

Without IB

With IB

Your current environment (if you think it is necessary)

I am running vLLM with Ray for distributed inference on Azure Kubernetes Service with nodepool of two GPU machines of type Standard_ND96asr_v4. These machines have 8 A100 GPUs each and are connected via Infiniband. I have deployed the GPU operator and network operator on the Kubernetes cluster.

Ray Head Environment

INFO 04-22 13:50:51 [__init__.py:239] Automatically detected platform cuda.
Collecting environment information...
PyTorch version: 2.6.0+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 4.0.0
Libc version: glibc-2.35

Python version: 3.12.9 (main, Feb  5 2025, 08:49:00) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-1084-azure-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.4.131
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA A100-SXM4-40GB
GPU 1: NVIDIA A100-SXM4-40GB
GPU 2: NVIDIA A100-SXM4-40GB
GPU 3: NVIDIA A100-SXM4-40GB
GPU 4: NVIDIA A100-SXM4-40GB
GPU 5: NVIDIA A100-SXM4-40GB
GPU 6: NVIDIA A100-SXM4-40GB
GPU 7: NVIDIA A100-SXM4-40GB

Nvidia driver version: 570.124.06
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        48 bits physical, 48 bits virtual
Byte Order:                           Little Endian
CPU(s):                               96
On-line CPU(s) list:                  0-95
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 7V12 64-Core Processor
CPU family:                           23
Model:                                49
Thread(s) per core:                   1
Core(s) per socket:                   48
Socket(s):                            2
Stepping:                             0
BogoMIPS:                             4890.86
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl tsc_reliable nonstop_tsc cpuid extd_apicid aperfmperf tsc_known_freq pni pclmulqdq ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw topoext perfctr_core ssbd vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves clzero xsaveerptr rdpru arat umip rdpid
Hypervisor vendor:                    Microsoft
Virtualization type:                  full
L1d cache:                            3 MiB (96 instances)
L1i cache:                            3 MiB (96 instances)
L2 cache:                             48 MiB (96 instances)
L3 cache:                             384 MiB (24 instances)
NUMA node(s):                         4
NUMA node0 CPU(s):                    0-23
NUMA node1 CPU(s):                    24-47
NUMA node2 CPU(s):                    48-71
NUMA node3 CPU(s):                    72-95
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Mitigation; untrained return thunk; SMT disabled
Vulnerability Spec rstack overflow:   Mitigation; safe RET, no microcode
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Retpolines; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] flashinfer-python==0.2.1.post2+cu124torch2.6
[pip3] numpy==2.1.3
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.4.0
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.51.0
[pip3] triton==3.2.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.8.3
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
	�[4mGPU0	GPU1	GPU2	GPU3	GPU4	GPU5	GPU6	GPU7	NIC0	NIC1	NIC2	NIC3	NIC4	NIC5	NIC6	NIC7	NIC8	CPU Affinity	NUMA Affinity	GPU NUMA ID�[0m
GPU0	 X 	NV12	NV12	NV12	NV12	NV12	NV12	NV12	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	SYS	24-47	1		N/A
GPU1	NV12	 X 	NV12	NV12	NV12	NV12	NV12	NV12	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	SYS	24-47	1		N/A
GPU2	NV12	NV12	 X 	NV12	NV12	NV12	NV12	NV12	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	NODE	0-23	0		N/A
GPU3	NV12	NV12	NV12	 X 	NV12	NV12	NV12	NV12	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	NODE	0-23	0		N/A
GPU4	NV12	NV12	NV12	NV12	 X 	NV12	NV12	NV12	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	72-95	3		N/A
GPU5	NV12	NV12	NV12	NV12	NV12	 X 	NV12	NV12	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	72-95	3		N/A
GPU6	NV12	NV12	NV12	NV12	NV12	NV12	 X 	NV12	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	48-71	2		N/A
GPU7	NV12	NV12	NV12	NV12	NV12	NV12	NV12	 X 	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	48-71	2		N/A
NIC0	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	 X 	NODE	SYS	SYS	SYS	SYS	SYS	SYS	SYS				
NIC1	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	 X 	SYS	SYS	SYS	SYS	SYS	SYS	SYS				
NIC2	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	 X 	NODE	SYS	SYS	SYS	SYS	NODE				
NIC3	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	 X 	SYS	SYS	SYS	SYS	NODE				
NIC4	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	 X 	NODE	SYS	SYS	SYS				
NIC5	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	 X 	SYS	SYS	SYS				
NIC6	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	 X 	NODE	SYS				
NIC7	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	 X 	SYS				
NIC8	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	 X 				

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NIC Legend:

  NIC0: mlx5_0
  NIC1: mlx5_1
  NIC2: mlx5_2
  NIC3: mlx5_3
  NIC4: mlx5_4
  NIC5: mlx5_5
  NIC6: mlx5_6
  NIC7: mlx5_7
  NIC8: mlx5_8

NVIDIA_VISIBLE_DEVICES=GPU-1a18aca5-b383-a074-4bc0-f7025b0bff6c,GPU-a4d0a79c-d538-2e80-9988-e5093de93bc8,GPU-3e5f5fb6-3b78-42c5-d38e-cf0e0ce1bc2e,GPU-089f42e2-1c0a-dcfb-7ceb-d24edf08742b,GPU-8707c7b7-0bb6-4e9c-61ed-ee4817d4f1a3,GPU-703db7c6-320b-536d-353e-5fed1b01c811,GPU-4361b490-d47b-4c91-7ed2-fd9883567cfe,GPU-e659ab71-db2e-6ca1-8977-01f72a2e53eb
NVIDIA_REQUIRE_CUDA=cuda>=12.4 brand=tesla,driver>=470,driver<471 brand=unknown,driver>=470,driver<471 brand=nvidia,driver>=470,driver<471 brand=nvidiartx,driver>=470,driver<471 brand=geforce,driver>=470,driver<471 brand=geforcertx,driver>=470,driver<471 brand=quadro,driver>=470,driver<471 brand=quadrortx,driver>=470,driver<471 brand=titan,driver>=470,driver<471 brand=titanrtx,driver>=470,driver<471 brand=tesla,driver>=525,driver<526 brand=unknown,driver>=525,driver<526 brand=nvidia,driver>=525,driver<526 brand=nvidiartx,driver>=525,driver<526 brand=geforce,driver>=525,driver<526 brand=geforcertx,driver>=525,driver<526 brand=quadro,driver>=525,driver<526 brand=quadrortx,driver>=525,driver<526 brand=titan,driver>=525,driver<526 brand=titanrtx,driver>=525,driver<526 brand=tesla,driver>=535,driver<536 brand=unknown,driver>=535,driver<536 brand=nvidia,driver>=535,driver<536 brand=nvidiartx,driver>=535,driver<536 brand=geforce,driver>=535,driver<536 brand=geforcertx,driver>=535,driver<536 brand=quadro,driver>=535,driver<536 brand=quadrortx,driver>=535,driver<536 brand=titan,driver>=535,driver<536 brand=titanrtx,driver>=535,driver<536
NCCL_VERSION=2.20.5-1
NCCL_NET_GDR_LEVEL=SYS
NVIDIA_DRIVER_CAPABILITIES=compute,utility
NCCL_DEBUG=INFO
NVIDIA_PRODUCT_NAME=CUDA
VLLM_USAGE_SOURCE=production-docker-image
CUDA_VERSION=12.4.0
LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
NCCL_IB_DISABLE=0
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

Ray Worker Environment

INFO 04-22 13:52:29 [__init__.py:239] Automatically detected platform cuda.
Collecting environment information...
PyTorch version: 2.6.0+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 4.0.0
Libc version: glibc-2.35

Python version: 3.12.9 (main, Feb  5 2025, 08:49:00) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-1084-azure-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.4.131
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA A100-SXM4-40GB
GPU 1: NVIDIA A100-SXM4-40GB
GPU 2: NVIDIA A100-SXM4-40GB
GPU 3: NVIDIA A100-SXM4-40GB
GPU 4: NVIDIA A100-SXM4-40GB
GPU 5: NVIDIA A100-SXM4-40GB
GPU 6: NVIDIA A100-SXM4-40GB
GPU 7: NVIDIA A100-SXM4-40GB

Nvidia driver version: 570.124.06
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        48 bits physical, 48 bits virtual
Byte Order:                           Little Endian
CPU(s):                               96
On-line CPU(s) list:                  0-95
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 7V12 64-Core Processor
CPU family:                           23
Model:                                49
Thread(s) per core:                   1
Core(s) per socket:                   48
Socket(s):                            2
Stepping:                             0
BogoMIPS:                             4890.87
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl tsc_reliable nonstop_tsc cpuid extd_apicid aperfmperf tsc_known_freq pni pclmulqdq ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw topoext perfctr_core ssbd vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves clzero xsaveerptr rdpru arat umip rdpid
Hypervisor vendor:                    Microsoft
Virtualization type:                  full
L1d cache:                            3 MiB (96 instances)
L1i cache:                            3 MiB (96 instances)
L2 cache:                             48 MiB (96 instances)
L3 cache:                             384 MiB (24 instances)
NUMA node(s):                         4
NUMA node0 CPU(s):                    0-23
NUMA node1 CPU(s):                    24-47
NUMA node2 CPU(s):                    48-71
NUMA node3 CPU(s):                    72-95
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Mitigation; untrained return thunk; SMT disabled
Vulnerability Spec rstack overflow:   Mitigation; safe RET, no microcode
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Retpolines; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] flashinfer-python==0.2.1.post2+cu124torch2.6
[pip3] numpy==2.1.3
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.4.0
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.51.0
[pip3] triton==3.2.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.8.3
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
	�[4mGPU0	GPU1	GPU2	GPU3	GPU4	GPU5	GPU6	GPU7	NIC0	NIC1	NIC2	NIC3	NIC4	NIC5	NIC6	NIC7	NIC8	CPU Affinity	NUMA Affinity	GPU NUMA ID�[0m
GPU0	 X 	NV12	NV12	NV12	NV12	NV12	NV12	NV12	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	SYS	24-47	1		N/A
GPU1	NV12	 X 	NV12	NV12	NV12	NV12	NV12	NV12	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	SYS	24-47	1		N/A
GPU2	NV12	NV12	 X 	NV12	NV12	NV12	NV12	NV12	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	NODE	0-23	0		N/A
GPU3	NV12	NV12	NV12	 X 	NV12	NV12	NV12	NV12	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	NODE	0-23	0		N/A
GPU4	NV12	NV12	NV12	NV12	 X 	NV12	NV12	NV12	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	72-95	3		N/A
GPU5	NV12	NV12	NV12	NV12	NV12	 X 	NV12	NV12	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	72-95	3		N/A
GPU6	NV12	NV12	NV12	NV12	NV12	NV12	 X 	NV12	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	48-71	2		N/A
GPU7	NV12	NV12	NV12	NV12	NV12	NV12	NV12	 X 	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	48-71	2		N/A
NIC0	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	 X 	NODE	SYS	SYS	SYS	SYS	SYS	SYS	SYS				
NIC1	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	 X 	SYS	SYS	SYS	SYS	SYS	SYS	SYS				
NIC2	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	 X 	NODE	SYS	SYS	SYS	SYS	NODE				
NIC3	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	 X 	SYS	SYS	SYS	SYS	NODE				
NIC4	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	 X 	NODE	SYS	SYS	SYS				
NIC5	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	 X 	SYS	SYS	SYS				
NIC6	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	 X 	NODE	SYS				
NIC7	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	 X 	SYS				
NIC8	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	SYS	SYS	NODE	NODE	SYS	SYS	SYS	SYS	 X 				

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NIC Legend:

  NIC0: mlx5_0
  NIC1: mlx5_1
  NIC2: mlx5_2
  NIC3: mlx5_3
  NIC4: mlx5_4
  NIC5: mlx5_5
  NIC6: mlx5_6
  NIC7: mlx5_7
  NIC8: mlx5_8

NVIDIA_VISIBLE_DEVICES=GPU-68dda3c9-81cb-8b2e-13f8-1e1b0cf85f62,GPU-7becc8c6-c086-1686-1411-31e459854ca1,GPU-ca8b7d4d-66f3-79da-628a-86f381fb8ceb,GPU-6eee3dae-945c-4707-4e6d-e0d7a99ad57d,GPU-8e5bd307-d9b9-981a-9bbb-8bbcc43641f9,GPU-af06d3de-09a8-86d2-71d7-353c1072c111,GPU-81cc8c9b-6fda-5d36-d287-8495bcfa2029,GPU-2af6f3df-1c02-5f55-b5f8-5e03ed6e49ac
NVIDIA_REQUIRE_CUDA=cuda>=12.4 brand=tesla,driver>=470,driver<471 brand=unknown,driver>=470,driver<471 brand=nvidia,driver>=470,driver<471 brand=nvidiartx,driver>=470,driver<471 brand=geforce,driver>=470,driver<471 brand=geforcertx,driver>=470,driver<471 brand=quadro,driver>=470,driver<471 brand=quadrortx,driver>=470,driver<471 brand=titan,driver>=470,driver<471 brand=titanrtx,driver>=470,driver<471 brand=tesla,driver>=525,driver<526 brand=unknown,driver>=525,driver<526 brand=nvidia,driver>=525,driver<526 brand=nvidiartx,driver>=525,driver<526 brand=geforce,driver>=525,driver<526 brand=geforcertx,driver>=525,driver<526 brand=quadro,driver>=525,driver<526 brand=quadrortx,driver>=525,driver<526 brand=titan,driver>=525,driver<526 brand=titanrtx,driver>=525,driver<526 brand=tesla,driver>=535,driver<536 brand=unknown,driver>=535,driver<536 brand=nvidia,driver>=535,driver<536 brand=nvidiartx,driver>=535,driver<536 brand=geforce,driver>=535,driver<536 brand=geforcertx,driver>=535,driver<536 brand=quadro,driver>=535,driver<536 brand=quadrortx,driver>=535,driver<536 brand=titan,driver>=535,driver<536 brand=titanrtx,driver>=535,driver<536
NCCL_VERSION=2.20.5-1
NCCL_NET_GDR_LEVEL=SYS
NVIDIA_DRIVER_CAPABILITIES=compute,utility
NCCL_DEBUG=INFO
NVIDIA_PRODUCT_NAME=CUDA
VLLM_USAGE_SOURCE=production-docker-image
CUDA_VERSION=12.4.0
LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
NCCL_IB_DISABLE=0
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

The pods are deployed with the following configuration:

apiVersion: leaderworkerset.x-k8s.io/v1
kind: LeaderWorkerSet
metadata:
  name: deepseek-v2-5
  namespace: default
spec:
  replicas: 1
  leaderWorkerTemplate:
    size: 2
    restartPolicy: RecreateGroupOnPodRestart
    leaderTemplate:
      metadata:
        labels:
          role: leader
      spec:
        containers:
        - name: vllm
          image: ghcr.io/surajssd/llm-k8s/lws-vllm:0.8.3
          imagePullPolicy: Always
          command:
          - sh
          - -c
          - "set -x;
            bash /vllm-workspace/examples/online_serving/multi-node-serving.sh leader
            --ray_cluster_size=$LWS_GROUP_SIZE
            --dashboard-host=0.0.0.0
            --metrics-export-port=8080;
            python3 -m vllm.entrypoints.openai.api_server
            --port 8000
            --model deepseek-ai/DeepSeek-V2.5
            --tensor-parallel-size 8
            --pipeline-parallel-size 2
            --enable-prefix-caching
            --max-model-len 8192
            --enforce-eager
            --trust-remote-code"
          env:
          - name: NCCL_DEBUG
            value: INFO
          - name: NCCL_NET_GDR_LEVEL
            value: SYS
          - name: NCCL_IB_DISABLE
            value: "0"
          readinessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 60
            periodSeconds: 15
            timeoutSeconds: 720
          ports:
          # VLLM port
          - containerPort: 8000
          # Ray dashboard port
          - containerPort: 8265
          # Ray metrics port
          - containerPort: 8080
          # Ray cluster address
          - containerPort: 6379
          resources:
            limits:
              nvidia.com/gpu: "8"
              rdma/shared_ib: 1
            requests:
              nvidia.com/gpu: "8"
              rdma/shared_ib: 1
          securityContext:
            capabilities:
              add: [ "IPC_LOCK" ]
          volumeMounts:
          - name: shm
            mountPath: /dev/shm
          - name: model
            mountPath: /root/.cache/huggingface
        volumes:
        - name: shm
          emptyDir:
            medium: Memory
        - name: model
          hostPath:
            path: /mnt/model
            type: DirectoryOrCreate
        affinity:
          podAntiAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
            - labelSelector:
                matchExpressions:
                - key: leaderworkerset.sigs.k8s.io/name
                  operator: In
                  values:
                  - deepseek-v2-5
              topologyKey: "kubernetes.io/hostname"

    workerTemplate:
      metadata:
        labels:
          role: worker
      spec:
        containers:
        - name: vllm
          image: ghcr.io/surajssd/llm-k8s/lws-vllm:0.8.3
          imagePullPolicy: Always
          command:
          - sh
          - -c
          - "set -x;
            bash /vllm-workspace/examples/online_serving/multi-node-serving.sh worker
            --ray_address=$LWS_LEADER_ADDRESS
            --metrics-export-port=8080"
          env:
          - name: NCCL_DEBUG
            value: INFO
          - name: NCCL_NET_GDR_LEVEL
            value: SYS
          - name: NCCL_IB_DISABLE
            value: "0"
          ports:
          # VLLM port
          - containerPort: 8000
          # Ray metrics port
          - containerPort: 8080
          resources:
            limits:
              nvidia.com/gpu: "8"
              rdma/shared_ib: 1
            requests:
              nvidia.com/gpu: "8"
              rdma/shared_ib: 1
          securityContext:
            capabilities:
              add: [ "IPC_LOCK" ]
          volumeMounts:
          - name: shm
            mountPath: /dev/shm
          - name: model
            mountPath: /root/.cache/huggingface
        volumes:
        - name: shm
          emptyDir:
            medium: Memory
        - name: model
          hostPath:
            path: /mnt/model
            type: DirectoryOrCreate
        affinity:
          podAntiAffinity:
            requiredDuringSchedulingIgnoredDuringExecution:
            - labelSelector:
                matchExpressions:
                - key: leaderworkerset.sigs.k8s.io/name
                  operator: In
                  values:
                  - deepseek-v2-5
              topologyKey: "kubernetes.io/hostname"

How do I reproduce this issue?

Follow the steps here: https://github.com/surajssd/llm-k8s/blob/fair-benchmarks/deepseek-v2.5/reproduce-benchmark-results.sh

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[Jeffwan]

I will rerun some benchmark and check the key metrics from our side and check whether the trend are consistent.

[surajssd]

Thanks @Jeffwan

[surajssd]

@Jeffwan I tried to do benchmarks with tp=16 (without pp) but ran into another issue: #15100 (comment)

[Jeffwan]

@surajssd finally get some time to reproduce the benchmark. could you share the client settings like prompt length and output size? If there're some reproducible scripts, that would be better.

[surajssd]

@Jeffwan I used leaderworkerset to deploy as mentioned in the issue description. I've also added a LWS YAML config in there, PTAL.

[surajssd]

And for the bechmarks itself, I have this script here: https://github.com/surajssd/llm-k8s/blob/fair-benchmarks/deepseek-v2.5/reproduce-benchmark-results.sh#L33-L62

[Jeffwan]

Setups

Here's my deployment yaml https://gist.github.com/Jeffwan/50e67dcdc35e2e14c71e053b0d07a7c9
I just tune --tensor-parallel-size 8 --pipeline-parallel-size 2 and

              - name: NCCL_IB_DISABLE
                value: "0" # or 1 

to generate different experiment configs

hardware:

• NVIDIA H20-SXM5-96GB * 8. TCP 96 Gbps, RDMA (RoCE v2)400 Gbps * 8
• Engine: AIBrix optimized v0.7.3 (pretty old one, we have a fix not upstreamed yet)

Benchmark

I used vLLM benchmark scripts - https://github.com/vllm-project/vllm/blob/main/benchmarks/benchmark_serving.py

python benchmark_serving.py \
    --backend vllm \
    --model deepseek-ai/DeepSeek-R1 \
    --served-model-name deepseek-r1-671b \
    --base-url http://${ENDPOINT}:80 \
    --endpoint /v1/completions \
    --num-prompts 100 \
    --request-rate 1 \
    --metric_percentiles '50,90,95,99' \
    --goodput ttft:5000 tpot:50 \
    --max-concurrency 200 \
    --random-input-len 2000 \
    --random-output-len 200 \
    --dataset-name random \
    --ignore-eos \
    --seed 61

Note: detail evaluation metrics is here https://gist.github.com/Jeffwan/50e67dcdc35e2e14c71e053b0d07a7c9?permalink_comment_id=5584332#gistcomment-5584332

Conclusion

I can reproduce similar patterns for TP=8, PP2, I didn't spend more time testing different QPS, but from basic quantitative analysis, the result would be similar. I think overall it meets the expectation

1. RDMA provides significant gains for TP=16 case. this is also the only setting can meet some goodput SLO (TTFT 5s, TPOT 50ms)
2. RDMA impact is smaller for TP=8, PP=2.
3. Slight latency overhead from PP configuration: TP=8 PP=2 introduces ~100ms additional TTFT compared to TP=16 RDMA, likely due to pipeline stage synchronization.
4. Recommendation:

• Use RDMA for large-scale TP setups to unlock its full bandwidth and latency advantages.
• TCP is sufficient for small PP-based deployments where communication is minimal.

Network bandwidth monitoring

[chewong]

@Jeffwan, thanks for sharing the benchmark. I ran a similar benchmark to yours (2x8xH100 with IB connectivity). While I got comparable numbers for 8TP+2PP, for TP16, I observed very suboptimal TPOT with or without InfiniBand. I’m not using RoCEv2, and I ran NVIDIA's nccl-tests before launching the workload, with the nccl-tests bandwidth appearing normal. For TP16, I had to turn on --enforce-eager to fix this issue. Any suggestions on what I should investigate to identify the cause of the suboptimal benchmark results?

https://gist.github.com/chewong/9e96fb4095a98c3d4d9b33c0f71f1749