[Mvapich-discuss] [MVAPICH-PLUS] RPM Request and Clarification

[Panda, Dhabaleswar]

Hi,

Thanks for your note. We will be sending you the requested RPM shortly.

Please note that no more active developments are taking place for MVAPICH2-X and MVAPICH2-GDR. You can use the latest MVAPICH-Plus 4.0 release. This version is optimized for fast communication using both CPU and GPU buffers. It provides support for all different interconnects (including InfiniBand, RoCE, Slingshot, and Omni-Path), GPUs (NVIDIA, AMD, and Intel), and CPUs (Arm and x86). We strongly encourage you to move to the MVAPICH-Plus series.

Thanks,

DK Panda


________________________________________
From: Mvapich-discuss <mvapich-discuss-bounces+panda.2=osu.edu at lists.osu.edu> on behalf of Le Viet Duc via Mvapich-discuss <mvapich-discuss at lists.osu.edu>
Sent: Monday, December 9, 2024 2:57 AM
To: mvapich-discuss at lists.osu.edu
Cc: 권민우; 안도식; KISTI 사용자지원
Subject: [Mvapich-discuss] [MVAPICH-PLUS] RPM Request and Clarification

Dear Prof. Panda and MVAPICH Group, I hope this email finds you well. With the recent release of MVAPICH-PLUS providing optimized collectives for Hopper GPUs, KISTI Supercomputing Center (KR) is planning to conduct OMB on H200 GPUs. Since all
ZjQcmQRYFpfptBannerStart
This Message Is From an External Sender
This message came from outside your organization.
<https://us-phishalarm-ewt.proofpoint.com/EWT/v1/KGKeukY!vOQfsSqtA6YgpRdxXw-kGXsy5n0Z2EoOW2S-tEEW2YSnSwcjUrf-bKXgUMAD4dbIIN8fvMgqH9g8xRZ8-QEnv12hV2cfC8wVzaWOPra1cibQc-EO7N1oQ7UnzfAtDaBHiv7JTEd9wsMLAfLzDCl_$>
Report Suspicious

ZjQcmQRYFpfptBannerEnd
Dear Prof. Panda and MVAPICH Group,

I hope this email finds you well.
With the recent release of MVAPICH-PLUS providing optimized collectives for Hopper GPUs, KISTI Supercomputing Center (KR) is planning to conduct OMB on H200 GPUs.

Since all the provided rpm files are based on RHEL 8/9, we have submitted a rpm request through the online submission form last week (Dec 5th).
Please allow me to submit the request again in case the gitlab mailing agent failed to deliver the first one.


  *
Organization:  KISTI (S. Korea)
  *
MVAPICH Version: MVAPICH-PLUS 4
  *
CPU Arch: x86
  *
GPU Arch :Hopper H200 (sm_90)
  *
OS: CentOS 7.9.2009
  *
MOFED: 23.10-0.5.5.0
  *
Compiler: gcc12.2.0
  *
Launcher: slurm
  *
Runtime: Cuda
  *
Runtime Version: CUDA 12.3
  *
Other: Slurm 24.05.3

Could you also clarify if my understanding below is correct ?

  *
MVAPICH2-X: optimized for CPU with SHARP support on CPU buffers.
  *
MVAPICH2-GDR: optimized for heterogenous systems without SHARP support on GPU buffers.
  *
MVAPICH-PLUS: X+GDR with SHARP support on GPU buffers
     *
In other words, is the PLUS version equivalent to NVIDIA's NCCL ?
     *
Does the PLUS version also support  all_reduce through NVLINK4 switches ?

Please pardon my confusion due to current product segmentation.

Regards.
Viet-Duc Le