Hello Kurisaki,
The REMD code was designed to execute with one replica per GPU. As the number of replicas has to necessarily be even, the only way to run REMD with 3 GPUs is if you are running more than one replica per GPU. This is very bad for computational efficiency, and can lead to problems like the one you are showing.
Just to make sure this is the only cause of your error: try to run the same calculation reduced to only 2 replicas (with 2 GPUs) and see if your error message still shows up.
I hope this helps,
All the best,
Vinícius Wilian D. Cruzeiro
PhD Candidate
Department of Chemistry, Physical Chemistry Division
University of Florida, United States
Voice: +1(352)846-1633
________________________________
From: Ikuo KURISAKI <kurisaki.bear.kobe-u.ac.jp>
Sent: Wednesday, March 6, 2019 3:24 AM
To: amber.ambermd.org
Subject: [AMBER] Error in REMD with pmemd.cuda.MPI
Dear Ambers,
Thank you for usual support.
I’m now facing with the following error message, that is,
“Error: out of memory launching kernel kCalculateNMREnergy”.
This appeared when I tried REMD with pmemd.cuda.MPI with three GPU cards.
I first doubted memory shortage of GPU cards,
although GPU memories are still available (2800/11000 MiB for each of three
GPU cards)
I’m happy If you give some advices to condition where this message is
returned.
Best regards,
Kurisaki
===========================================
栗崎以久男/ Ikuo KURISAKI
神戸大学大学院システム情報学研究科
計算科学専攻 特命講師/
Graduate School of System Informatics, Kobe University
Department of Computational Science Project-Oriented Lecturer
E-mail: kurisaki.bear.kobe-u.ac.jp
===========================================
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.ambermd.org_mailman_listinfo_amber&d=DwICJg&c=sJ6xIWYx-zLMB3EPkvcnVg&r=vg8iTdivJL1PEwCvcH8P2DFF-Rtc9lAQvIqnaSWm1Pc&m=rW9ETb64Kh6XZ8NLi-5KX8bFIp_J3cdXcw50OeaEpYs&s=TFaHYw95Ky-0wIbBQc5fDZx1Cu7rxKlK8rcfJuTlMgQ&e=
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Mar 06 2019 - 07:00:02 PST
