From f5d126570f236cb7d14003db2e165c965c5525d3 Mon Sep 17 00:00:00 2001
From: Ali Omar Abdelazim Mohammed
<mohammal@hotlum-login.head.cm.hpcrb.rdlabs.ext.hpe.com>
Date: Tue, 23 Aug 2022 07:04:44 -0500
Subject: [PATCH] add clarifications on FI_OFI_RXM_USE_SRX and
FI_VERBS_PREFER_XRC for MPI apps with >=64 ranks
---
maestro/core.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/maestro/core.c b/maestro/core.c
index c22908f9..a40d60fa 100644
--- a/maestro/core.c
+++ b/maestro/core.c
@@ -610,6 +610,7 @@ mstro_core_init(const char *workflow_name,
ERR("If your job terminates after this error please check\n"
" ... that firewalling is not prohibiting libfabric communication\n"
" ... that ulimit -n is high enough to keep a couple of FDs per network endpoints open (ulimit -n 1024 is a good start)\n"
+ " ... that FI_OFI_RXM_USE_SRX=1 and FI_VERBS_PREFER_XRC=0 with large MPI applications (>=64 ranks) when using verbs provider, especially with Cray MPICH\n"
" ... that your workflow-name/component-name/component-index triple is correct\n");
goto BAILOUT;
}
--
GitLab