EFA not working with CentOS 7 and Intel MPI

Question

I have been trying to resolve this problem all week.  
  
I followed all the steps to create a 2 instance system with EFA drivers installed.  
https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/efa-start.html  
  
I started with the official CentOS 7 HVM image in the marketplace and updated it before installing the efa driver.  
  
I have separately and on different images, tried both IntelMPI 2019.5 and 2019.6 (I installed them with the provided aws_impi.sh that points to the amazon efa driver library)  
  
Any test I run hangs with efa enabled, the jobs start on both instances but MPI communication is not occurring.  
export FI_PROVIDER=efa  
  
Things run fine using the sockets fabric though.  
export FI_PROVIDER=sockets  
  
c5n.18xlarge instances.  
  
I am using an updated CentOS7 image with Kernel revision:  
3.10.0-1062.4.1.el7.x86_64  
(I was unable to get it working on the previous kernel version 957 also)  
  
I'm running in a subnet in a VPC with an attached EFS volume.   
efa driver version 1.4.1  
  
I did not install or test with openmpi.  
  
The efa_test.sh fails, but it works if I use "-p sockets" instead of "-p efa".  
  
Another test example:  
#!/bin/bash  
module load mpi  
export FI_PROVIDER=efa  
#export FI_PROVIDER=sockets  
mpirun -np 2 -ppn 1 -f $PWD/hosts $PWD/mpi/pt2pt/osu_latency  
  
Additional details:  
fi_info:  
Instance 1:  
provider: efa  
    fabric: EFA-fe80::a0:1eff:febf:9407  
    domain: efa_0-rdm  
    version: 2.0  
    type: FI_EP_RDM  
    protocol: FI_PROTO_EFA  
provider: efa  
    fabric: EFA-fe80::a0:1eff:febf:9407  
    domain: efa_0-dgrm  
    version: 2.0  
    type: FI_EP_DGRAM  
    protocol: FI_PROTO_EFA  
provider: efa;ofi_rxd  
    fabric: EFA-fe80::a0:1eff:febf:9407  
    domain: efa_0-dgrm  
    version: 1.0  
    type: FI_EP_RDM  
    protocol: FI_PROTO_RXD  
  
Instance2:  
provider: efa  
    fabric: EFA-fe80::70:3bff:fee1:52ad  
    domain: efa_0-rdm  
    version: 2.0  
    type: FI_EP_RDM  
    protocol: FI_PROTO_EFA  
provider: efa  
    fabric: EFA-fe80::70:3bff:fee1:52ad  
    domain: efa_0-dgrm  
    version: 2.0  
    type: FI_EP_DGRAM  
    protocol: FI_PROTO_EFA  
provider: efa;ofi_rxd  
    fabric: EFA-fe80::70:3bff:fee1:52ad  
    domain: efa_0-dgrm  
    version: 1.0  
    type: FI_EP_RDM  
    protocol: FI_PROTO_RXD  
  
ibstat  
CA 'efa_0'  
	CA type:   
	Number of ports: 1  
	Firmware version:   
	Hardware version:   
	Node GUID: 0x0000000000000000  
	System image GUID: 0x0000000000000000  
	Port 1:  
		State: Active  
		Physical state: LinkUp  
		Rate: 100  
		Base lid: 0  
		LMC: 1  
		SM lid: 0  
		Capability mask: 0x00000000  
		Port GUID: 0x00a01efffebf9407  
		Link layer: Unknown  
  
CA 'efa_0'  
	CA type:   
	Number of ports: 1  
	Firmware version:   
	Hardware version:   
	Node GUID: 0x0000000000000000  
	System image GUID: 0x0000000000000000  
	Port 1:  
		State: Active  
		Physical state: LinkUp  
		Rate: 100  
		Base lid: 0  
		LMC: 1  
		SM lid: 0  
		Capability mask: 0x00000000  
		Port GUID: 0x00703bfffee152ad  
		Link layer: Unknown  
  
Some other errors I have seen but I'm not familiar with:  
  
ibsysstat -G 0x00703bfffee152ad  
ibwarn: \[24340] mad_rpc_open_port: can't open UMAD port ((null):0)  
ibsysstat: iberror: failed: Failed to open '(null)' port '0'  
  
ibhosts   
ibwarn: \[11769] mad_rpc_open_port: can't open UMAD port ((null):0)  
/codebuild/output/src613026668/src/rdmacore_build/BUILD/rdma-core-25.0/libibnetdisc/ibnetdisc.c:802; can't open MAD port ((null):0)  
/usr/sbin/ibnetdiscover: iberror: failed: discover failed  
  
Any ideas?  
  
Thanks - Patrick

Answer

Hi Raghu,  
  
Thank you. Adding the outbound rule resolved the issue.  
  
- Patrick

Answer

Hi Patrick - I suspect the problem lies with your Security Group rules. Like mentioned in this document:  
https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/efa-start.html#efa-start-security   
  
EFA requires a security group that allows inbound and outbound traffic to itself. In other words, you need an explicit inbound _and_ outbound rule to allow traffic to **sg-xxxx**. A rule to allow All Traffic to **0.0.0.0** does not suffice.  
  
Let us know if this resolves your issue.

Answer

Hi Raghu,  
I am facing the same problem, but my security group configuration already added the rule for inbound and outbound from itself. Would you mind give some other suggestions I can have a try?  
  
BTW, an additional question is about the Capacity of the EFA, is that ok to have 0x0 in "Capability mask" field?   
  
issue:   
  
$ ibping -S  
ibwarn: \[20355] mad_rpc_open_port: can't open UMAD port ((null):0)  
ibping: iberror: failed: Failed to open '(null)' port '0'  
  
Edited by: zarzen on Mar 2, 2020 11:30 AM

EFA not working with CentOS 7 and Intel MPI

Relevanter Inhalt