[lustre-discuss] lctl ping node28 at o2ib report Input/output error
adilger at whamcloud.com
Thu Jun 28 23:06:08 PDT 2018
On Jun 28, 2018, at 21:14, yu sun <sunyu1949 at gmail.com> wrote:
> all server and client that fore-mentioned is using netmasks 255.255.255.224. and they can ping with each other, for example:
> root at ml-gpu-ser200.nmg01:~$ ping node28
> PING node28 (10.82.143.202) 56(84) bytes of data.
> 64 bytes from node28 (10.82.143.202): icmp_seq=1 ttl=61 time=0.047 ms
> 64 bytes from node28 (10.82.143.202): icmp_seq=2 ttl=61 time=0.028 ms
> --- node28 ping statistics ---
> 2 packets transmitted, 2 received, 0% packet loss, time 999ms
> rtt min/avg/max/mdev = 0.028/0.037/0.047/0.011 ms
> root at ml-gpu-ser200.nmg01:~$ lctl ping node28 at o2ib1
> failed to ping 10.82.143.202 at o2ib1: Input/output error
> root at ml-gpu-ser200.nmg01:~$
> and we also have hundreds of GPU machines with different IP Subnet, they are in service and it's difficulty to change the network structure. so any material or document can guide me solve this by don't change network structure.
The regular IP "ping" is being routed by an IP router, but that doesn't
work with IB networks, AFAIK. The IB interfaces need to be on the same
subnet, you need to have an IB interface on each subnet configured on
each subnet (which might get ugly if you have a large number of subnets)
or you need to use LNet routers that are connected to each IB subnet to
do the routing (each subnet would be a separate LNet network, for example
10.82.142.202 at o2ib23 or whatever).
The other option would be to use the IPoIB layer with socklnd (e.g.
10.82.142.202 at tcp) but this would not run as fast as native verbs.
> Mohr Jr, Richard Frank (Rick Mohr) <rmohr at utk.edu> 于2018年6月29日周五 上午3:30写道：
> > On Jun 27, 2018, at 4:44 PM, Mohr Jr, Richard Frank (Rick Mohr) <rmohr at utk.edu> wrote:
> >> On Jun 27, 2018, at 3:12 AM, yu sun <sunyu1949 at gmail.com> wrote:
> >> client:
> >> root at ml-gpu-ser200.nmg01:~$ mount -t lustre node28 at o2ib1:node29 at o2ib1:/project /mnt/lustre_data
> >> mount.lustre: mount node28 at o2ib1:node29 at o2ib1:/project at /mnt/lustre_data failed: Input/output error
> >> Is the MGS running?
> >> root at ml-gpu-ser200.nmg01:~$ lctl ping node28 at o2ib1
> >> failed to ping 10.82.143.202 at o2ib1: Input/output error
> >> root at ml-gpu-ser200.nmg01:~$
> > In your previous email, you said that you could mount lustre on the client ml-gpu-ser200.nmg01. Was that not accurate, or did something change in the meantime?
> (Note: Received out-of-band reply from Yu stating that there was a typo in the previous email, and that client ml-gpu-ser200.nmg01 could not mount lustre. Continuing discussion here so others on list can follow/benefit.)
> For the IPoIB addresses used on your nodes, what are the subnets (and netmasks) that you are using? It looks like servers use 10.82.143.X and clients use 10.82.141.X. If you are using a 255.255.0.0 netmask, you should be fine. But if you are using 255.255.255.0, then you will run into problems. Lustre expects that all nodes on the same lnet network (o2ib1 in your case) will also be on the same IP subnet.
> Have you tried running a regular “ping <IPoIB_address>” command between clients and servers to make sure that part is working?
> Rick Mohr
> Senior HPC System Administrator
> National Institute for Computational Sciences
> lustre-discuss mailing list
> lustre-discuss at lists.lustre.org
Principal Lustre Architect
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 235 bytes
Desc: Message signed with OpenPGP
More information about the lustre-discuss