[lustre-discuss] routed LNET connection between server and virtualized client is failing
Stephane Thiell
sthiell at stanford.edu
Thu Aug 13 21:40:54 PDT 2020
Hi Uwe,
We had a similar problem in the past and our conclusion was that the in-kernel OFED drivers (provided by distribution) likely doesn’t have all the patches required for SR-IOV to work correctly in VMs (when using VFs on IB HBAs and KVM). Mellanox OFED is required at least on the VMs when using SR-IOV, otherwise you get these random network errors.
MOFED 4.9 is a LTS release for ConnectX-3 HBAs and should work great with Lustre 2.12.5 on your VMs. It's definitely worth the effort in the long term in my opinion.
Best,
Stephane
> On Aug 13, 2020, at 5:20 AM, Uwe Sauter <uwe.sauter.de at gmail.com> wrote:
>
> Dear all,
>
> (TL;DR at the bottom)
>
> I have the following situation:
>
>
> +----------------+
> | +--------------+ +-------------------------------------------------+
> | Lustre servers | | | |
> | @ o2ib20 | +---------+--------+ | Virtualization host: |
> | | | | | * Proxmox 6.2, up-to-date |
> +----------------+ | o2ib20 | | ** Debian 10.5 based |
> | 10.148.0.0/16 | | ** Ubuntu based kernel 5.4.44-2-pve |
> | | | * ConnectX-3 (MCX354A-FCBT) |
> +---------+--------+ | ** 15 VFs configured |
> | | ** SR-IOV |
> +----------------------+ | * OFED provided by distribution |
> | | |
> +--------+-------+ | |
> | LNET router | | Virtual machines and LNET routers: |
> +--------+-------+ | * CentOS 7.8 based |
> | | * OFED provided by CentOS |
> +----------------------+ | * Lustre 2.12.5 |
> | | * Kernel 3.10.0-1127.18.2.el7 |
> +---------+--------+ | |
> | |-----------------+--------------+--------------+ |
> | o2ib43 | | | | | |
> +----------------+ | 10.225.0.0/16 | | | | | |
> | | | | | +------+-----+ +------+-----+ +------+-----+ |
> | Lustre servers | +---------+--------+ | | | | | | | |
> | @ o2ib43 | | | | VM 1 | | VM 2 | | VM 3 | |
> | +--------------+ | | @ o2ib43 | | @ o2ib43 | | @ o2ib43 | |
> +----------------+ | | | | | | | |
> | | | | | | | |
> | +------------+ +------------+ +------------+ |
> | |
> +-------------------------------------------------+
>
>
> Lustre @ o2ib20 is a Sonexion appliance based on CentOS 7.2 and Lustre version 2.11.0.300_cray_43_gd35e657_dirty.
>
> Lustre @ 02ib43 is CentOS 7.6 based setup with kernel 3.10.0-957.1.3.el7_lustre and Lustre version lustre-2.10.7.1nec-1.el7.x86_64.
>
> The issue I currently see is that once more that one VM is running on the virualization host
> then access to the Lustre file system behind the LNET routers is stuck.
>
> The errors I can see on the VM is e.g.:
>
> [ 1297.470192] LustreError: 2477:0:(events.c:200:client_bulk_callback()) event type 1, status -5, desc ffff89365cebb800
> [ 1297.472058] LustreError: 2478:0:(events.c:200:client_bulk_callback()) event type 1, status -5, desc ffff89365cebb800
> [ 1297.473909] Lustre: 2490:0:(client.c:2133:ptlrpc_expire_one_request()) @@@ Request sent has failed due to network error: [sent
> 1597316593/real 1597316593] req at ffff89365cee0900 x1674906532108800/t0
> (0) o4->snx11167-OST001a-osc-ffff893468b2f800 at 10.148.240.33@o2ib20:6/4 lens 488/448 e 0 to 1 dl 1597316688 ref 2 fl
> Rpc:eX/0/ffffffff rc 0/-1
> [ 1297.479055] Lustre: snx11167-OST001a-osc-ffff893468b2f800: Connection to snx11167-OST001a (at 10.148.240.33 at o2ib20) was lost;
> in progress operations using this service will wait for recovery to com
> plete
> [ 1299.470205] LustreError: 2478:0:(events.c:200:client_bulk_callback()) event type 1, status -5, desc ffff89365cebb800
> [ 1299.472403] LustreError: 2477:0:(events.c:200:client_bulk_callback()) event type 1, status -5, desc ffff89365cebb800
> [ 1299.474395] Lustre: 2490:0:(client.c:2133:ptlrpc_expire_one_request()) @@@ Request sent has failed due to network error: [sent
> 1597316595/real 1597316595] req at ffff89365cee0900 x1674906532108800/t0
> (0) o4->snx11167-OST001a-osc-ffff893468b2f800 at 10.148.240.33@o2ib20:6/4 lens 488/448 e 0 to 1 dl 1597316690 ref 2 fl
> Rpc:eX/2/ffffffff rc 0/-1
> [ 1299.479830] Lustre: snx11167-OST001a-osc-ffff893468b2f800: Connection to snx11167-OST001a (at 10.148.240.33 at o2ib20) was lost;
> in progress operations using this service will wait for recovery to com
> plete
> [ 1299.496826] Lustre: snx11167-OST001a-osc-ffff893468b2f800: Connection restored to 10.148.240.33 at o2ib20 (at 10.148.240.33 at o2ib20)
> [ 1301.470102] LustreError: 2478:0:(events.c:200:client_bulk_callback()) event type 1, status -5, desc ffff89365cebb800
> [ 1301.472096] LustreError: 2477:0:(events.c:200:client_bulk_callback()) event type 1, status -5, desc ffff89365cebb800
> [ 1301.474135] Lustre: 2490:0:(client.c:2133:ptlrpc_expire_one_request()) @@@ Request sent has failed due to network error: [sent
> 1597316597/real 1597316597] req at ffff89365cee0900 x1674906532108800/t0
> (0) o4->snx11167-OST001a-osc-ffff893468b2f800 at 10.148.240.33@o2ib20:6/4 lens 488/448 e 0 to 1 dl 1597316692 ref 2 fl
> Rpc:eX/2/ffffffff rc 0/-1
> [ 1301.479772] Lustre: snx11167-OST001a-osc-ffff893468b2f800: Connection to snx11167-OST001a (at 10.148.240.33 at o2ib20) was lost;
> in progress operations using this service will wait for recovery to com
> plete
> [ 1301.483576] LNetError: 2486:0:(lib-move.c:1999:lnet_handle_find_routed_path()) no route to 10.148.240.33 at o2ib20 from <?>
>
>
>
> Access to the Lustre file system which is on the same IB fabric is still possible so I suspect that this is somehow related to
> LNET routing.
>
> If I run LNET selftests as explained at http://wiki.lustre.org/LNET_Selftest between one of the LNET routers and the VMs I can see
> that RPCs get dropped.
>
>
> Access from a client running on native hardware is possible for both file systems.
>
>
>
> Has someone a comparable setup? What kind of logs is needed to debug this? I'll gladly provide any info…
>
>
> TL;DR
> * native access is possible in the same IB fabric as well as when being routed between different fabrics
> * if only one VM is running then access is possible to both file systems, too
> * if more VMs are running on the same virtualization host than access is only possible on the file system attached to the same
> fabric as the VMs
> * access to the routed file system gets stuck
>
>
>
> Any help is appreciated.
>
> Thanks,
>
> Uwe Sauter
>
> _______________________________________________
> lustre-discuss mailing list
> lustre-discuss at lists.lustre.org
> http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
More information about the lustre-discuss
mailing list