Author Archives: eremo

NUMA node error for Nvidia cards on Linux PCs

Posted on by

You may have experienced it in various contexts: CUDA, Tensorflow, gaming applications or complex 3D graphics applications may warn you that your Nvidia card is associated with an unexpected negative NUMA value. The warning often refers to a value of “-1”. And the clever application replaces this value by a default value of “0”.

The problem is particularly annoying when dealing Machine Learning, e.g. in Jupyter notebooks. There warnings may repeatedly clatter the output of some cells – e.g. during the setup of the graphics card for some ML experiments.

Besides the question why the Nvidia drivers for Linux and/or CUDA drivers do not fix this problem by detecting just one NUMA node on the system and setting the value for the card to “0”, the question for us users is how we can get rid of the warnings.

A basic idea is that we set the right value by ourselves. I have described this simple measure in the sister blog, which unfortunately still is under construction. See:
Setting NUMA node to 0 for Nvidia cards on standard Linux PCs.

There I also briefly discuss what NUMA basically is thought for – and why it normally does not affect consumer PCs.

 

More fun with veth and Linux network namespaces – VII – two namespaces connected by a veth based VLAN

Posted on by

With the last one of the previous posts in this series

we started to look at VLANs based on veths and Linux bridges. I presented multiple options to configure a veth endpoint in a network namespace for enabling communication with other namespaces via tagged Ethernet packets.

We saw that we can potentially start splitting network traffic already from within a multi-homed namespace which is connected to different VLAN-segments. However, a veth endpoint with multiple interfaces also poses a basic ambiguity problem for the direction of both ARP and ICMP requests into the right one of the various attached (V)LAN-segments. From the results of previous experiments we would assume that the Linux kernel solves this problem by following routes.

In this post we will study the most simple VLAN configuration one can think of: We connect two network namespaces directly with a veth based VLAN line. I.e. we use a veth connection to transmit tagged packets along it. As this would be a lit boring, we add some pepper to the scenario by allowing for untagged packets, too.

To achieve this we reduce “option 4” discussed in the last post to a one-armed solution: We allow for only one VLAN interface per veth endpoint (see a sketch of the scenario below). We start a series of experiments with assigning an IP to the veth’s main interface, only. The setting for the source validation kernel parameter rp_filter will be relaxed.

In the experiments of this post we focus on

  • some claimed aspects regarding the role of the main trunk interface of a veth endpoint,
  • the potential impact of routes on the ARP communication.
  • the fact that working ARP does not mean working ICMP for a variety of different reasons.

Regarding the first point we will see that tagged packets just traverse the trunk interfaces on their way from and to the VLAN interface. Regarding route settings we will look at 36 possible variants. We will see that under the given conditions asymmetric route settings may or may not disable communication – already on the ARP level. But even if ARP works and even if we had symmetric routes in the namespaces, ICMP may not function. I will try to isolate the causes of positive and negative results for ARP and ICMP requests.

I recommend that readers who want to perform these experiments on their own to watch the system log in parallel.

Continue reading →

Leap/SLES 15.5 – strange compatibility problem between tcpdump, libpcap and arping from iputils

Posted on by

My readers know that I presently work again with virtual networks. A part of my studies is related to ARP and routes on veth devices with VLAN-interfaces. I follow the packet transfer across VLANs with tcpdump, which itself depends on libpcap. On Leap 15 the relevant package is: libpcap1. ARP commands were generated by the arping command, which gets available after an installation of the package “iputils“.

This worked perfectly on a laptop. I know for a presentation had to use another system with the same kernelversion and virtual networking support. There I got strange messages for ARP packets passing a veth endpoint’s main device (in my example: veth2V) on their way to a VLAN interface (“veth2V.30) on the same veth endpoint: 

netns2:~ # tcpdump -n -e -i any -v
tcpdump: data link type LINUX_SLL2
tcpdump: listening on any, link-type LINUX_SLL2 (Linux cooked v2), snapshot length 262144 bytes

15:15:36.334692 veth2V B   ifindex 2 46:b9:81:00:00:1e ethertype ARP (0x0806), length 52: Unknown Hardware (36461) (len 0), Unknown Protocol (0x0000) (len 1), Unknown (2048) 
        0x0000:  8e6d 0000 0001 0800 0604 0001 46b9 8b4b  .m..........F..K
        0x0010:  8e6d c0a8 0501 ffff ffff ffff c0a8 0502  .m..............
15:15:36.334692 veth2V.30 B   ifindex 4 46:b9:8b:4b:8e:6d ethertype ARP (0x0806), length 48: Ethernet (len 6), IPv4 (len 4), Request who-has 192.168.5.2 (ff:ff:ff:ff:ff:ff) tell 192.168.5.1, length 28
15:15:36.334720 veth2V.30 Out ifindex 4 d2:85:39:c8:43:fc ethertype ARP (0x0806), length 48: Ethernet (len 6), IPv4 (len 4), Reply 192.168.5.2 is-at d2:85:39:c8:43:fc, length 28
15:15:36.334721 veth2V Out ifindex 2 d2:85:81:00:00:1e ethertype ARP (0x0806), length 52: Unknown Hardware (17404) (len 0), Unknown Protocol (0x0000) (len 1), Unknown (2048) 
        0x0000:  43fc 0000 0001 0800 0604 0002 d285 39c8  C.............9.
        0x0010:  43fc c0a8 0502 46b9 8b4b 8e6d c0a8 0501  C.....F..K.m....

I had never seen similar messages in comparable experiments with veths before. And these messages about “Unknown Hardware”. In addition the length of the Ethernet packets were wrong. I did not get such errors on my laptop where I had prepared the setups of the virtual VLANs.

It took me some time to find the difference between the systems: iputils as well as tcpdump on both systems came from the Network:Utilites-repository
“https://download.opensuse.org/ repositories/ network: /utilities/ 15.5/“.

However, libpcap1 on the presentation system came from the main SLES 15.5 OSS repository in version 1.10.1-150400.1.7. On my laptop I had instead fetched the library from the Network:Utilites-repository, too.

Changing libpcap1 to the present version 1.10.4-lp155.92.1 from the Network:Utilites-repository led to correct tcpdump information: 

netns2:~ # tcpdump -n -e -i any -v
tcpdump: data link type LINUX_SLL2
tcpdump: listening on any, link-type LINUX_SLL2 (Linux cooked v2), snapshot length 262144 bytes

15:38:44.899645 veth2V B   ifindex 2 f2:5b:23:ba:16:8a ethertype ARP (0x0806), length 48: Ethernet (len 6), IPv4 (len 4), Request who-has 192.168.5.2 (ff:ff:ff:ff:ff:ff) tell 192.168.5.1, length 28
15:38:44.899645 veth2V.30 B   ifindex 4 f2:5b:23:ba:16:8a ethertype ARP (0x0806), length 48: Ethernet (len 6), IPv4 (len 4), Request who-has 192.168.5.2 (ff:ff:ff:ff:ff:ff) tell 192.168.5.1, length 28
15:38:44.899667 veth2V.30 Out ifindex 4 9a:88:24:0c:f9:99 ethertype ARP (0x0806), length 48: Ethernet (len 6), IPv4 (len 4), Reply 192.168.5.2 is-at 9a:88:24:0c:f9:99, length 28
15:38:44.899669 veth2V Out ifindex 2 9a:88:24:0c:f9:99 ethertype ARP (0x0806), length 48: Ethernet (len 6), IPv4 (len 4), Reply 192.168.5.2 is-at 9a:88:24:0c:f9:99, length 28

Interestingly, even if one changed both tcpdump, iputils and libpcap1 to the main Leap /SLES 15.5 repository the problem would come up, too.

So, there seems to be something severely wrong with libpcap1 of the main Leap /SLES 15.5 repositories.