Leveraging open technologies to monitor packet drops in AI cluster fabrics

In this talk from the recent OCP Global Summit, Aldrin Isaac, eBay, describes the challenge, AI clusters operate most efficiently over lossless networks for optimum job completion times which can be significantly impacted by dropped packets. Although networks can be designed to minimize packet loss by choosing the right network topology, optimizing network devices and protocols, an effective monitoring and troubleshooting network performance tool is still required. Such tool should capture packet drops, raise notifications and identify various drop reasons and pin point where the drops caused congestions. In turn, it allows the governing management application to tune configurations of relevant infrastructure components, including switches, NICs and GPU servers.

The talk shares the results and best practices of a TAM (Telemetry and Monitoring) solution being prepared for deployment at eBay. It leverages OCP’s SAI and open sFlow drop notification technologies as part of eBay’s ongoing initiatives to adopt open networking hardware and community SONiC for its data centers.

The sFlow Dropped Packet Notification Structures extension mentioned in the talk adds real-time packet drop notifications (including dropped packet header and drop reason) as part of an industry standard sFlow telemetry feed, making the data available to open source and commercial sFlow analytics tools.

For example, Dropped packet metrics with Prometheus and Grafana describes how to incorporate sFlow dropped packet notifications into operational dashboards using current implementations by Arista Networks, VyOS, FD.io / VPP, and Linux servers. Current network hardware is capable of reporting on dropped packets, so ask your network equipment vendor about their plains to support the sFlow extension so that you can befit from this transformational capability.

OCP Global Summit 2024

AI networking is a popular topic at the up coming OCP Global Summit in San Jose, California, with an entire morning on Wednesday October 16 devoted to the subject.

Of particular interest is the talk, Leveraging open technologies to monitor packet drops in AI cluster fabrics, by Aldrin Isaac, eBay, describing the challenge, AI clusters operate most efficiently over lossless networks for optimum job completion times which can be significantly impacted by dropped packets. Although networks can be designed to minimize packet loss by choosing the right network topology, optimizing network devices and protocols, an effective monitoring and troubleshooting network performance tool is still required. Such tool should capture packet drops, raise notifications and identify various drop reasons and pin point where the drops caused congestions. In turn, it allows the governing management application to tune configurations of relevant infrastructure components, including switches, NICs and GPU servers.

The talk will share the results and best practices of a TAM (Telemetry and Monitoring) solution being prepared for deployment at eBay. It leverages OCP’s SAI and open sFlow drop notification technologies as part of eBay’s ongoing initiatives to adopt open networking hardware and community SONiC for its data centers.

The sFlow Dropped Packet Notification Structures extension mentioned in the talk adds real-time packet drop notifications (including dropped packet header and drop reason) as part of an industry standard sFlow telemetry feed, making the data available to open source and commercial sFlow analytics tools.

For example, Dropped packet metrics with Prometheus and Grafana describes how to incorporate sFlow dropped packet notifications into operational dashboards using current implementations for Arista, VyOS, and Linux servers. The availability of drop monitoring in SONiC will extend this capability to the wide range of hardware platforms supporting the SONiC network operating system.

Vector Packet Processor (VPP)

VPP with sFlow - Part 1 and VPP with sFlow - Part 2 describe the journey to add industry standard sFlow instrumentation to the Vector Packet Processor (VPP) an Open Source Terabit Software Dataplane for software routers running on commodity x86 / ARM hardware.

The main conclusions based on testing described in the two VPP blog posts are:

1. If sFlow is not enabled on a given interface, there is no regression on other interfaces.
2. If sFlow is enabled, copying packets costs 11 CPU cycles on average
3. If sFlow takes a sample, it takes only marginally more CPU time to enqueue.
• No sampling gets 9.88Mpps of IPv4 and 14.3Mpps of L2XC throughput,
• 1:1000 sampling reduces to 9.77Mpps of L3 and 14.05Mpps of L2XC throughput,
• and an overly harsh 1:100 reduces to 9.69Mpps and 13.97Mpps only.

The VPP sFlow plugin provides a lightweight method of exporting real-time sFlow telemetry from a VPP based router. Including the plugin with VPP distributions has no impact on performance. Enabling the plugin provides real-time visibility that opens up additional use cases for VPPs programmable dataplane. For example, VPP is well suited to packet filtering use cases where the number of ACL entries would exceed the capabilities of an ASIC. Combined with real-time visibility to identify DDoS attacks, VPP provides an effective means of mitigating the attacks by scrubbing the unwanted traffic.

Emulating congestion with Containerlab

The Containerlab dashboard above shows variation in throughput in a leaf and spine network due to large "Elephant" flow collisions in an emulated network, see Leaf and spine traffic engineering using segment routing and SDN for a demonstration of the issue using physical switches.

This article describes the steps needed to emulate realistic network performance problems using Containerlab. First, using the FRRouting (FRR) open source router to build the topology provides a lightweight, high performance, routing implementation that can be used to efficiently emulate large numbers of routers using the native Linux dataplane for packet forwarding. Second, the containerlab tools netem set command can be used to introduce packet loss, delay, jitter, or restrict bandwidth of ports.

The netem tool makes use of the Linux tc (traffic control) module. Unfortunately, if you are using Docker desktop, the minimal virtual machine used to run containers does not include the tc module.

multipass launch docker

Instead, use Multipass as a convenient way to create and start an Ubuntu virtual machine with Docker support on your laptop. If you are already on a Linux system with Docker installed, skip forward to the git clone step.

multipass ls

List the multipass virtual machines.

Name                    State             IPv4             Image
docker                  Running           192.168.65.3     Ubuntu 22.04 LTS
                                          172.17.0.1

Make a note of the IP address(es) of the docker virtual machine.

multipass shell docker

Run a shell inside the docker virtual machine.

git clone https://github.com/sflow-rt/containerlab.git

Install sflow-rt/containerlab project.

cd containerlab
./run-clab

Run Containerlab using Docker.

In this example we will be using the 3 Stage Close Topology shown above.

env SAMPLING=10 containerlab deploy -t clos3.yml

Start a leaf and spine topology emulation, but use a sampling rate of 1-in-10 rather than the default of 1-in-1000. See Large flow detection for a discussion of scaling sampling rates with link speed to get consistent results between the emulation and a physical network.

./bw.py clab-clos3

Rate limit the links in the topology to 10Mbps.

./topo.py clab-clos3

Post the topology to the sFlow-RT real-time analytics container. Access the Containerlab Dashboard shown at the top of this page using a web browser to connect to http://192.168.65.3:8008/ (where 192.168.65.3 is the IP address of the docker container noted earlier).

docker exec -it clab-clos3-h1 iperf3 -c 172.16.2.2 --parallel 2

Run a series of iperf3 tests to create pairs of large flows between h1 and h2. When the flows take different paths across the fabric the total available bandwidth is 20mbps. If the flows hash onto the same path, then they share 10mbps bandwidth and the throughput is halved.

This example demonstrates that Containerlab is not restricted to emulating and validating configurations, but can also be used to emulate performance issues. In this example, the effects of large flow collisions are relevant to the performance of data center fabrics handling AI/ML workloads where large flow collisions can significantly limit performance, see RoCE networks for distributed AI training at scale.

Dropped packet metrics with Prometheus and Grafana

Dropped packets due to black hole routes, buffer exhaustion, expired TTLs, MTU mismatches, etc. can result in insidious connection failures that are time consuming and difficult to diagnose. Dropped packet notifications with Arista Networks, VyOS dropped packet notifications and Using sFlow to monitor dropped packets describe implementations of the sFlow Dropped Packet Notification Structures extension for Arista Networks switches, VyOS routers, and Linux servers respectively, providing end to end visibility into packet drop events (including switch port, drop reason and packet header for each dropped packet).

Flow metrics with Prometheus and Grafana describes how define flow metrics and create dashboards to trend the flow metrics over time. This article describes how the same setup can be used to define and trend metrics based on dropped packet notifications.

  - job_name: sflow-rt-drops
    metrics_path: /app/prometheus/scripts/export.js/flows/ALL/txt
    static_configs:
      - targets: ['sflow-rt:8008']
    params:
      metric: ['dropped_packets']
      key:
        - 'node:inputifindex'
        - 'ifname:inputifindex'
        - 'reason'
        - 'stack'
        - 'macsource'
        - 'macdestination'
        - 'null:vlan:untagged'
        - 'null:[or:ipsource:ip6source]:none'
        - 'null:[or:ipdestination:ip6destination]:none'
        - 'null:[or:icmptype:icmp6type:ipprotocol:ip6nexthdr]:none'
      label:
        - 'switch'
        - 'port'
        - 'reason'
        - 'stack'
        - 'macsource'
        - 'macdestination'
        - 'vlan'
        - 'src'
        - 'dst'
        - 'protocol'
      value: ['frames']
      dropped: ['true']
      maxFlows: ['20']
      minValue: ['0.001']

The Prometheus scrape configuration above is used to keep track of drop notifications. The highlighed dropped setting is used to select drop notifications for the metric (the default dropped:['false'] creates flow metrics based packet samples and is used to trend normal traffic).

Deploy real-time network dashboards using Docker compose is the simplest way to deploy an sFlow-RT, Prometheus, and Grafana stack with some basic dashboards. Install sFlow-RT Dropped Packets dashboard, code 21721, in Grafana to see the dashboard shown at the top of this page, displaying Drop Locations, Drop Reasons and Dropped Packet Details.

Dropped packet notifications with Arista Networks

Visibility into dropped packets is essential for Artificial Intelligence/Machine Learning (AI/ML) workloads, where a single dropped packet can stall large scale computational tasks, idling millions of dollars worth of GPU/CPU resources, and delaying the completion of business critical workloads. Enabling real-time sFlow telemetry provides the observability into traffic flows and packet drops needed to effectively manage these networks.

The availability of the Arista EOS 4.31.4M maintenance release brings sFlow dropped packet monitoring (previously demonstrated using the 4.30.1F feature release - see SC23 Dropped packet visibility demonstration) to production networks, see EOS Life Cycle Policy

sflow sampling 50000
sflow polling-interval 20
sflow vrf mgmt destination 203.0.113.100
sflow vrf mgmt source-interface Management0
sflow run

The above Arista EOS commands enable sFlow counter polling and packet sampling on all ports, sending the sFlow telemetry to the sFlow analyzer at 203.0.113.100

flow tracking mirror-on-drop
  sample limit 100 pps
  !
  tracker SFLOW
    exporter SFLOW
      format sflow
      collector sflow
      local interface Management0
  no shutdown

The above commands add sFlow Dropped Packet Notification Structures to the sFlow telemetry feed using Broadcom Mirror on Drop (MoD) instrumentation. Broadcom implements mirror-on-drop in Jericho 2, Trident 3, and Tomahawk 3, or later ASICs.

In this example, the sFlow-RT real-time analytics engine receives sFlow telemetry from switches/routers and creates metrics to drive the real-time Grafana dashboard shown at the top of the article. Deploy real-time network dashboards using Docker compose describes how to quickly deploy a monitoring stack consisting of sFlow-RT, a Prometheus time series database, and Grafana dashboards.

VyOS 1.4 LTS released

Protectli Vault - 4 Port

The VyOS 1.4.0 (Sagitta) LTS release announcement is exciting news! VyOS is an open source router operating system based on Linux that can be installed on commodity PC hardware - for optimal performance at least 1GB RAM and 4GB of storage space is recommended.

The new 1.4 LTS release includes a significantly enhanced implementation of industry standard sFlow telemetry based on the open source Host sFlow agent.

set system sflow interface eth0
set system sflow interface eth1
set system sflow interface eth2
set system sflow interface eth3
set system sflow polling 30
set system sflow sampling-rate 1000
set system sflow drop-monitor-limit 50
set system sflow server 192.0.2.100

Enter the commands above to enable sFlow monitoring on interfaces eth0, eth1, eth2, and eth3. Interface counters will be exported every 30 seconds, packets will be sampled with probability 1/1000, and up to 50 packet headers (and drop reasons) per second will collected from packets dropped by the router. The sFlow telemetry stream will be sent to an sFlow collector at 192.0.2.100.

Running Docker on the sFlow collector makes it easy to run a variety of sFlow analytics tools.

docker run --rm -p 6343:6343/udp sflow/sflowtool

Run the sflow/sflowtool image to decode and print the contents of the sFlow telemetry stream and verify receipt of data.

docker run --rm -p 6343:6343/udp sflow/tcpdump tcp port 80

Run the sflow/tcpdump image to decode and filter sampled packet headers. For more complex packet analysis tasks, try the sflow/tshark image.

Run the sflow/sflowtrend image to trend interface counters and top flows.

Deploy real-time network dashboards using Docker compose describes how to configure Prometheus and Grafana to capture time series data and create custom dashboards.

Dropped packet reason codes in VyOS describes how the new Linux kernel in VyOS 1.4 provides detailed visibility into every dropped packet (including the reason it was dropped). This cabability is used by the new sFlow agent implement the sFlow Dropped Packet Notification Structures extension to provide network-wide visibility into dropped packets.

Download VyOS today to try out the new features. Pre-built LTS images are available with paid support, but anyone can build an image from sources or download the latest rolling release.