netobserv

Network Observability Real-Time Per Flow Packets Drop

logo

By: Amogh RD, Julien Pinsonneau and Mohamed S. Mahmoud

In OCP ensuring efficient packet delivery is crucial for maintaining smooth communication between applications. However, due to various factors such as network congestion, misconfigured systems, or hardware limitations, packets might occasionally get dropped. Detecting and diagnosing these packet drops is essential for optimizing network performance and maintaining a high quality of service. This is where eBPF (extended Berkeley Packet Filter) comes into play as a powerful tool for real-time network performance analysis. In this blog, we’ll take a detailed look at how network observability using eBPF can help in detecting and understanding packet drops, enabling network administrators and engineers to proactively address network issues.

Detecting Packet Drops with eBPF

eBPF enables developers to set up tracepoints at key points within the network stack. These tracepoints can help intercept packets at specific events, such as when they are received, forwarded, or transmitted. By analyzing the events around packet drops, you can gain insight into the reasons behind them. In network observability we are using the tracepoint/skb/kfree_skb tracepoint hook to detect when packets are dropped, determine the reason why packets drop and reconstruct the flow by enriching it with drop metadata such as packets and bytes statistics, For TCP, only the latest TCP connection state as well as the TCP connection flags are added. The packet drops eBPF hook supports TCP, UDP, SCTP, ICMPv4 and ICMPv6 protocols. There are two main categories for packet drops. The first category, core subsystem drops,covers most of the host drop reasons; for the complete list please refer to drop-reason. Second, there are OVS-based drops, which is a recent kernel enhancement; for reference please checkout the following link OVS-drop-reason.

Kernel support

The drop cause tracepoint API is a recent kernel feature only available from RHEL9.2 kernel version. Older kernel will ignore this feature if its configured.

How to enable packet drops

By default packets drops detection is disabled because it requires privileged access to the host kernel. To enable the feature we need to create a FlowCollector object with the following fields enabled in eBPF config section

apiVersion: flows.netobserv.io/v1beta1
kind: FlowCollector
metadata:
  name: cluster
spec:
  agent:
    type: EBPF
    ebpf:
      privileged: true
      features:
        - PacketDrop

A quick tour in the UI

Once the PacketDrop feature is enabled, the OCP console plugin automatically adapts to provide additional filters and show information across Netflow Traffic page views.

Open your OCP Console and move to Administrator view -> Observe -> Network Traffic page as usual.

Now, a new query option is available to filter flows by their drop status:

drop filter query option

Two new filters, Packet drop TCP state and Packet drop latest cause are available in the common section:

drop state & cause filters

The first one will allow you to set the TCP state filter:

state filter

The second one will let you pick causes to filter on:

cause filter

Overview

New graphs are introduced in the advanced options -> manage panels popup:

advanced options

Select the desired graphs to render them in the overview panel:

drop graphs 1 drop graphs 2 drop graphs 3

Note that you can compare the top drops against total dropped or total traffic in the last graph using the kebab menu drop graph option

Traffic flows

The table view shows the number of bytes and packets sent in green and the related numbers dropped in red. Additionally, you can get details about the drop in the side panel that brings you to the proper documentation.

drop table

Topology

Last but not least, the topology view displays edges containing drops in red. That’s useful especially when digging on a specific drop reason between two resources.

drop topology

Potential use-case scenarios

while : ; do curl <another nodeIP>:<unknown port>; sleep 5; done

The drops can be observed on the console as seen below:

NO_SOCKET drop table

NO_SOCKET drop overview

OVS drop table

OVS drop topology

OVS drop overview

Resource impact of using PacketDrop

The performance impact of using PacketDrop enabled with the Network Observability operator is noticeable on the flowlogs-pipeline(FLP) component using ~22% and ~9% more vCPU and memory respectively when compared to baseline whereas the impact on other components in not significant (less than 3% increase).

Feedback

We hope you liked this article !

Netobserv is an OpenSource project available on github. Feel free to share your ideas, use cases or ask the community for help.