/topk-ebpf

Primary LanguageC

TopK ebpf implementation

This repository provides a eBPF implementation of the polylog TopK algorithm (https://dl.acm.org/doi/10.1145/948205.948227)

There are 2 ways to compile and run eBPF programs:

  • Using bcc[https://github.com/iovisor/bcc] and python bindings. This is the easier way of running stand alone bpf programs.
  • Using llvm + clang directly and libbpf. This allows better integration of bpf into larger C/C++ projects.

Background readings

Network Telemetry at the SmartNIC

Useful Information and Tools

Kernel Versions Newer kernel is always preferred as new features are added with almost every kernel minor version.[https://github.com/iovisor/bcc/blob/master/docs/kernel-versions.md]

ip link ip link show can be used to check the hook point of xdp programs

12: enp4s0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdp qdisc mq state UP mode DEFAULT group default qlen 1000
    link/ether 00:15:4d:13:3d:12 brd ff:ff:ff:ff:ff:ff
    prog/xdp id 151

Possible hookpoint tags are xdp native XDP mode, using XDP hookpoint with driver support xdpgeneric generic XDP mode, using TC hookpoint without driver support xdpoffload offloaded XDP mode, using supported nics, e.g. Netronome

sudo ip link set dev enp4s0 xdp off can also be used to purge all ebpf/xdp programs hooked to a specific nic.

Running ebpf/XDP with python and bcc

NOTE: different bcc tool may have slightly different API, so the example program may need some tweek to work with latest bcc relase.

TopK

There are 2 files in this example:

  • topk_xdp.c contains the eBPF program.
  • xdp_loader.py is the bcc binding that does compile, load, interact with ebpf all-in-one.