---
title: "Discussion on \"Changing a threshold without detaching XDP, and events that cannot flood\" | Hashnode"
description: "Discussion on \"Changing a threshold without detaching XDP, and events that cannot flood\". Building AgentXDP, part 17. Part 16 closed week 3 with a testbed that times every RDMA request. This part is about the requests that are too slow: one threshold per interface, changed while the agent"
url: "https://hashnode.com/posts/changing-a-threshold-without-detaching-xdp-and-events-that-cannot-flood/6afee680f7c5713dfb060187"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Changing a threshold without detaching XDP, and events that cannot flood](https://pegasus-nathanael-tech.hashnode.dev/changing-a-threshold-without-detaching-xdp-and-events-that-cannot-flood?utm_source=hashnode&utm_medium=feed)

Building AgentXDP, part 17. Part 16 closed week 3 with a testbed that times every RDMA request. This part is about the requests that are too slow: one threshold per interface, changed while the agent

1d ago · 7 min read · pegasus-nathanael-tech.hashnode.dev  
Read the article

-   [#ebpf](https://hashnode.com/tag/ebpf)
-   [#cpp](https://hashnode.com/tag/cpp)
-   [#linux](https://hashnode.com/tag/linux)
-   [#observability](https://hashnode.com/tag/observability)
-   [#buildinpublic](https://hashnode.com/tag/buildinpublic)

## Written by

[Nathanael FETUE](https://hashnode.com/@nathanael-fetue)

@nathanael-fetue · Toulouse

  
[View profile](https://hashnode.com/@nathanael-fetue)

From Cameroon to Cloud Engineering at ISAE-SUPAERO/ENSEEIHT |DeepTech & Distributed Systems

Hi, I'm Nathanael. MSc graduate from Toulouse INP - ENSEEIHT (N7), France. I work on hardware/software co-design and low-level performance. I just completed a six-month research internship at INRIA Rennes, where I designed a hardware isolation mechanism for processing-in-memory processors (UPMEM), measured in a cycle-accurate simulator at 5.54% overhead at most, against 26 to 120% for a software scheme, and validated on FPGA. I write about computer architecture, operating systems, networking (RDMA/InfiniBand) and the move from research to industry. I am looking for my first role in systems and performance engineering, AI infrastructure or low-level software. Stack: C/C++, Go, Python, Verilog, Linux, RDMA, Terraform, Ansible. Contact: nathanaelfetuefoko@gmail.com | linkedin.com/in/nathanael-fetue-foko | github.com/NathanaelFetue

## Responses 2

Markdown supported · ⌘↵ to post

[Ehsan](https://hashnode.com/@ibuildinpublic)

[1d ago](https://hashnode.com/posts/changing-a-threshold-without-detaching-xdp-and-events-that-cannot-flood/6afee680f7c5713dfb060187/comment/6b000e5ba33f96a400bdb3ef)

Keeping the threshold in a BPF map so you never detach XDP is a nice fix. When a link degrades, does GCRA drop the noisy burst fast enough that the Prometheus alert still fires on the real counts?

[Nathanael FETUE](https://hashnode.com/@nathanael-fetue)

[23h ago](https://hashnode.com/posts/changing-a-threshold-without-detaching-xdp-and-events-that-cannot-flood/6afee680f7c5713dfb060187/comment/6b000e5ba33f96a400bdb3ef#reply-6b0028f620343d339bba5859)

Thanks for reading it this closely, good question!

The alert doesn't depend on GCRA at all. Every slow request is counted before the limiter gets involved, so the counter always has the real number. GCRA only decides how many JSON lines get written. The alert reads the counter, so it still fires on the real counts.

The only place I can lose events is earlier, if the kernel ring buffer or the analysis queue fills up. I count those losses separately and have an alert on them, because that's exactly when things are going badly. I haven't tested this on real hardware yet, only on my Soft-RoCE testbed, so I'm curious whether you've seen it behave differently in practice.
