High-Performance Computing (HPC) and modern AI workloads (like LLM training) depend on RDMA (Remote Direct Memory Access) to keep latency near zero and free up CPU overhead.
While InfiniBand has traditionally dominated this space, RoCE (RDMA over Converged Ethernet)—pronounced "Rocky"—is rapidly becoming the standard for open, Ethernet-based AI fabrics.
Whether creating training materials, writing technical documentation, or building Proof-of-Concept (PoC) demonstrations—for public courses or for enterprise internal training—we always need example IP addresses.
You might ask: Are there any reserved IP blocks that are guaranteed never to overlap with live production environments?
The answer is yes. The IETF has already defined dedicated blocks across several RFCs. This article summarizes the reserved Unicast address ranges you should be using.
【Quick copy/paste】
Google: (Not recommended: static IP addresses: 216.239.35.0 or 216.239.35.4)
Configuring NTP servers with hostnames is a better approach.
TIME.google.com
TIME1.google.com
TIME2.google.com
TIME3.google.com
TIME4.google.com
Cloudflare:
TIME.cloudflare.com
National Time and Frequency Standard Laboratory of Taiwan:
TICK.stdtime.gov.tw
TOCK.stdtime.gov.tw
Maintaining an accurate and synchronized time state across all hardware components is a non-negotiable operational standard in enterprise infrastructure. Network Time Protocol (NTP)—a UDP-based, user-mode network protocol—serves as the foundational baseline for achieving this requirement across servers, network devices, and client nodes.
Most computing systems rely on battery-powered internal electronic clocks that naturally drift over time, and few organizations can justify the exorbitant expense of dedicated atomic clock hardware for every node. This time variance becomes highly problematic when diagnosing system issues. It is extremely difficult to reconstruct the timeline about what events happen first and what events come after when analyzing log files recorded on different nodes without synchronized system time.
Because a standard NTP implementation delivers millisecond-level precision, it allows engineers to accurately merge, and correlate logs collected across entirely separate infrastructure nodes. Without this synchronization, reconstructing historical event timelines for troubleshooting is nearly impossible.
In the world of high-performance networking, engineers and operators often focus their attention on the switch chassis itself. When evaluating power budgets or structural installation, it is common to view the switch as the primary, expensive "brain" of the system, while treating the SFP transceiver modules as mere auxiliary accessories—side components that add little to the overall system load. However, a closer analysis of high-density systems, such as the 48-port Cisco Catalyst 9500, proves that this assumption is a significant misconception.
This article examines the validity of this assumption. Let us look closer at all transceivers versus the switch main body.
 |
| Cisco Catalyst 9500 (C9500-48Y4C) Source (Thanks!): Cisco Catalyst 9500 Series Switches Hardware Installation Guide |
 |
| 25G transceiver (SFP-25G-SR-S) Source (Thanks!): Cisco 25GBASE SFP28 Modules Data Sheet |
I choose Cisco Catalyst 9500 (C9500-48Y4C) as an example target for analysis. This Catalyst 9500 has forty-eight user ports, and I install forty-eight 25G transceivers (SFP-25G-SR-S) into each port. We analyze in three dimensions: Power Consumption, Heat Generation, and Financial Cost.
To achieve higher transmission bit rate, Ethernet fiber optic transceivers often utilize parallel optics to transmit data across multiple fiber lanes simultaneously.
Think of a single pair of fibers as one data lane for sending and receiving. By bundling four pairs of fibers, a transceiver can transmit and receive four times the amount of data simultaneously. For instance, if a single fiber pair delivers a throughput of 10 Gbps, a four-pair configuration can scale up that speed fourfold to 40 Gbps.
Standard transceivers that only require a single pair of fibers typically use LC (Lucent Connector) interfaces. However, when a transceiver needs to connect multiple fiber pairs through a single port, it must use high-density MPO/MTP connectors.
MPO is an open industry standard. Its name perfectly reflects its design philosophy: a high-density, multi-fiber connector with a simple push-on mating mechanism. Common configurations include MPO-8 (8 parallel fibers) and the more widespread MPO-12 (12 parallel fibers).
 |
| Figure 1: MPO-12 vs MPO-8 Source (Thanks!): https://www.flukenetworks.com/blog/cabling-chronicles/getting-12-and-8-fiber-polarity-right |
Most people know about Internet Service Providers (ISP). People must subscribe to at least one ISP to connect to the Internet. Operating behind the scenes, specialized network infrastructures play a crucial role in ensuring the Internet works correctly and efficiently. This role is the Internet Exchange Point (IXP). In this post, I will explain in greater depth why we need this new role, and what problem this role solves.
On March 23, 2026, Federal Communications Commission (FCC) of the United States announced, “Covered List to Include Foreign-Made Consumer Routers, Prohibiting Approval of New Models”. When a product is in Covered List, it means FCC will not inspect that product. That is, no permission at all from FCC for us to import, sell, or use that product inside the United States. Effectively, the United States bans that product. That is why most press said, “US bans foreign-made routers.”
FACT SHEET: FCC Updates Covered List to Include Foreign-Made Consumer Routers, Prohibiting Approval of New Models
It looked frightening when I first read it. The impacts are not immediate, however.
 |
| Jianshan Hill Observation Deck (尖山觀景平台) overlooking Taipei City, New Taipei City, and Tamsui River of Taiwan. |
【 BGP Millionaire】
 |
| I used this tool to generate this plot. Plot Range: 01-Feb-2025 0001 to 30-Apr-2025 0030. |
Recently I observed and noticed that, global IPv4 BGP table size has already grown over 1 million entries, around February to April time in this year, 2025.
A submarine cable cutting incident happened recently at the Baltic Sea. Another similar incident also happened north of Taiwan.
I wanted to know more about how to repair a broken submarine fiber cable. I did my research on the Internet. I summarize my findings and list the key fixing steps below.
1. The submarine cable was working.
Initially, the working submarine cable was below the sea water and above the seabed. The submarine cable follows and fits the terrain of seabed.
2. A fault happens.
Examples of faults are punctured insulation, fiber core breaks, and cable cut.
Faults could be natural factors, such as the aging of cable itself, earthquakes, falling rocks, and shark bites. Faults could also be human factors, such as fishing nets dragging, anchoring hits, and malicious sabotage.
How many of you are working with BGP in your daily job? Let me know in the comments below this post.
I also created a live demonstration about filtering based on BGP AS-Path with Regular Expressions. I hope it is helpful to you.
I was so amazed about this cooling technology. It is new to me.
Thinking about liquid cooling for servers and networking devices, coming into my mind are tightly sealed water pipes around the circuit boards. This is different. The circuit board and everything simply immerse in liquid.
This is an interesting milestone for IPv6. That means a massive majority of people are using IPv6 today. I want to note down this moment. And I want to share 3 of my own observations about the IPv6 BGP table.
Simple visual tool to calculate Cisco IOS Wildcard Mask
