HN Debrief

Debugging my new network, when 10 Gigabit Ethernet Runs at 300 Megabits

  • Infrastructure
  • Hardware
  • Networking
  • Developer Tools

The post is a hands-on debugging story about a new 10GbE home network that stubbornly ran at roughly 300 Mbps on Windows. The eventual fix was not exotic hardware replacement. It was tuning the network adapter, especially receive buffers and offload settings, after basic checks showed the cabling and switching were fine. That made the piece useful less as a one-off Windows rant and more as a reminder that high-speed Ethernet still fails in boring ways. Driver defaults, hidden counters, and half-exposed tuning knobs can wipe out most of the expected gain.

If fast local links are underperforming, do not stop at cabling and switch checks. Test each hop with iperf3, inspect NIC driver settings and counters, and assume defaults on consumer gear may be badly tuned for 10G.

Discussion mood

Positive on the writeup and strongly pragmatic. People liked the concrete debugging process, but the mood was also weary because network performance still gets kneecapped by cheap hardware choices, stale driver defaults, flaky power management, and terrible device feedback.

Key insights

  1. 01

    ASPM can break internet speed only

    PCIe Active State Power Management can create a nasty false negative where a NIC benchmarks well on a local network but slows badly on real internet paths. The key detail is latency. On a nearby LAN the card never has time to enter a low-power state, while a few milliseconds of WAN delay can trigger sleep and make exit latency the bottleneck.

    If LAN tests look healthy but WAN throughput collapses, check PCIe power-management settings before blaming your ISP or router. Reproduce with both low-latency and higher-latency targets so you can tell whether the failure is speed-related or latency-triggered.

      Attribution:
    • ValdikSS #1
  2. 02

    Slow ports hurt through switch behavior

    A 100 Mbps device does not automatically cap every other link on the network. The real danger is broken flow control or shallow switch buffers. Pause frames can stall unrelated outbound traffic, and a 10G sender aimed at a 100M receiver can overflow cheap switch memory fast enough to cause drops elsewhere.

    When one low-speed endpoint seems to poison the whole LAN, investigate switch flow control and buffer behavior instead of assuming Ethernet itself is the problem. Turning off pause frames or replacing the switch can be more effective than touching endpoints.

      Attribution:
    • toast0 #1
    • adastra22 #1
    • blahlabs #1
  3. 03

    Old driver defaults are the hidden bottleneck

    High-speed Ethernet often inherits tuning from an earlier era. Too few receive buffers and aggressive interrupt moderation were acceptable on 100M or 1G cards, then quietly became disastrous as the same driver stack grew into 10G. The comments also pushed back on blindly increasing buffers, since oversized queues can trade throughput gains for latency and bufferbloat.

    For 10G links, inspect driver-era assumptions like ring sizes, coalescing, and offloads rather than trusting stock settings. Optimize for the workload you care about, because the throughput-maximizing knob setting may also make interactive traffic worse.

      Attribution:
    • toast0 #1
    • iknowstuff #1
  4. 04

    iperf3 is the first sanity check

    The most useful tool named was iperf3 because it separates raw link performance from application weirdness. One commenter used it immediately to validate their 10GbE cabling and NIC path, which is the right move before chasing OS tuning myths or swapping hardware at random.

    Keep iperf3 in your standard bring-up checklist for new networks. Prove the link can move traffic at line rate first, then debug storage, application, or remote-server issues with a narrower search space.

      Attribution:
    • tacomagick #1
    • Arainach #1
  5. 05

    Physical layer mistakes still win

    Even on modern gear, something as simple as an SFP+ module not being fully seated can produce bizarre partial functionality instead of a clean failure. That is exactly why experienced operators keep repeating the unfashionable advice to check the physical layer first.

    Before changing drivers or registry settings, reseat optics, DACs, and transceivers and verify link training details on both ends. Partial insertion and marginal physical connections can mimic software bugs surprisingly well.

      Attribution:
    • alexnewman #1
    • chasd00 #1
    • evulhotdog #1
  6. 06

    Console downloads need packet-level debugging

    Slow PlayStation downloads were framed as a separate but related case where link speed is often not the bottleneck at all. The useful suggestions were to capture packets, look for path MTU issues, small TCP windows, bad IPv6 behavior, distant mirrors, or loss, and in some cases front the console with Squid because Sony's own networking path appears to behave poorly.

    If a device downloads slowly despite a fast link, sniff traffic and measure the path instead of treating it like a raw Ethernet problem. Consumer platforms can have broken stack behavior that only shows up with packet captures or a proxy in the middle.

      Attribution:
    • toast0 #1
    • robotmay #1
    • OptionOfT #1

Against the grain

  1. 01

    This is not just Windows being bad

    Blaming Windows alone misses the broader pattern. The sharper point is that many operating systems and drivers expose poor defaults and weak diagnostics for high-speed networking. Windows may be where this failure surfaced, but the failure mode is generic enough that Linux, VMs, and appliances deserve the same scrutiny.

    Do not let OS tribalism narrow your investigation. Use the same measurement-first process across platforms and look for vendor driver quality, firmware quirks, and hidden tuning limits everywhere.

      Attribution:
    • anonymousiam #1
    • toast0 #1
  2. 02

    100 Mbps TV ports are often fine

    The complaints about televisions shipping with Fast Ethernet ignore the actual workload. Streaming video rarely needs gigabit, and the lower-speed PHY saves money, power, and heat. In plenty of homes, good Wi‑Fi will outperform the wired port on a TV anyway.

    Do not treat every 100 Mbps port as a product flaw. Match the hardware choice to the use case, and reserve outrage for cases where the workload really can exceed Fast Ethernet or where the device degrades the rest of the network.

      Attribution:
    • Aurornis #1
    • toast0 #1
    • somehnguy #1
    • tshaddox #1

In plain english

10GbE ↩
10 Gigabit Ethernet, a network standard that carries data at up to 10 billion bits per second over Ethernet.
ASPM ↩
Active State Power Management, a PCIe power-saving feature that puts devices into lower-power states when idle.
bufferbloat ↩
Excessive latency caused by network buffers that are too large and stay full under load.
iperf3 ↩
A command-line tool used to measure maximum network throughput between two machines.
IPv6 ↩
Internet Protocol version 6, the newer internet addressing system designed to replace IPv4.
LAN ↩
Local Area Network, a network covering a small area such as a home, office, or data center rack.
MTU ↩
Maximum Transmission Unit, the largest packet size a network link can carry without fragmentation.
NIC ↩
Network Interface Card, the hardware that connects a computer or device to a network.
offload settings ↩
NIC features that move parts of packet processing from the CPU to the network card, which can improve speed but sometimes cause bugs.
PCIe ↩
Peripheral Component Interconnect Express, the high-speed internal bus used to connect devices like network cards to a computer.
PHY ↩
Physical layer transceiver, the hardware part of a network interface that handles the electrical or optical signaling on the cable.
SFP+ ↩
Small Form-factor Pluggable Plus, a modular transceiver format commonly used for 10 gigabit network links.
Squid ↩
A caching and forwarding proxy server often used to mediate web traffic.
TCP ↩
Transmission Control Protocol, the main internet transport protocol that provides reliable ordered delivery of data.
WAN ↩
Wide Area Network, a network connection that spans larger distances, such as internet paths between cities or countries.

Reference links

Networking diagnostics and tuning

Background analogy