The speed test of 900 is for operators. File transfer over LAN is so slow that you need to test the intranet first. I used iperf3 to find out.

The reason for this is very simple. The broadband at home was upgraded to Gigabit last year, and the Speedtest ran over 940. I took a screenshot and sent it to the family group to show off, and I was slapped in the face that night.: When dragging a 4G movie from the desktop to the shared folder of the laptop, the progress bar stays at around 35MB/s all year round, and occasionally drops to 20. My first reaction was that the broadband was falsely marked, and my second reaction was that the router was not working, so I almost ordered a new router. Later, I thought about it a little more and realized that these two things are not the same path at all.

Speedtest tests the section from my home to the operator's computer room, which is the broadband itself. When the computer transfers the data to the computer, the data does not leave the door at all.: Starting from the desktop network card, entering the router LAN port, being exchanged, and then to the laptop's wireless network card, the whole process is at home. A broadband test of 900 can only prove that the "way out" is wide, but not the "way inside the house". The indoor part needs to be measured separately. The tool is iperf3, which is free and open source and available for Windows and Linux. I have used it and it is much more reliable than the "WiFi Speed ​​Test Master", and the numbers can match the theoretical values.

Make it clear first: Divide a data path into three sections and test each section

I drew my home network into three sections: The first section is broadband. The optical modem goes to the operator. Speedtest manages this section. ； The second section is the wired intranet, where devices plugged into the router's LAN port can communicate with each other. The upper limit of this section is determined by the switching capabilities of the network card and router. ； The third segment is wireless. Laptops and mobile phones use WiFi. The upper limit of this segment is determined by the negotiated rate. It is the easiest segment to shrink among the three segments. File transfer is slow and 80% of the time it is stuck in the second or third paragraph, which has nothing to do with the first paragraph.

Therefore, the correct order of investigation is to isolate segment by segment.: First test the wired pair and lock the LAN port of the router. ； Then test the wired versus wireless test and isolate the WiFi. Repair whichever section is down, don't change the router if it's slow. I later finished the test in this order and found that my questions were two overlapping sections.: The network cable on the desktop is an old 100M cable, and the laptop is still connected to 2.4G. When testing separately, the accounts of each section are calculated clearly. Each of the two pits occupies one. When they are overlapped, only the vague number "35MB/s" can be seen.

How does iperf3 run? Two commands are enough

The usage of iperf3 is to run from both ends: One is used as a server to wait, and the other is used as a client to receive traffic. After installing it, type on your laptop iperf3 -s , it hangs there and listens to port 5201 ； Then go to the desktop and type iperf3 -c 192.168.1.123 (The following string is the internal IP of the notebook), run for ten seconds, and spit out a line of results. Look at two numbers: The Mb/s in the Interval column is the rate, and the last Sum is the general ledger. Pay attention to the unit. What is given here is Mb (megabit). Divide by 8 to get what we usually call MB. The negotiated rate of 2402Mb/s theoretically reaches a transmission speed of 300MB/s. Many people are fooled by the unit here. When they see that the numbers do not match up, they think the measurement is wrong.

There is a small pit under Windows: The server firewall blocks port 5201 by default. When running the client for the first time, it will report unable to connect to server , don’t suspect that the installation is broken, go to the firewall to release the port or temporarily turn it off before running. In addition, the single-thread numbers occasionally jitter, so run two or three more rounds to get a stable state. Just remember a comparison: On my Gigabit wired-to-wired machine, the iperf3 is stable at 930 to 940Mb/s. This is the real ceiling of the Gigabit port after deducting the protocol overhead. It is not the same as the 940 used by Speedtest to measure broadband. One number is inside the house and the other is outside the door.

Test the line segments first: My 100Mb cable measured 94Mb

In the first round, both the desktop and laptop were plugged in with network cables, and both were connected to the LAN port of the router. It ran out to 94Mb/s. I stared at this number for a long time - Gigabit port versus Gigabit port, how come it was cut down to a fraction? Later I found the answer in the router background: In the device details, the network port status of each machine is clearly written in the desktop column. 100M full duplex . The network cable was given by the master when the machine was installed in the early years. Only four cores are valid. Gigabit requires all eight cores. If the negotiation fails, it will automatically fall back to 100M.

It is easy to change the way of thinking and verify: I put on another spare line, and the background immediately became 1000M full duplex , iperf3 is run again, the number jumps from 94 to 938. This broken cable has been in my house for four years, and I always thought it was fine - because 100Mbps is enough for watching videos and browsing the web, but it didn't show its original shape until I transferred large files. Lessons learned: The negotiated speed of the network port is more honest than the appearance of the cable. This is especially true for long-distance embedded cables. If the crystal head on the panel is oxidized and crimped poorly, it will silently drop to 100 Mbits, and the appearance cannot be seen at all.

Wireless segment retest: The gap between 2.4G and 5G is bigger than imagined

After the network cable was repaired, the wired to wireless test was repeated again, and this time it was the laptop's turn to reveal the secret. It has been hanging on 2.4G, and iperf3 is only about 90Mb/s - worse than that 100M broken line. Switch the notebook to 5G and keep the same position, and the number will directly reach around 700. By the way, the negotiated rate corresponding to 700 is 2402Mb/s, which is the "theoretical speed" reported by the WiFi6 notebook network card and the router when handshaking. It is considered healthy if it can actually run at around 30%. The loss in the middle comes from signal quality, retransmission and airtime competition. The negotiated rate itself is a statement of the upper limit, not a promise.

Here is an explanation of an old question: Why is my phone still slow when connected to "full" WiFi? The number of grids only reflects the signal strength, not the frequency band. For the same notebook, the measured 2.4G full capacity is 90, and the 5G full capacity is 700, a difference of nearly eight times, without mentioning a word about the grid number. When the router combines dual bands into one name in the background, which band the device chooses depends entirely on its own algorithm. Older devices prefer 2.4G because it can transmit farther. This is why my laptop sticks to 2.4G - it's not that it's broken, it just "feels" that 2.4G is more stable. If you want your device to use 5G, you have to either disassemble the dual-band and manually connect it with a name for each, or see if the router has a band guidance switch. Different brands call it different. Xiaomi's one is called "WiFi Signal Enhancement", which is actually a guidance algorithm. There is a 100Mbit difference between the two mobile phones and the WiFi speed test. I am The article about the speed difference between two mobile phones I've looked at the more detailed ones, and the same phone can be quite different because it's locked into different frequency bands.

After measuring a bunch of numbers, how do I get them right?

After getting to the bottom of it, the problem will be more specific.: My "35MB/s" actually = the 94Mb÷8≈11.7MB/s section of the 100M line, superimposed on the 2.4G wireless section of 90Mb÷8≈11MB/s. After taking the shortcomings, the bottleneck is about 11MB/s, plus the overhead of the SMB protocol itself and the disk writing speed, the final result is 35MB/s - No, I need to correct myself here.: Before changing the line, the file drag rate when I was dragging was 35MB/s. According to the shortcomings, it should only be about 11MB/s. I looked back through the records and realized that the laptop was actually connected to 5G that day. 35MB/s was the shortcomings of 100M wired 11.7 and wireless 87.5 plus the protocol overhead. I forgot the scene. You see, this is the disadvantage of not testing section by section - relying on impressions to logarithms, even remembering your own pitfalls. Later, I measured all four sections and recorded them in a table. It was clear at a glance which number had been moved or not.

There is only one logic for getting seated: The slowest section between the two ends . Wired 94 with wireless 700, actually 94 ； Gigabit broadband is equipped with 2.4G 90, no matter how wide it is when you go out, it will be useless. Therefore, the word "slow Internet speed" is too bold. Later, I broke down the six slow methods at home one by one and wrote them in Six slow methods for symptomatic investigation In that article, slow file transfer is only one of them, and speed testing of the video card is another matter. The judgment caliber is completely different. This round of addition is the "test the internal network by yourself" link. Among the six slow methods, anyone who doubts the wiring in the house can use iperf3 to test it by himself, without having to guess.

write at the end: Find out first, then spend money

The final bill for this round of trouble is zero yuan: The spare parts left over from the installed machine are used to replace the network cable, and switching to 5G is free of charge. If I had ordered a new router based on the first reaction, it would have cost more than a thousand yuan, and the 35MB/s would probably not move at all, because the bottleneck is not on it at all. My order is fixed and I recommend it to people around me now.:

• The first step is to measure the bandwidth with Speedtest and confirm that the section outside the door has not shrunk. This is the baseline.
• In the second step, plug in the cable at both ends and run iperf3. Check the network port negotiation speed in the background to check the cable and port.
• The third step is to switch to WiFi on one end and run again. Test 2.4G and 5G once each to see how much the negotiated rate is different from the actual rate.
• The fourth step is to reconcile the shortcomings, change the line when it is necessary, and switch the frequency band when it is necessary. If everything is normal, then consider upgrading the equipment.

By the way, I would like to remind you of two directions where it is easy to work in vain. First, don’t use Speedtest to draw conclusions about the local area network. It can’t measure the area inside the house. No matter how high the number is, it has nothing to do with transferring files. The caliber of broadband repair reports is another matter. I’m here The article about repair speed test I wrote about why the master only recognizes the number in the business hall App. Second, don’t be superstitious about the negotiated speed. 2402Mb/s is just a handshake statement. Let’s see how much iperf3 actually outputs before talking about the bottleneck. Who will be held back if there are too many devices? Will devices with weak signals affect the whole family? That is another topic. There is data in the article about the actual measurement of the corridor camera. Do some research first, then talk about spending money. This sequence can save most of the wasted money.