The cheapest speed-up for an old NAS is the port, not the CPU

Every "make your NAS faster" article starts with the CPU. On the machines this site is about, that is the wrong end of the wire. Copying a file to a NAS crosses four things — the client's disk, the network link, the server's disk, the server's CPU — and the slowest one sets the pace. On a 2011 laptop board, that is almost always the Ethernet port, and it is the one part you cannot fix by installing anything.

We have two boxes from that era. One has a gigabit port. The other has a 100 Mbps port soldered to the board, and it is the machine that taught us this lesson, twice, in the wrong order. That second box is the laptop written up in An old laptop is a NAS with a UPS already inside it — this is the same port that caps it.

What we measured, and what we did not

Everything in the tables below came off our own hardware: hdparm, dd with conv=fsync, ethtool, /proc and /sys, and a byte-for-byte comparison after each transfer. We did not run iperf3 between the two boxes, we have not measured a USB Ethernet adapter on our own machine, and we have not tested a mini-PCIe replacement card in the slot. Where a number is a specification or an estimate, the sentence says so.

The one habit worth copying: measure the disk first, then the link. We got that order wrong once and spent an evening blaming the network for what was a dying 5400 rpm drive.

The same file, three different ceilings

One JavaScript bundle, 2.89 MB, served three ways from the same machine:

Read the third row carefully. Once the drive was fixed, that one file moved at 93 Mbps on a 100 Mbps link — the port was full. The box had not become faster in any general sense; it had run out of network exactly as designed, and the browser stopped waiting.

Now the second row. The same port, the same file, 24 seconds. Nothing about the network changed between those two rows. What changed was that the file had been read off a 2011 5400 rpm notebook drive that could not serve a few dozen concurrent small reads in parallel. That is the first lesson: a slow transfer is a queueing problem before it is a bandwidth problem, and you cannot tell which by watching the progress bar.

The numbers that decide it

Put those two facts together and the conclusion is uncomfortable: on a 100 Mbps machine, a faster CPU, more RAM and a faster SSD all improve things you cannot see from the client. They make thumbnails generate sooner and the web UI respond quicker. They do not move a file across the wire any faster, because the wire is already saturated.

The order to spend money in

1. Nothing. Run the two commands below and find out which of the four stages is slow. It is free, and roughly half the people who ask us this question discover the bottleneck is the client's Wi-Fi, not the NAS.
2. A cable and a switch port. Cat5 was specced for 100 Mbps and often works at gigabit; Cat5e is the safe minimum. A switch that only does 100 Mbps on that port is more common than people expect, and it is invisible unless you look.
3. A USB 3 gigabit adapter — if the machine has USB 3. Both of ours are USB 2.0 only (EHCI controllers, no xHCI), so an adapter is capped by the bus at roughly 25–35 MB/s in practice. That is still two to three times faster than 100 Mbps, and for a 30-40 currency unit part it is the best value on this list. On a laptop with only USB 2.0 ports, it is also the end of the road.
4. A mini-PCIe or PCIe card. Only where a slot is free, and only after checking the physical size. Our laptop's single half-height slot (26.8 mm) is occupied by the Wi-Fi card, so the trade is Wi-Fi for Ethernet. The slot is PCIe x1 Gen1, about 250 MB/s each way — enough for gigabit, not enough to saturate 2.5 GbE.
5. A different machine. At some point the honest answer is that a 100 Mbps board from 2011 is a 100 Mbps board, and the money is better spent on a box with the port you want. We say this reluctantly; see the mini PC guide for what that costs.

Two commands that answer the question

# 1. What did the link actually negotiate?
cat /sys/class/net/eno1/speed          # 100 or 1000, in Mbit/s
sudo ethtool eno1 | grep -E 'Speed|Duplex|Advertised link modes'

# 2. What can the disk do, with the page cache out of the way?
sudo hdparm -t --direct /dev/sdX

If the first command prints 100, stop reading buying guides and go check the cable and the switch. If it prints 1000 and transfers still crawl, the second command tells you whether the drive is the reason — and if the drive is fine, the remaining suspects are SMB itself and the client.

Two more things worth knowing before you blame hardware:

• dmesg | grep -i 'link is up' shows what the port negotiated at boot. Ours logs a Link is Down / Link is Up - 100Mbps/Full pair every time the cable is touched, which is how we found a marginal connector.
• Small files are latency, not bandwidth. Copying 20,000 photos will never reach the link's ceiling — each file costs a round trip. Test with one file of a few hundred megabytes, or you will be measuring protocol overhead and concluding your hardware is broken.

What the port does not explain

Two things on our boxes look like network problems and are not:

• The web UI loading slowly on the first visit. That is the drive, not the link: dozens of small asset reads behind one another. It went from 24 seconds to a quarter of a second on the same 100 Mbps port. Details in the SSD guide.
• Thumbnail and OCR queues. Those are pure CPU. A photo library generates thumbnails at roughly 91 images per minute on a 2011 two-core chip, and no network upgrade changes that number by a single image.

What we would not buy

• A 2.5 GbE adapter for a machine whose disks top out at 107 MB/s. You would be buying a wider pipe into the same bucket. Gigabit is the right target for any 2.5" platter.
• A 2.5 GbE adapter in a USB 2.0 port. It negotiates gigabit at best and usually less; the bus is the ceiling, not the PHY.
• "Cat8" cables for a 15-metre run. Cat5e does gigabit at that distance. The extra money buys shielding you do not need in a cupboard.
• Powerline adapters marketed on their PHY rate. Real throughput depends on your mains wiring, and the number on the box is not the number you get.
• A CPU upgrade to fix transfer speed. We have watched people do this. The port does not care.

What we are still measuring

Whether a half-height mini-PCIe gigabit card physically fits the laptop and survives its BIOS. The driver support is already there — the kernel on that machine ships the r8169 aliases and firmware for the RTL8125 family, so a card from that family should come up without a build — but a 26.8 mm slot, a card that may be full height, and a vendor BIOS that has historically rejected unsupported wireless cards is three unknowns too many to recommend before we have tried it. When we do, this page gets a new date.

Until then the honest ranking stands: check the cable, check the switch port, then buy the cheapest adapter your bus can actually feed — and do not touch the CPU. The bottleneck is a number you can read in one command, and it is almost never the one the upgrade guides are selling.