Running a NAS, Part 5: Tuning the NAS Network

A faster drive does nothing for a file copy that the network can’t carry. Part 4, SSDs in a NAS, put flash in the box as cache, as ZFS special devices, and as whole pools, and one SSD can outrun the link in front of it. Part 5 of Running a NAS follows the path from the NASNetwork-attached storage is a storage system that gives other computers access to its files over a network. to your PC. It covers what each Ethernet speed delivers, which settings speed up one transfer and which only help many, and what to check first. Each setting comes with the condition under which it helps.

My own server runs Unraid with dual parity on a 10GbE network, and it holds 20TB hard drives and six 3.2TB SAS SSDs. At that speed the link is no longer the obvious limit, which is why the order below starts with measuring.

Part 6 compares Server Message Block (SMB) with the Network File System (NFS) and with iSCSI. Here a protocol matters only where it changes how many network lanes one transfer can use. Storage networks for AI clusters are a different job, covered in the Networking for SI series, which starts with NVMe-oF and the Network Behind SI Storage.

What each link speed delivers

Ethernet speeds are written in bits and file sizes in bytes, so the first step is dividing by eight. The table does that for four common speeds, using the rate the IEEE standard behind each one defines.

Source: IEEE 802.3bz-2016, 802.3an-2006 and 802.3by-2016 (IEEE Standards Association pages), and Realtek’s RTL8125BG page for 802.3ab. The raw rate is the line rate divided by eight, before any protocol overhead.

Real transfers land below the raw rate

Protocol overhead takes a share of those raw rates. NetApp’s knowledge base puts 1GbE with TCP and standard 1,500-byte frames at about 116MB/s, and with 9,000-byte jumbo frames at about 123MB/s, which is a gain of about 6 percent by this guide’s arithmetic on NetApp’s calculated figures. That article covers 1GbE only. For 2.5GbE, 10GbE, and 25GbE the raw rates in the table are the most a transfer can reach, and what you get is something to measure.

The drive at the other end

Whatever sits at the other end sets the other limit. Seagate’s IronWolf Pro data sheet lists a maximum sustained transfer rate of 285MB/s for the hard drive. Samsung’s 870 EVO data sheet lists 560MB/s sequential read for the SATA SSD, and its 990 PRO data sheet lists up to 7,450MB/s for the NVMe SSD. Set against the table, those ratings give an estimate of which link a drive can fill. One hard drive at 285MB/s is more than twice what 1GbE carries and sits just under the raw rate of 2.5GbE. The SATA SSD fills 2.5GbE with room to spare and uses less than half of 10GbE. The NVMe SSD’s read rate is more than twice the raw rate of 25GbE.

On my own server’s 10GbE link, a hard drive rated like that Seagate would leave most of the link idle, since 285MB/s is under a quarter of 1,250MB/s. Link speed limits one transfer rather than the NAS, though. Several clients reading from different drives can together use far more of the link than one drive can. I read 2.5GbE as the first upgrade that pays off for a NAS of hard drives, because one drive already fills 1GbE. I see 10GbE as the point where SSDs or several simultaneous users start to matter. Part 4 covers the all-flash boxes and the network ports they ship with.

Cables and how far each link goes

A link is only as good as its cable. The 10GBASE-T standard covers up to 100m of balanced twisted-pair cabling, but a card maker can set a tighter rule. Intel rates its X550-T2 10GBASE-T adapter for Category 6 cable up to 55m and Category 6A up to 100m. A run longer than 55m therefore needs Category 6A for that card, even though the standard allows 100m.

For 2.5GbE, Realtek’s RTL8125BG page says the controller transmits over Cat 5e cable. It gives no length, so existing Cat 5e wiring is the first thing to try, and an iperf3 test on the real run tells you whether it holds.

Short runs inside a rack or on a desk can skip twisted pair. Enhanced small form-factor pluggable (SFP+) ports take a direct-attach copper cable (DAC) or a plug-in module. Cisco sells passive twinax cables up to 5m and active ones at 7m and 10m. Its copper SFP+ module reaches 30m, and its 10GBASE-SR optical module reaches 300m over OM3 multimode fiber. Between a NAS and a switch in the same rack, a short DAC avoids the heat that copper modules add, covered below.

Measure before you change anything

A file copy depends on the network, on the protocol, on the NAS’s processor and on its drives at once, so a slow copy doesn’t say which one is slow. Running iperf3 takes the drives out of the picture. Its server listens on TCP port 5201 by default and a client connects to it.

iperf3 -s                         # on the NAS, or another machine on the same switch
iperf3 -c <server address>        # the PC sends
iperf3 -c <server address> -R     # the server sends, which tests downloads
iperf3 -c <server address> -P 4   # four parallel streams

Run it in both directions, because adding -R is what measures the download speed from the server. Parallel streams with -P can give higher throughput than one stream when the test is CPU-limited, and since version 3.16 each stream gets its own thread. As of November 2026 the project’s news page lists 3.22, released September 29, 2026, as the newest release, so an older package may predate that change.

Once the plain test runs near the raw rate for the link, the -F option sends or receives a real file, which shows whether the storage is the bottleneck for file transfers. If iperf3 reaches close to the raw rate but a file copy doesn’t, the limit sits in the drives, the protocol, or the NAS’s processor. Available system resources and the read and write speeds of the NAS’s storage devices greatly influence overall bandwidth, as QNAP puts it. If iperf3 itself runs slow, the fault is in the link, so go through the cable, the port speeds, the switch, and the settings below.

SMB Multichannel: more lanes for one client

SMB Multichannel lets one client open several network connections to the same share, so a single copy can use more than one lane. Whether it applies depends on the client, on the server software, and on the adapters at both ends.

Windows clients

Microsoft’s SMB Multichannel page says the feature is on by default, with nothing to install. It needs at least two computers running Windows 10, Windows Server, or Azure Local, and network adapters that support Receive Side Scaling (RSS), or several adapters, or teamed adapters. With adapters of different speeds it picks the fastest and leaves the slower ones idle, so a 10GbE port beside a 1GbE port adds no lane. A single adapter without RSS gets no benefit, and Multichannel isn’t used.

Samba

On the server side, the Samba manual says the server multi channel support parameter was added in version 4.4 and was still experimental up to 4.14. It is on by default from 4.15, which means older builds don’t have it switched on. It works only on Linux and FreeBSD because it depends on kernel APIs. The interfaces parameter takes per-interface options such as speed= and capability=RSS, which let an admin describe each interface to Samba.

NAS makers

As of November 2026 the makers differ in the details. Synology’s DSM has an Enable SMB3 Multichannel box in its SMB settings. It is supported only on x64 platforms, so models with ARM processors don’t get it. It isn’t supported while Link Aggregation is on, and it may lower write speed on encrypted shared folders. Its supported clients include macOS 11.3 or above, beside Windows 8 and Windows Server 2012 or above, and Synology recommends that the number of adapters on the NAS divide evenly by the number on the client. A single adapter without RSS makes no difference there either.

QNAP’s SMB Multichannel page dates support to QTS 5.1.0 and QuTS hero h5.1.0. TrueNAS’s networking recommendations say Multichannel needs no switch configuration or link aggregation support, and that enabling both LACP and Multichannel isn’t best practice.

What link aggregation adds, and what it doesn’t

Link aggregation bundles several ports into one logical link. IEEE’s 802.1AX standard describes it as parallel point-to-point links used as if they were a single link, and the makers label the usual mode 802.3ad or LACP, short for Link Aggregation Control Protocol. It serves a crowd rather than one person. Port trunking on a QNAP cannot break the speed limit of a single Ethernet device but gives enough bandwidth for several users at once. Its example is two 1Gb NICs that aggregate to 2Gb while one transfer stays at 1Gb.

Synology’s comparison table agrees. Link Aggregation doesn’t enhance single-client throughput, where Multichannel does, and its Link Aggregation help says it balances traffic for each connected device. On a single-user system LACP brings no benefit at all, according to TrueNAS, and two network cards on one subnet don’t load balance by themselves, although users may assume they do. QNAP’s 802.3ad setup guide makes the default hash policy layer 2, which works with every switch but balances only by MAC address. With that default, one PC has one MAC address, so its traffic stays on one link.

Source: Synology’s SMB3 Multichannel comparison and Link Aggregation help, QNAP’s trunking pages, and TrueNAS’s networking recommendations.

This guide reads LACP as worth its switch ports only when several people use the NAS at once. For one PC, money goes further on a faster link or on Multichannel.

Jumbo frames: a small gain for a risk across the path

A frame is the unit a link sends, and its maximum size is the maximum transmission unit (MTU). Standard frames carry 1,500 bytes and jumbo frames carry 9,000, so a big file crosses in fewer, larger frames. NetApp’s 1GbE figures are about 116MB/s with standard frames and 123MB/s with jumbo frames, roughly 6 percent more by this guide’s arithmetic. Both are NetApp’s calculated figures, not measurements. I found no maker figure for 2.5GbE or 10GbE, so that gain is unmeasured there.

The risk is that every device on the path has to agree. All the computers and devices across the network that reach a Synology NAS must support jumbo frames and use the same MTU value. TrueNAS requires the same of all network hardware, switches and client systems included, to prevent fragmentation and performance degradation, and tells users to set the MTU to 9000 on the parent interface. One device left at the standard size is a reason to leave the setting alone. Six percent rarely justifies a mismatch, so make this the last change and run iperf3 before and after.

Network cards, ports and the watts they draw

A network interface card (NIC) sets the speed, the connector, and the heat. Intel’s I226-V 2.5GbE controller has a TDP of 1.3W. For 10GbE, Intel’s product brief puts the dual-port X550-T2 10GBASE-T adapter at 11.2W typical and 13.0W maximum at 10GbE. It puts the dual-port X710-DA2 SFP+ adapter with a direct-attach cable at 3.3W typical and 3.7W maximum. By that estimate the copper card draws more than three times the typical power of the SFP+ card. Both are one maker’s cards, and each figure is for the cable type its row names.

MikroTik shows the same gap at the module level. It gives its S+RJ10 copper SFP+ module an average of 2.7W on a 30m link, against a maximum of 0.8W for its optical SFP+ module. It also recommends putting S+RJ10 modules in every second port, with an optical module or an empty port between them, because they run hot. The trade is reach: a DAC stops at a few meters, while copper cabling runs up to 100m. Check the NAS’s spec page for which ports it has before buying a card. A 2.5GbE port also calls for a look at the controller chip, not just the port speed. Intel’s specification update for the I225 says a gap-handling error in its first silicon can drop a 2.5GbE link to below 1Gbps, and Part 9, NAS Problems Found Too Late, covers the check.

Switches and virtual LANs

Every transfer crosses the switch, and the slowest port in the path sets the speed for that transfer. TP-Link’s managed-switch guide says an unmanaged switch is plug-and-play with no VLANs, no QoS, and no monitoring. A managed switch adds virtual LANs (VLANs) under IEEE 802.1Q, quality of service (QoS), and link aggregation. QNAP’s 802.3ad mode needs a switch that supports it, which makes LACP the setting that forces a managed switch.

IEEE’s 802.1Q-2022 covers the operation of VLAN bridges. TrueNAS recommends VLANs to isolate NAS traffic, since they split one physical switch into logical segments that each form their own broadcast domain. Synology warns that if the switch isn’t configured to match a NAS VLAN ID, the NAS may become inaccessible. Write down the VLAN ID on both sides before changing either.

What to check and change, in order

Work down this list and stop when a transfer reaches the raw rate of your link, because every later step costs effort or money for no gain.

1. Run iperf3 in both directions. If it reaches near the raw rate, the network isn’t the limit, so look at the drives, the protocol, and the processor, and use -F to test the storage.
2. If it falls short, confirm the speed that each end and each switch port negotiated, then the cable category and length against the limits above.
3. Check the switch. Its ports need to be at least as fast as the link, and it needs to be managed if you plan on LACP or VLANs.
4. For one fast client, turn on SMB Multichannel. Windows has it on by default, DSM has a checkbox, and QNAP supports it from QTS 5.1.0. Confirm the adapters support RSS or come in matching counts.
5. For many clients, set up LACP on a managed switch, and don’t enable it together with Multichannel.
6. Try jumbo frames last, and only when every device on the path agrees on the MTU.
7. Buy hardware for the limit you found. 2.5GbE suits one hard drive once the controller passes the check in Part 9, 10GbE suits SSDs and several users, and SFP+ with a DAC suits short runs where heat and watts matter.

The switch tiers of Best NAS cover the copper speeds in that list, starting at the slowest, and none of their picks takes SFP+.

Once the link is right, the next question is who can reach the NAS and by what route. Part 6, Sharing and Remote Access, compares the sharing protocols and covers reaching the NAS from outside the house, and locking it down against the ransomware campaigns that hit exposed NAS boxes.