Inconsistent SMB Throughput on 10Gb Network (Windows Host vs Laptop vs VMs)

[tabby27]

Issue Summary: Inconsistent SMB Throughput on 10Gb Network (Windows Host vs Laptop vs VMs)

I’m troubleshooting a persistent throughput issue on what should be a near pro-grade 10GbE network (unamnaged switches), and I’m trying to identify whether this is a Windows SMB / Hyper-V / NIC driver interaction or a deeper architectural problem.

Environment (simplified):

Server: Supermicro H12 platform
AMD EPYC 7002/7003
Dual Broadcom BCM57416 10GBase-T NICs
Windows host running Hyper-V
this has a SSD that has all the VMs on it
this has the spinning disk with the source files

Windows 11 PC
this has the destination files, on a spinning disk

Windows laptop with USB 2.5G adapter

Windows VMs on Hyper-V (on the server)

Network: All links negotiate at 10Gb, no known 1G hops, same destination for all tests
laptop is 2.5G


The transfer test I do is the same source to destination. A network drive on the supermicro server (spinning disk) to a destination on another windows 11 machine also a spinning disk. I just operate the transfer from different machines for the test. I don't understand why the speeds vary so much.

Observed behavior:
Laptop : ~150+ MB/s (the best for some reason)

Supermicro host: ~50–70 MB/s (not bad, but so much slower than laptop, it's the host!)

Windows VM: ~20MB/s (ridiculous slow)


Key confirmations:

No cable or link negotiation issues

No 1G switch hops
Disk performance is not the bottleneck (laptop proves this)
SR-IOV is not available under Windows Hyper-V for the Broadcom NICs (SriovSupport : NotSupported)
Hyper-V external vSwitch is correctly configured
Problem persists even when the Supermicro host is the source of the data

Why this seems wrong:
If the laptop can read/write the same spinning disks at ~150 MB/s over the network, the Supermicro host (with more CPU, more RAM, and 10Gb NICs) should not be limited to ~30–60 MB/s as an SMB source.

I can't figure out why doing the transfer from the VM is so slow.

Thank you.

 

[MrGuvernment]

set up iperf on each system, do a server and then client and use the -P switch to do multiple threads (iperf3) Usually do a thread per physical core in your system, and focus only on the networking stack.

What are the brand of switches? Pro Grade switches would be more managed actual switches, but there are lots of cheap 10Gb switches out there and 10BaseT tends to run hot, which could be throttling things.

You could test doing a direct 10Gb link from the server to the desktop to eliminate the switch as an issue.

 

[louie1961]

I know you said unmanaged switch but are all the end points on the same subnet? If they are on different subnets, the router comes into play and can slow things down.