PassMark USB3.0 Loopback benchmark test under the Linux

[karry]

I want to use the USB3.0 Loopback to test the performance of my computer's USB3.0 interface. Since you didn't provide the Passmark USB3 testing software for the Linux system, I referred to the API provided on the official website and wrote a simple test program. The test results are as follows. I would like you to help me confirm whether this performance is a little lower compared to the test results under Windows?（USB 3.0 is directly derived from the PCH and does not go through a conversion chip.）
API：PassMark USB3 Loopback plug Application Programming Interface - API

================================================== ==============================
Sweep Results
================================================== ==============================
Direction burst threads URB Avg (Mbps) Peak (Mbps) of ceiling
----------------------------------------------------------------------------
READ 12 4 64K 3145.50 3145.66 78.6%
READ 12 8 64K 3145.49 3145.67 78.6%
READ 12 4 256K 3145.46 3146.71 78.6%
READ 12 2 64K 3145.45 3145.65 78.6%
READ 12 8 256K 3145.38 3146.70 78.6%
READ 12 2 256K 3145.38 3146.69 78.6%
READ 12 2 1M 3144.69 3150.88 78.6%
READ 12 8 1M 3144.68 3150.88 78.6%
READ 12 4 1M 3144.63 3150.89 78.6%
READ 12 1 1M 2851.59 2865.96 71.3%
READ 12 1 256K 2812.68 2819.86 70.3%
READ 12 1 64K 2514.64 2517.15 62.9%
WRITE 12 2 64K 3353.14 3353.04 83.8%
WRITE 12 4 64K 3353.10 3354.01 83.8%
WRITE 12 8 64K 3352.92 3353.06 83.8%
WRITE 12 4 256K 3293.92 3297.55 82.3%
WRITE 12 8 256K 3293.71 3297.33 82.3%
WRITE 12 2 256K 3291.01 3293.35 82.3%
WRITE 12 4 1M 3275.20 3284.98 81.9%
WRITE 12 2 1M 3275.08 3284.95 81.9%
WRITE 12 8 1M 3274.91 3284.97 81.9%
WRITE 12 1 1M 3042.96 3050.33 76.1%
WRITE 12 1 256K 2943.67 2945.45 73.6%
WRITE 12 1 64K 2652.94 2661.61 66.3%

================================================== ==============================
Overall Best Configuration
================================================== ==============================
READ : burst=12 threads=4 URB=64K avg 3145.50 Mbps (78.6% of the 4 Gbps ceiling)
WRITE: burst=12 threads=2 URB=64K avg 3353.14 Mbps (83.8% of the 4 Gbps ceiling)

[David (PassMark)]

There are some example speeds on the FAQ page
https://www.passmark.com/support/usb3loopback_faq.php

On different PC hardware we got between 3472 and 2401 Mb/Sec with Windows and the Window device drivers. (actual useful data, ignoring the protocol overhead).

But these are all old machines now. I would expect more modern PC hardware to be more consistently at the upper end of this range. (or even slightly higher)

[karry]

I had already seen these data before. The data we have for Windows is also consistent with the data you provided. Now, we would like to confirm whether there will be significant differences in the test results when the same hardware platform is tested under Windows versus Linux?
For example, the test result of the Intel 11th generation platform under Windows is 3400 Mb/s, but under Linux, it only reaches 3100 Mb/s.

[David (PassMark)]

We did do some testing like this years ago, but I don't have the numbers now.

But there is totally a difference in the performance of Windows / Linux device drivers.

There are also significant differences based on the host controller interface standard used. From oldest to newest, there are UHCI, OHCI, EHCI, xHCI. Then within xHCI there are multiple versions from 0.9 to 2.0. The 1.0 & 1.1 standards were common, but didn't perform as well as 2.0 does. So if the device driver only uses the 1.1 features on Linux, but the 2.0 features on Windows, this can also cause differences. We saw big performance differences at higher speeds (20Gbit/s) when difference interfaces were used.

[karry]

Under the condition that the same hardware environment remains unchanged (the same Intel hardware platform, the same USB interface, the same USB 3.0 Loopback), only the test system (Windows10 and Ubuntu 22.04) was modified. The obtained test results are as follows. It seems that the results are quite close.

[David (PassMark)]

Given the overheads in the USB protocol, this is a reasonable result.
Our tools report useful data transferred. If you looked on the wire then you could find more bits than this (Packet headers, checksums, polling, etc..). And there is the 8b/10b encoding which is a 20% loss.