Drivers & Software

[amdmax]

On my AMD Threadripper PRO 3955wx  system I have 4 non OS additional 1TB NVMe's PCIe 4*gen4.
When I measure sequential read and write I get:

Individual

Performance 6.7 GB/s read and 5 GB/s write.

2 drives MS Windows 10 striped volume
Approx. 12 GB/s read and 10GB/s write.
So far so good.

4 drives MS Windows 10 striped volume
Using 4 NVMe drives I get no more than

approx 10 GB/s seq read and 10 GB/s seq write.

The drives are on there own PCI Express root complex and all running at 4 PCIe lanes *16GT/s

Using taskmanager - performance, I can see that the individual drive performance in this striped volume is approx 2.5 GB/s r/w and is way below the max individual performance of 6.7 GB/s read and 5 GB/s write the drives are capable of.

I understand there is some overhead to take into account for the stiped volume but I expected results in the range of 17-20 GB/s.

(The same drives in an AMD RaidXpert2 RAID 0 array perform 12GB/s read 17GB/s write. That read is significantly lower than write also raises some questions)

Does anyone have an idea what could be the bottleneck that is holding back the 4 drives striped volume from achieving 17-20GB/s

[amdmax]

I did some testing and it appears to me that the use of 1 thread is the bottleneck.

When i increase the number of threats in the benchmark, I get the results you would expect (at least for read)

On a Windows 10 stiped volume using 4 gen 4 NVMe's I got:

10GB/s seq read - 10 GB/s seq write using 1 Thread

18GB/s seq read - 9 GB/s seq write using 2 Threads

26GB/s seq read - 9 GB/s seq write using 4 Threads

26GB/s comes down to 6.5GB/s read per NVMe. That is close to the max NVMe performance.

https://docs.microsoft.com/en-us/answers/questions/386725/striped-volume-bottleneck.html?childToView...