Performance degradation on diskless systems (tested on pktgen)

Dear community,

I recently tested the performance of the kernel pktgen sample application on my diskless system and compared it to a local installation. The details about my hardware and diskless setup are described below.

Details about hardware (tested on 2 systems):

• AMD Ryzen 5 5600G, 32GB DDR-4 2133 MHz, Intel X520-DA2 10G NIC
• Intel(R) Xeon(R) CPU E5-2620 v3, 16GB DDR-4 2133MHz, Intel X520-DA2 10G NIC

The details about my diskless setup:

• Kernel (6.8) and rootfs reside on NFS host (nfsvers=3)
• Client PXE boots into kernel with root=/dev/nfs for rootfs
• NFS host and client are in the same server rack in a local 1Gbps network

I get the following performance numbers for the pktgen sample application with the following command:
Code:./pktgen_sample03_burst_single_flow.sh -i ens2f0Results:

• AMD Local install: ~ 8 Mpps
• AMD Diskless: ~ 4.5 Mpps
• Intel Local install: ~ 13.8 Mpps
• Intel diskless: ~ 9.3 Mpps

Besides the substantial difference between AMD and Intel processors (probably will be discussed in a new thread), I am interested in the performance degradation on diskless systems. Has anyone experienced similar problems on diskless systems and is it possible to close this gap by different configuration?

Is it possible to cache the rootfs somehow?

Appreciate any help on this matter!

Best regards,
Marco