Performance issues on the NFS client with sec=krb* and nconnect>1

Issue

Performance is evaluated with a test script:

#!/bin/sh
echo 3 > /proc/sys/vm/drop_caches
dd if=/dev/zero of=/mnt/nfs/testfile.bin bs=1M count=50000 conv=fsync

The results for this are in the expected range as we are aware of the performance implications which Kerberos has:

NFS Version sec=sys     sec=krb5    sec=krb5i   sec=krb5p
NFS 4.1     871 MB/s    869 MB/s    409 MB/s    349 MB/s
NFS 4.2     870 MB/s    872 MB/s    407 MB/s    262 MB/s

We were then evaluating the impact/improvement of nconnect in this setup. Here the numbers with nconnect=4. With sec=sys we are close to hardware saturation so, all good. But as soon as sec=krb* the best results we measured are lower then without nconnect:

NFS Version sec=sys     sec=krb5    sec=krb5i   sec=krb5p
NFS 4.1     1.1 GB/s    323 MB/s    327 MB/s    422 MB/s
NFS 4.2     1.0 GB/s    297 MB/s    373 MB/s    380 MB/s