Trending links: [0] [1].
Other testbeds do not show such structure. The structure is visible also in 1c 1tnl test, so worker placement is not an issue. When one test has good/bad performance, other test cases in the same robot run also have good/bad performance (in bisect job, performance between robot invocations in the same job fluctuates).
Looking at runtime stats [2] [3] it shows the difference in processing, but not clear where exactly (perfmon needs newer Linux kernel to work on ARM).

The task is to find the cause, so CSIT can prepare more consistent environment or a bug in VPP can be fixed.

[0] https://csit.fd.io/trending/#eNrdlU1uwyAQhU_jbKqRDMFh1UUT3yMieJqggj0C0sY9fXESibhSmy4q1S4LfvRGvDefkAix87gNaB-Lal3IdcGladJULJ8e0vJKBMsWlI0gyj1yYicrWQkNNS-gfU-xg5XYAdOA8WBIGAqoWWxteIN03KmAQJ01ugeFgVervXbgvB8s-GawaI5x5JcVOvRZ-SpFLlceVa6_hMtqxHDj8_Ow-YZnrxwG8475mtR61nUimSWmx96xpxv1SkDW54pfwi7KOXG_l3ay4I0zp88Pvkxj3I5p42Rf_J2035Ifup8a-lmR_y_gxWyoi79EXtWLtvPu_L1yOWyVvZpU9QfvFSfo
[1] https://s3-docs.fd.io/csit/master/trending/trending/ipsec-3n-alt-xl710.html#b-ipsec-ip4base-policy-aes256gcm-sw
[2] https://s3-logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-mrr-daily-master-3n-alt/149/log.html.gz#s1-s1-s1-s1-s14-t1-k2-k13-k8-k14-k1-k1-k1-k1
[3] https://s3-logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-mrr-daily-master-3n-alt/150/log.html.gz#s1-s1-s1-s1-s14-t1-k2-k13-k8-k14-k1-k1-k1-k1