Impact on rls2606 results is in failure probability od NDRPDR tests, for example this [0] figure shows IMIX is significantly more sensitive than 1518B, and larger number of cores are somewhat more sensitive. In stats [1], the typical counter was bad-hmac in crypto-deq. Looking at trending [2] there is a notable regression in March, although later stats [3] show engine-err in crypto-deq and crypto-enq instead, I guess due to a refactor instead. CX6 is also tested on 3n-icx where trending [4] shows a similar regression, but 3na-spr with CX7 [5] does not.
[0] 3n-alt CX6 release
[1] https://logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-report-coverage-2606-3n-alt/28158576877/log.html.gz#s1-s1-s1-s1-s12-t3-k3-k7-k1-k1-k1-k8-k14-k1-k1-k1-k1
[2] 3n-alt CX6 trending
[3] https://logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-ndrpdr-weekly-master-3n-alt/29623724752/log.html.gz#s1-s1-s1-s1-s6-t4-k3-k7-k1-k1-k1-k8-k14-k1-k1-k1-k1
[4] 3n-icx CX6 trending
[5] 3na-spr CX7 trending
Impact on rls2606 results is in failure probability od NDRPDR tests, for example this [0] figure shows IMIX is significantly more sensitive than 1518B, and larger number of cores are somewhat more sensitive. In stats [1], the typical counter was
bad-hmacincrypto-deq. Looking at trending [2] there is a notable regression in March, although later stats [3] showengine-errincrypto-deqandcrypto-enqinstead, I guess due to a refactor instead. CX6 is also tested on 3n-icx where trending [4] shows a similar regression, but 3na-spr with CX7 [5] does not.[0] 3n-alt CX6 release
[1] https://logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-report-coverage-2606-3n-alt/28158576877/log.html.gz#s1-s1-s1-s1-s12-t3-k3-k7-k1-k1-k1-k8-k14-k1-k1-k1-k1
[2] 3n-alt CX6 trending
[3] https://logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-ndrpdr-weekly-master-3n-alt/29623724752/log.html.gz#s1-s1-s1-s1-s6-t4-k3-k7-k1-k1-k1-k8-k14-k1-k1-k1-k1
[4] 3n-icx CX6 trending
[5] 3na-spr CX7 trending