ruo91

[root@bastion yamls]# oc get pod
NAME                                           READY   STATUS        RESTARTS   AGE
gemma-2-9b-it-fp8-predictor-75fb9658bb-725mb   1/2     Running       0          15s

[root@bastion yamls]# oc logs -f gemma-2-9b-it-fp8-predictor-75fb9658bb-725mb
Defaulted container "kserve-container" out of: kserve-container, kube-rbac-proxy
[preflight] tcmalloc: /usr/lib64/libtcmalloc_minimal.so.4
[preflight] cpu_limit=8 world=1 (tp=1 pp=1 dp=1) api_servers=1 reserved=1
[preflight] OMP_NUM_THREADS=7 (per worker process, total <= 7x1+1)

ruo91

[preflight] tcmalloc: not available, using glibc malloc
[preflight] OMP_NUM_THREADS=8
[preflight] /etc/passwd is read-only - skipped (HOME=/tmp is set, expected)
(APIServer pid=1) INFO 08-17 11:30:31 [api_utils.py:345] 
(APIServer pid=1) INFO 08-17 11:30:31 [api_utils.py:345]        █     █     █▄   ▄█
(APIServer pid=1) INFO 08-17 11:30:31 [api_utils.py:345]  ▄▄ ▄█ █     █     █ ▀▄▀ █  version 0.27.1
(APIServer pid=1) INFO 08-17 11:30:31 [api_utils.py:345]   █▄█▀ █     █     █     █  model   /mnt/models
(APIServer pid=1) INFO 08-17 11:30:31 [api_utils.py:345]    ▀▀  ▀▀▀▀▀ ▀▀▀▀▀ ▀     ▀
(APIServer pid=1) INFO 08-17 11:30:31 [api_utils.py:345] 

ruo91

I0815 09:44:11.669608       1 qat_plugin.go:44] QAT device plugin started
I0815 09:44:11.671088       1 dpdkdrv.go:629] Device 0000:e8:00.1 with cy capabilities found (Healthy)
I0815 09:44:11.671450       1 dpdkdrv.go:629] Device 0000:e8:00.2 with cy capabilities found (Healthy)
I0815 09:44:11.671711       1 dpdkdrv.go:629] Device 0000:e8:01.3 with cy capabilities found (Healthy)
I0815 09:44:11.671959       1 dpdkdrv.go:629] Device 0000:e8:01.4 with cy capabilities found (Healthy)
I0815 09:44:11.672202       1 dpdkdrv.go:629] Device 0000:e8:01.5 with cy capabilities found (Healthy)
I0815 09:44:11.672458       1 dpdkdrv.go:629] Device 0000:e8:01.6 with cy capabilities found (Healthy)
I0815 09:44:11.672710       1 dpdkdrv.go:629] Device 0000:e8:01.7 with cy capabilities found (Healthy)
I0815 09:44:11.672976       1 dpdkdrv.go:629] Device 0000:e8:02.0 with cy capabilities found (Healthy)

ruo91

The istio-proxy pod gets stuck in the init phase when using OpenShift AI 3.4.2 in an environment running OpenShift 4.20.8 and Service Mesh 3.2.1

OpenShift 4.20.8 버전에서 Service Mesh 3.2.1을 사용하고 있고,
OpenShift AI 3.4.2 버전을 사용하는 경우 기존 또는 신규 Pod 배포 시 istio-proxy가 init 단계에서 timeout이 발생하는 상황이다.

이는 Knowledge base 내용 처럼 OpenShift 4.19, 4.20.0-4.20.30 또는 4.21.0-4.21.21를 사용하는 버전에서 발생되며,
해당 버전에서는 Service Mesh 3.4.x 버전으로 강제로 업그레이드 해서는 안된다.

해결 방법은 OpenShift 4.20, 4.21, 4.22 버전의 마이너 최신 버전으로 업그레이드 후
Service Mesh Operator를 3.4.x 버전 이상으로 업그레이드 한다.

ruo91

2026/08/14 01:42:10 INFO GOMEMLIMIT is already set, skipping package=github.com/KimMachineGun/automemlimit/memlimit GOMEMLIMIT=1073741824
2026-08-14T01:42:10.158884Z	info	FLAG: --concurrency="0"
2026-08-14T01:42:10.158909Z	info	FLAG: --domain="nginx.svc.cluster.local"
2026-08-14T01:42:10.158915Z	info	FLAG: --help="false"
2026-08-14T01:42:10.158918Z	info	FLAG: --log_as_json="false"
2026-08-14T01:42:10.158924Z	info	FLAG: --log_caller=""
2026-08-14T01:42:10.158926Z	info	FLAG: --log_output_level="default:info"
2026-08-14T01:42:10.158929Z	info	FLAG: --log_stacktrace_level="default:none"
2026-08-14T01:42:10.158939Z	info	FLAG: --log_target="[stdout]"
2026-08-14T01:42:10.158943Z	info	FLAG: --meshConfig="./etc/istio/config/mesh"

ruo91

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: block-istiod
  namespace: istio-system
spec:
  podSelector:
    matchLabels:
      app: istiod
  policyTypes:

ruo91

00. Arista 7150S-64 (Data Sheet)

https://www.arista.com/assets/data/pdf/Datasheets/7150S_Datasheet.pdf

1. ZTP(Zero Touch Provisioning) 비활성화

arista> zerotouch cancel

2. MGMT IP 설정

arista> enable

ruo91

OpenShift 4.18 환경에서 오픈소스 HAProxy를 기반으로 사용자 정의 컨테이너 이미지를 구성하여
CPU 최적화 및 성능 향상을 실현한 사례를 정리한 내용이다.

1. HAProxy 성능 최적화를 위한 CPU L1 Cache 기반 설정 전략

HAProxy를 순수 TCP Pass-through 모드로 사용하며, 데이터 패킷 크기를 최신 CPU의 L1 Cache 사이즈에 맞춤으로써
Memory Access Latency와 Context Switching을 Zero에 수렴하게 만드는 전략이다.

HAProxy의 주요 버퍼 설정 값들을 16KB(16,384 bytes) 수준으로 맞추면, L1 Cache 구조에 최적화되어 처리 성능이 크게 향상된다.

1.1. CPU 최적화: L1 캐시 기반 최적화

ruo91

1. ETCD 인증서 저장

1.1. Server CA Bundle

$ ssh core@master01.ocp4.local "sudo cat /etc/kubernetes/static-pod-resources/etcd-certs/configmaps/etcd-all-bundles/server-ca-bundle.crt" > certs/server-ca-bundle.crt

1.2. ETCD Serving Cert / Key

$ ssh core@master01.ocp4.local "sudo cat /etc/kubernetes/static-pod-resources/etcd-certs/secrets/etcd-all-certs/etcd-serving-master01.ocp4.local.crt" > certs/etcd-serving.crt
$ ssh core@master01.ocp4.local "sudo cat /etc/kubernetes/static-pod-resources/etcd-certs/secrets/etcd-all-certs/etcd-serving-master01.ocp4.local.key" > certs/etcd-serving.key

ruo91

apiVersion: networking.istio.io/v1alpha3
kind: EnvoyFilter
metadata:
  name: proxy-protocol
  namespace: istio-system
spec:
  workloadSelector:
    labels:
      istio: ingressgateway
  configPatches: