implementing-fine-grained-and-pluggable-container-resource-management-leveraging-nri-27d-nri-jiong-ken-yong-zha-jpian-fa-lia-qiang-ren-intel-he-cao-bytedance.pdf

1、Implementing Fine-Grained and Pluggable Container Resource Management Leveraging NRIQiang RenIntelHe CaoByteDanceAgendaKatalyst OverviewPlugin-Based Resource Management Mechanism of KatalystNRI MechanismApplication of NRI in KatalystCommunity 1Katalyst OverviewCapacity Planning ChallengesUsageBuffer

2、WasteThe resource utilization of online services exhibits a tidal pattern,with very low utilization during the night Users tend to over-request resources to ensure service stability,leading to resource wastageColocationOnline services usageBufferThe resource utilization patterns of online services a

3、nd batch jobs are inherently complementary:Online services prioritize CPU and RPC latency Batch jobs prioritize memory and throughputBatch jobs usageNot used by batch jobsReclaimed resourcesKatalyst:Resource Management SystemKatalyst,derived from the“catalyst”in chemical reactions,provides enhanced

4、resource management capabilitiesfor workloads running on Kuberneteshttps:/ KernelMasterNodePlugin-Based Resource Management Mechanism of Katalyst 2Fine-Grained Resource Management Strategies4 Extended QoS Classes Expressing services requirements for resource quality Naming based on CPU as the primar

5、y resource dimensionMore QoS Enhancements NUMA binding NUMA exclusive Network class QoS ClassesAttributesSuitable for workload typesRelationship with K8s QoSdedicated_cores Dedicated CPU cores,not shared with other workloads Supports binding to NUMA nodes for improved performanceExtremely latency-se

6、nsitive workloads,such as ads,search,and recommendationGuaranteedshared_cores Shared CPU pool Supports further dividing CPU pools based on business types Also supports NUMA bindingWorkloads that can tolerate a certain degree of CPU throttling or interference,such as microservicesGuaranteed/Burstable