CHT Project Progress Report 11/18 Simon

CHT Project Develop a resource management scheduling algorithm for CHT datacenter. Two types of jobs, interactive/latency- sensitive and batch/computation-intensive. Minimize SLA violation with limited resources. # of servers Based on Red Hat Openshift

Current Plan Two components Scheduler Container Scaler Deploy container/pod to server Inside Kubernetes Container Scaler Decide the number of container/pod for each service. Output the results in JSON As an input of the scheduler

Container Scaler Decide the number of container/pod for each service according to the monitoring data. Ex: add one container if the average CPU utilization is higher than 85%. Add/adjust rules to make better decisions.

Adjust According to Monitor If the QoS of an application can be measured from the monitor. Add container according to the QoS. Ex: if the average response time is over 80% of the SLA, add new container. If not, need to “derive” from the monitoring data. Ex: if the CPU utilization over 75% and the network I/O is larger than 800Mbps, add new container.

根據硬體資源使用率，判斷是否可減少容器數量 量測資料 資料包含效能數據 比較目前效能及 SLA 效能 > SLA*0.9 利用硬體資源使用率，推測目前效能 根據硬體資源使用率，推測增加容器數量 開始 以 json 格式寫入輸出檔 結束 紀錄服務目前容器數量 處理完所有服務 更新容器數量 根據硬體資源使用率，判斷是否可減少容器數量 F T

根據硬體資源使用率，判斷是否可減少容器數量 量測資料 資料包含效能數據 比較目前效能及 SLA 效能 > SLA*0.9 利用硬體資源使用率，推測目前效能 根據硬體資源使用率，推測增加容器數量 開始 以 json 格式寫入輸出檔 結束 紀錄服務目前容器數量 處理完所有服務 更新容器數量 根據硬體資源使用率，判斷是否可減少容器數量 F T Utilization of key resource, regression, SVM, …etc Pareto_frontier https://en.wikipedia.org/wiki/Pareto_efficiency#Pareto_frontier

Improving the Rules Apply data analysis techniques to improve the derivation. SVM(Support Vector Machine) Find the hyper-plane that separate the data points into “safe” and “violation”. Add containers for those applications being classified into “violation”. Find another hyper-plane that separate the data points into “safe” and “over-provision”. Remove containers for those applications being classified into “over-provision”.

Implementation Working on container scaler with basic rules. No data analysis techniques are involved in current state. Will coordinate the data format with Prof. Lin’s group. JSON format.

JSON Format

Next Keep working. Progress meeting with CHT Midterm report Early December Midterm report 12/12