110 likes | 121 Views
NTU/CHT 合作案子計劃: 大型資料中心之硬體資源分配管理. Motivation. There are tasks with different characteristics in a datacenter. Different in importance, computation demand, execution period, … etc. Should be completed before some time constraints.
E N D
NTU/CHT合作案子計劃:大型資料中心之硬體資源分配管理NTU/CHT合作案子計劃:大型資料中心之硬體資源分配管理
Motivation • There are tasks with different characteristics in a datacenter. • Different in importance, computation demand, execution period, … etc. • Should be completed before some time constraints. • How to decide the number of servers that execute each tasks in order to meet task deadline is an important issue.
Goal • An resource manage component that can dynamically adjust the number of (dedicated) server for each task. • Decisions are made according to task attributes. • Different decision policies. • Fixed-size heterogeneous server environment.
Models • We model a task Tk as: • R : Remaining workloads/data to be processed. • P : Priority of Tk. • D : Deadline of Tk.
Models(Cont.) • We model each server Sn as a vector: • M : number of tasks • tn,i : throughput of the i-th task on this server • Obtain tn,i by pre-processing • Ex: run some small test cases. • High CPU demand, low processing speedtask=> small throughput
Some Solutions • Priority-based • Task priority is related to its deadline. • Assign one server to each task, and allocate the rest servers to the task with highest priority. • “Earliest Deadline First”.
Some Solutions(Cont.) • Workload-based • For each task, calculate the ratio of its workload to the workload summation of all tasks. • Distributed servers according to the ratio of each task. • “Fair distribution”
Throughput-Aware • The previous policies make decisions according to only task characteristics. • However, in heterogeneous environment, servers can have different capabilities. • We propose a throughput-aware strategy that consider both task and server characteristics.
Another Scenario • We only consider the situation that the total workload of each task is fixed. • However, the total workload of tasks may increase during execution in real life. • Ex: sensor data analyzing task. • We will modify our throughput-aware strategy to deal with such tasks.
Future Plan • Oct. ~ Dec., 2013 • Modify throughput-aware for tasks with increasing workloads. • Policy for tasks with dependencies. • Midterm report • 2014 • Policy for non-dedicated server • A server can host two or more tasks. • Experiments