Ensieea Rizwani

28
ENSIEEA RIZWANI 1 Green Scheduling: A Scheduling Policy for Improving the Energy Efficiency of Fair Scheduler By: Tao Zhu1,2, Chengchun Shu1, Haiyan Yu1

description

Ensieea Rizwani. Green Scheduling: A Scheduling Policy for Improving the Energy Efficiency of Fair Scheduler By : Tao Zhu1,2, Chengchun Shu1, Haiyan Yu1. Motivation - PowerPoint PPT Presentation

Transcript of Ensieea Rizwani

Page 1: Ensieea Rizwani

1

ENSIEEA RIZWANI

Green Scheduling: A Scheduling Policy for Improving the Energy

Efficiency of Fair SchedulerBy:

Tao Zhu1,2, Chengchun Shu1, Haiyan Yu1

Page 2: Ensieea Rizwani

2

Motivation

Reducing energy consumption of data centers is critical to cutting down operational costs as well minimizing its impact to the environment. On one hand, if performance per watt ofserver doesn’t improve, power cost could easily overtake hardware cost . On the other hand, CO2 emissions of globaldata centers will be up to 259 million tons by 2020 , which will accelerate global warming.

Page 3: Ensieea Rizwani

3

Outline

Introductions Overview Power conservation Mechanism Structure Simulation and Measurement Conclusion Related Work

Page 4: Ensieea Rizwani

In the last few years, a lot of effort has been devoted toimprove the energy efficiency of data centers.

Hardware (efficient building block) Reference to last presentation

Software Techniques

At the software level, improve the energy efficiency of MapReduce . MapReduce has been the dominant framework deployed in data center forprocessing large data sets: by 2010, Google processedapproximate 1000 PB of data daily using MapReduce [11];Yahoo had 38000 servers running hadoop(an open-sourceimplementation of MapReduce) in production [12]. So itsenergy efficiency promotion will benefit the data center'senergy consumption reduction.

Page 5: Ensieea Rizwani

5

Data Center Fact

The fact is that servers in data center are non-powerproportional (the energy consumed is not proportional to the work completed). In our experiments, the slave consumes 54.5 W at idle and 87.5 W at peak utilization. For servers, their peak energy efficiency occurs at peak utilization and improves as utilization increases.

Page 6: Ensieea Rizwani

6

Management System of HPC

MapReduce's energy efficiency is closely tied to its scheduler, we find that fair scheduler outperforms FIFO scheduler in energy efficiency when CPU-intensive job and IO-intensive job running simultaneously on the cluster, because fair scheduler achieves better resource utilization by overlapping resource complementary tasks on slaves.

We propose an energy-efficient scheduling policy called green scheduling which relaxes fairness slightly to create asmany opportunities as possible for overlapping resourcecomplementary tasks. The results show that green scheduling can save between 7% and 9% energy consumption of fair scheduler.

Page 7: Ensieea Rizwani

7

We believe the energy saving is the result of the better resource utilization achieved by fair scheduler by overlapping CPU-intensive task and IO-intensive task on slaves. The two types of tasks are complementary : IO-intensive task causes CPU to be idle, letting CPU-intensive task run can increase CPU utilization. In contrast, the effect on I/O performance is opposite: CPU-intensive task leaves IO idle while IO-intensive task can keep IO busy.

Page 8: Ensieea Rizwani

8

Simulation to Validate

we compare our cluster's CPU and IO utilizations under FIFO scheduler and fair scheduler when the CPU-intensive job Pi estimator and the IO-intensive job RandomWriter are running simultaneously on it. Experimental results are demonstrated in Figure 1.

Under FIFO scheduler, CPU utilization fluctuates

between 60% and 100% while IO utilization is below 10% until job Pi estimator finishes. But after job RandomWriter starts, CPU utilization drops dramatically and IO utilization increases significantly. In contrast, fair scheduler keeps both CPU and IO at high utilization over the two jobs' duration. Clearly, fair scheduler leads to better resource utilization than FIFO scheduler.

Page 9: Ensieea Rizwani

9

Scheduler

Page 10: Ensieea Rizwani

10

Pi estimator

Page 11: Ensieea Rizwani

11

Relaxing Fairness

Page 12: Ensieea Rizwani

12

This motivates us to propose an energy efficient scheduling policy called green scheduling: when a slave asks for new task, if the loss of fairness is in permissible range, ourscheduler will choose the job whose resource requirement is the most complementary to the slave's current resource utilization, maximizing the slave's utilization while having a minimal impact on fairness.

Page 13: Ensieea Rizwani

13

Fair Scheduler

Page 14: Ensieea Rizwani

Priority

The default scheduler in Hadoop is FIFO scheduler. All running jobs are sorted and queued according to their priority and submit time. Five priority levels are defined:

very high, High normal low very low

When a slave is ready to accept a new task, FIFO scheduler always picks up the first job in the queue and assigns its required task to the slave.

Note: UB Data center CCR, implements group priority

Page 15: Ensieea Rizwani

15

Starvation

One drawback of FIFO scheduler is its poor response time. Let's look at a concrete example,: Job i at time t duration: 3 days Job j at time t+1 duration: 10 min

Under FIFO scheduler, the response time of jobj is almost 433 times of its job duration. To address this problem, propose fair scheduler which assigns each job a certain share to avoid starving.

Page 16: Ensieea Rizwani

Comparison

Page 17: Ensieea Rizwani

17

Page 18: Ensieea Rizwani

IV. GREEN SCHEDULINGFair scheduler is often more energy efficient than FIFO scheduler when complementary jobs are running simultaneously on the cluster. However, this scheduler itself does not take the slave's and tasks’ resource utilization into account when scheduling jobs. To investigate the opportunity to improve the energy efficiency of fair scheduler, we analyze slots allocation on one slave under FIFO and fair sharing.

Page 19: Ensieea Rizwani

D. Green Scheduling

To achieve better energy efficiency, green scheduling takesinto account slave’s resource utilization and task’s resourceutilization when choosing which job should be scheduled next.

However, this may violate the primary design goal of fairscheduler: fairness. To minimize the impact on fairness, weonly consider slave’s resource utilization as an important factorof choosing job in two scenarios:

both of the two jobs areneedy and neither of them is needy. The justification is that thetwo jobs have got relative fair shares in the two scenarios. Inthe scenario where one job is needy and other one is not, theshares that two jobs have got are absolute unfair. Consequently,relaxing fairness in this scenario will aggravate unfairness.

Page 20: Ensieea Rizwani

Pseudo code for comparing naiive job sorting

Page 21: Ensieea Rizwani

21

A MapReduce job usually consists of a set of maptasks and reduces tasks. For simplicity, we only consider scheduling map tasks to achieve better utilization.

Page 22: Ensieea Rizwani

Green Scheduling algorithm

Page 23: Ensieea Rizwani
Page 24: Ensieea Rizwani
Page 25: Ensieea Rizwani
Page 26: Ensieea Rizwani

26

Conclusion

This paper presented a new scheduling policy called

green scheduling to improve the energy efficiency of fair

scheduler. Knowing the job’s resource requirement and slave’ resource utilization, green scheduling can create as many opportunities as possible for overlapping CPU-intensive task and IO-intensive

task. The key insight it is that overlapping complementary tasks can achieve better energy efficiency as well as utilization. We perform an

evaluation using different workloads that consist of

CPU-intensive job and IO-intensive job, and the results show that fair sharing with green

scheduling can reduce 7%-9% energy consumption over naïve fair sharing.

Page 27: Ensieea Rizwani

27

Related Work

Energy efficiency of Hadoop: Chen et al. [5]

Overlapping CPU-intensive job with IO-intensive job in scheduling: Overlapping CPU-intensive job with IO-intensive job leads to better resource utilization. Wiseman et al. [17]

Page 28: Ensieea Rizwani

28

Thank You