ANALYZING ENERGY CONSUMPTION OF ELASTIC HPC APPLICATIONS IN THE - PowerPoint PPT Presentation

ANALYZING ENERGY CONSUMPTION OF ELASTIC HPC APPLICATIONS IN THE CLOUD Gustavo Rostirolla, Vinicius Facco Rodrigues, Rodrigo da Rosa Righi, and Cristiano André da Costa Contact: grostirolla1@gmail.com

SUMMARY • Introduction • Energy Consumption • Elastic Energy Consumption Model • Methodology • Results Analysis • Conclusion

INTRODUCTION • Elasticity can be a double-edged sword involving performance and energy consumption; Resources Time Energy A user can achieve a good performance considering the time to execute its application, but using a large amount of resources, resulting in a waste of energy.

INTRODUCTION • Measuring performance and energy consumption accurately are not easy tasks; Administrators can suffer with resource sharing among the users, besides a waste on energy consumption. Can be measured by the server. But how much power each user is consuming in a moment?

MODEL • Deploying energy sensors or wattmeters can be costly if not done at the time the whole infrastructure is set up ( besides being time consuming as the infrastructure scales up); • We present an elastic energy consumption model which gives data about energy when executing HPC applications in elastic-based cloud environments.

MODEL The proposed model extracts energy consumption data from a malleable infrastructure of resources , enabling relationships among energy consumption, resource consumption and performance. 1. Collect samples of resource usage, and the machine energy consumption using a smart power meter; 2. Perform regression methods to generate the energy model; 3. Test the model in a different set of data collected from another homogeneous machine.

MODEL • We obtained a mean and median accuracy of 97.15% and 97.72%, respectively. Measured Power Predicted Power 80 70 Instantaneous Power Consumption (W) 60 50 40 30 20 10 0 0 1000 2000 3000 4000 5000 6000 7000 8000 Sample

MODEL 2 VMs 4 VMs 2 VMs 4 VMs 8 VMs • Considers the cloud elasticity; • Can measure shared resources power consumption; • Scales as the infrastructure grows up (homogeneous).

MODEL f ( CPU, Memory ) = α + β × CPU + δ × Memory α represents an IDLE power consumption. β and δ represent the variable power consumption determined by the amount of resources that is used in the moment MC ( m, i ) = f ( CPU ( m, i ) , Memory ( m, i )) Energy consumption of machine m according to the logged CPU and memory in a instant i

MODEL Machines ⇢ x = 0 if machine i is not active in the instant t ; X ETC ( t ) = MC ( i, t ) × x x = 1 if machine i is active in the instant t. i =0 ETC calculates the total power consumption of all machines allocated in an instant t , taking into account elasticity using the previous equation MC t � 0 ≤ t ≤ TotalApplicationTime X TC ( t ) = ETC ( i ) i =0 TC calculates the total energy consumption using the previous equation ETC in a determine time interval.

MODEL AppT ime ⇢ y = 0 if in instant i the total of active machines 6 = z ; X NEC ( z ) = ETC ( i ) ⇥ y y = 1 if in instant i the total of active machines = z. i =0 Presents the application power consumption when employing an specific amount of nodes represented by z . It allows the analysis of how much energy has been spent using a specific amount of nodes during the application execution 500 400 Energy (kJ) 300 2 VMs 4 VMs 6 VMs 8 VMs 10 VMs 200 100 0 30 50

METHODOLOGY SSH Connection and Cloud-supported M S S S S Application S S S S Application Program Interface (API) Virtual VM VM 2c-1 VM 0 VM c-1 VM c VM 2c VM 3c-1 VM n-1 VM (m-1)c Machines Master AutoElastic Area Manager for Data Computational Share Resources Cloud Front-End Node 0 Node 1 Node 2 Node m-1 M Master process S Slave process Interconnection Network AutoElastic Cloud (b) • 6 Nodes - 1 FrontEnd 5 Computing • 2.9 GHz Dual Core • 4 GB RAM • 100 Mbps

METHODOLOGY Constant Ascending Descending Wave Number of subintervals 10 9 [load(x)] x 100000 8 7 6 5 4 3 2 1 0 1 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 6000 6500 7000 7500 8000 8500 9000 9500 10000 Iteration Load Pattern Constant Ascending Descending Wave 80 Power Consumption (W) 70 60 50 40 30 20 10 0 0 300 600 900 1200 1500 1800 2100 2400 2700 3000 3300 3600 3900 4200 Time (seconds) Power Consumption of a Single Node Varying the Load Pattern

RESULTS ANALYSIS 500 500 400 400 Energy (kJ) Energy (kJ) 300 300 Constant Wave 200 200 100 100 0 0 30 50 30 50 30 50 30 50 70 70 90 90 70 70 90 90 Thresholds (lower / upper) Thresholds (lower / upper) 2 VMs 4 VMs 6 VMs 8 VMs 10 VMs 500 500 400 400 Energy (kJ) Energy (kJ) 300 300 200 200 Ascending Descending 100 100 0 0 30 50 30 50 30 50 30 50 70 70 90 90 70 70 90 90 Thresholds (lower / upper) Thresholds (lower / upper)

��� ��� ��� � � �� ��������������������� ��� ��� ��� ���� ���� ���� ���� ���� ���� ���� ��� ��� ���� ����� ���� ��� �������������� ��� RESULTS ANALYSIS (i) Host allocation; (ii) Virtual machines booting; (iii) Processing stop to incorporate new resources; (iv)Host Deallocation.

CONCLUSION • A model estimates energy consumption based on CPU and Memory traces with mean and median accuracy of 97.15% and 97.72% ; • Equations to analyze HPC application power consumption on elastic cloud environment; • Best energy saving with a threshold close to 90% ; • Worst energy saving with an upper threshold equal to 70% , but it reaches the best performance rates.

Contact: grostirolla1@gmail.com • Extend the proposed model to include heterogeneous machines;