1 / 25

Self-Tuning Energy-Aware Ensemble Model for Server Clusters

Self-Tuning Energy-Aware Ensemble Model for Server Clusters. Team Members : Sanket Dangi (MT2009032) Team Lead Gaurav Kapoor (MT2009059) Deepthi Karnam(MT2009060) Sudha Mani (MT2009073) Celina Madhavan (MT2009075). Architecture. Master. Clients. Server Cluster.

frieda
Download Presentation

Self-Tuning Energy-Aware Ensemble Model for Server Clusters

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Self-Tuning Energy-Aware Ensemble Model for Server Clusters Team Members : SanketDangi (MT2009032) Team Lead GauravKapoor(MT2009059) Deepthi Karnam(MT2009060) SudhaMani(MT2009073) Celina Madhavan(MT2009075)

  2. Architecture Master Clients Server Cluster Reference :Energy Efficient Real Time Heterogeneous Server Clusters by CosminRusu,AlexendreFerrerira,ClaudioScordino,AaronWatson,Rami Melham

  3. Steps to achieve goal according to Development plan • Identifying User processes…………………………..………..……DONE • Identifying CPU idle % and send system to hibernation if it crosses some threshold………………………………………………….…….DONE • Setting up wakeup alarm in BIOS……………………………….…DONE • WakeonLan Configuration…………………………………………..DONE • Energy Consumption Calculation in different modes…………DONE • Port Filtering and Storing Traffic in Database……………….…DONE • Traffic Pattern Recognition Algorithm-kNN Algorithm……….DONE • Hibernation Decision depending on Traffic Pattern and Energy Consumption Calculation………………………………YET TO BE DONE • Scalability of the Solution………………………………YET TO BE DONE

  4. Experimental Setup to Test the Various Modules Developed THIS LAPTOP ACTS AS SERVER 2 AND IS SENT INTO HIBERNATION AND WAKES UP WHEN A CERTAIN NUMBER OF PACKETS ARE FORWARDED TO IT FROM THE MASTER THIS LAPTOP ACTS AS SERVER 1 WHICH COLLECTS THE PACKETS GENERATED BY THE CLIENT AND FORWARDED BY MASTERAND DISPLAYS THE COUNT OF PACKETS GENERATED ROUTER THIS LAPTOP ACTS AS A CLIENT FROM WHICH THE TRAFFIC IS GENERATED THIS LAPTOP ACTS AS THE MASTER AND GETS THE REQUESTS FROM THE CLIENT AND FORWARDS IT TO THE SERVERS

  5. Deliverable 1a,b :Identifying Number of User Processes and Idle CPU Utilization • If there is no client processes running in the server, then the server can be sent to hibernation. • Identifying CPU Idle % and send system to hibernation using ACPI hibernation script if it crosses some threshold limit • Shell Script is used • Commands Used: • ps • grep • wc • mpstat Process (Number of User Process in server) Process (Identifying Idle CPU %) Master

  6. Deliverable 1c :Wakeup Alarm • 1.Depending upon Historical Pattern, no. of active servers would be decided and wakeup alarm will wakeup server(s). • 2.Alarm is set in BIOS. The alarm IRQ (interrupt) is set • 3.Shell Script is used • 4.Files used: • “/sys/class/rtc/rtc0/wakealarm” • “/proc/driver/rtc” At set alarm time the servers wakeup Master Servers

  7. Deliverable 2 :Testing Deliverables on Client Server Architecture Server Master Key generated • To establish Communication between Master and Server : • 1. SSH configuration • 2. PasswordLess Login by Generating RSA Key using ssh-keygen command • Testing of Previous Deliverables on client server architecture

  8. Deliverable 3:Wakeup On LAN Active Server No more requests can be serviced by the active server Master Client1 Requests Client n When active server can’t service anymore requests another server is activated using WOL Server

  9. Deliverable 3:Wakeup On LAN(Continued) • Assumptions: Wired LAN • Package Used: wakeonlan • Command Used: wakeonlan <MAC_Address>

  10. Deliverable 4a :Energy Measurements • Energy Meter Specification: • MACO EM 08 single phase multifunction meter 1mpA 230V • The multifunction meter was used to measure the power consumed by a laptop in watts during normal functioning, sleep mode, hibernation mode, and shutdown. • The energy consumed in watthrs by the laptop while going into hibernation and waking up was also measured. • Based on the measurements we have arrived at the following conclusions • The laptop consumes less than a watt in sleep and hibernation modes. This was measured as 3W and 1W for a desktop PC. • To save energy it is advisable to keep a server in idle mode for a period of 1.4 mins when a laptop is configured as a server rather than send it to hibernation. V A KW KWHr MULTIFUNCTION METER Power socket

  11. Deliverable 4b:Port Filtering And Storing Traffic In Database If any request hits, increment RequestCounter Client 1 80 Wait for Fixed Duration(15 mins) Store RequestCounter in Database 20 Client 2 Ports Reset RequestCounter Clients Master

  12. Deliverable 4b:Port Filtering And Storing In Database(Continued) Packages Used: JDK6,MySql5.1,packETH(traffic generator) Libraries Used: jpcap library Procedure: 1.Open Ethernet device for sniffing 2.Generate Random Traffic through packETH 3.Capture Packets and Count for every fixed duration 4.Store count in database corresponding to that duration

  13. Screenshot :

  14. Deliverable 4c:Pattern Recognition Algorithm-kNN Algorithm • The data mining algorithm uses a pattern recognition technique proposed by Taehyung Kim, Hyoungsoo Kim, Cheol Oh and Bongsoo Son in their paper Traffic Flow Forecasting Based on Pattern Recognition to Overcome Memoryless Property available at http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=04197439 • Since traffic flow is not completely random in nature, there should be some patterns in which the past traffic flow repeats itself. This pattern recognition technique, enables us to consider the past sequences of traffic flow patterns to predict the future state. • Consider a discrete time-series q= {q1, q2, ..., qn} (e.g., 5 min. traffic volumes) where n is the total number of points in the series.This is our available historical data for say N days, which forms an estimation set. • The algorithm works as follows. We define a pattern size ‘l’ and search for the pattern defined by {qn-l,qn-l+1, ….qn} in the historical data for a maximum number of ‘k’ nearest neighbours. • If a match is found at qj we predict the value of qn+1 by taking the average of qj+1 for all the ‘k’ nearest neighbours.

  15. Deliverable 4c:Pattern Recognition Algorithm-kNN Algorithm • The algorithm first takes a minimum value of ‘k’ and ‘l’ and predicts the traffic for each qn+1 in the estimation set based on the available qn values • The root mean squared error (RMSE) between the actual and predicted values, for these choices of neighborhood size and pattern size,is calculated for the entire estimation set. • The process is repeated incrementing ‘l’ upto a maximum size of lmax • The process is again repeated incrementing values of ‘k’ upto a maximum of kmax.(kmax and lmax are chosen by us) • The optimum values of ‘k’ and ‘l’ for the particular network is arrived at by plotting ‘l’ versus RMSE for various ‘k’. • These optimum values of ‘k’ and ‘l’ are then used for predicting the pattern for the (N+1)th day. • The pattern for the (N+1)th day is stored in a database and used by the load balancing algorithm to take decisions on sending the servers to hibernation.

  16. TRAFFIC STATISTICS FROM GERMAN INTERNET EXCHANGEhttp://www.de-cix.net/content/network/Traffic-Statistics.html MATCHING PATTERNS

  17. FLOW CHART START Increment ‘k’ SET k=1,l=1 For each qn match the pattern{qn,….qn-l} with ‘k’nearest matches in historical data (N days) is ‘k’<kmax? yes no Find qn+1 based on the available qn values for each qn in the estimation set Plot graph of ‘l’ versus RMSE and find optimum value of ‘k’ and ‘l’ Calculate RMSE Predict graph for (N+1)th day Increment ‘l’ yes END is l< lmax? no

  18. Optimal Pattern Size as a Function ofNeighborhood Size

  19. Deliverable 4c:Request Forwarding Web Server 1 Client1 HTTP Request coming to Master Server is forwarded to Web Server 1 Master Client 2 Requests Web Server 2 If requestcounter exceeds capacity of webserver 1, WOL signal sent and further requests are forwarded to it Client n

  20. Deliverable 4c:Request Forwarding (Continued) • Iptables used • Start port forwarding by editing /etc/sysctl.conf file • net.ipv4.ip_forward=1 • For port forwarding,use command • iptables –t nat –A PREROUTING –p tcp –d <source_ip> --dport <port_no> -j DNAT –-to <destination_ip>:<dest_port_no>

  21. Work To Be Done Decision Regarding whether to send a system into hibernation or not? While sending server to hibernation, historical traffic pattern and energy consumption reading would be check and decision regarding sending it to hibernation or not would be done. b) Scalability of solution Try the solution for variable and Heterogeneous servers.

  22. Thank You

More Related