1. LOAD BALANCING IN A CENTRALIZED DISTRIBUTED SYSTEM BY ANILA JAGANNATHAM ELENA HARRIS

2. OUTLINE PROBLEM STATEMENT ARCHITECTURE LOAD BALANCING METHODS EXPERIMENTAL RESULTS CONCLUSION

3. PROBLEM STATEMENT To study different load balancing methods in order to improve CPU utilization and overall execution time of requests in a centralized distributed system with an environment that requires extensive calculations. To study different load balancing methods in order to improve CPU utilization and overall execution time of requests in a centralized distributed system with an environment that requires extensive calculations.

4. ARCHITECTURE MASTER – SLAVE ARCHITECHTURE CENTRALIZED CLIENT ACTS AS THE MASTER MACHINE SERVERS ACT AS SLAVES SOCKETS ARE USED FOR COMMUNICATION SERVER CLIENT

5. LOAD BALANCING APPROACHES RANDOM LOAD BALANCING Requests are routed to servers at random. Used for homogeneous systems. Requests are routed to servers at random. Used for homogeneous systems. ROUND ROBIN LOAD BALANCING Simple and no overhead Simple and no overhead Load may not be evenly balanced in the system Load may not be evenly balanced in the system CPU UTILIZATION BASED LOAD BALANCING Client queries CPU utilization of each of the servers and chooses the least loaded server to assign incoming user request. Client queries CPU utilization of each of the servers and chooses the least loaded server to assign incoming user request.

6. CPU UTILIZATION METRIC Client requests statistics from /proc/stat file maintained by each server to estimate current CPU utilization. Statistics from /proc/stat: – Total time spent by the CPU on all processes – Total time spent by the CPU on all processes – Total number of processes running. – Total number of processes running. Average time per process,  T /  P is used as metric  T – Difference in CPU time at two time points  T – Difference in CPU time at two time points  P – Difference in number of processes at two time points  P – Difference in number of processes at two time points The larger  T /  P, the less loaded a CPU is.

7. USER APPLICATIONS Four applications are chosen to simulate user requests. Run time of these applications is 8 min, 3 min, 30 sec and 3 sec respectively. 8 min, 3 min, 30 sec and 3 sec respectively. Client has dual functionality - Generates user requests - Generates user requests - Performs load balancing.

8. EXPERIMENTAL RESULTS EXPERIMENTAL RESULTS  T/  P TIME PERIODS RANDOM LOAD BALANCING RANDOM LOAD BALANCING

9.  T/  P TIME PERIODS

11. TIME POINTS  T/  P

12. EXPERIMENT -II In this experiment, 1000 tasks were distributed. From the graph, average time/process using the proposed method changes less drastically.

13. CONCLUSION LOAD BALANCING USING CURRENT CPU AVERAGE TIME/PROCESS RESULTS IN BETTER OVERALL CPU UTILIZATION AS COMPARED TO RANDOM OR ROUND ROBIN METHODS. LOAD BALANCING USING CURRENT CPU AVERAGE TIME/PROCESS RESULTS IN BETTER OVERALL CPU UTILIZATION AS COMPARED TO RANDOM OR ROUND ROBIN METHODS.