Download presentation
Presentation is loading. Please wait.
Published byGerard Porter Modified over 8 years ago
1
Web Servers load balancing with adjusted health-check time slot
2
Content Objective Realization Load balancing methods Stateful load balancing techniques Adjusting health check time slot to web traffic Web Server load estimation Simulation of algorithms Conclusions and future work
3
Thesis major goal Increase of performance in web servers load balancing
4
Approaches taken Load balancer work reduction Reduce number of health checks by adjusting health check interval with fluctuations in web traffic Efficient load balancing Weighted estimation of load metrics resulting in overload situations prevention and a very good load balancing
5
Load balancing Stateless load balancing Servers health isn’t considered when distributing traffic Easy implementation Non-optimal load balancing Packet loss Stateful load balancing Incoming traffic is distributed according to the load each server has Good load balancing Huge data processing and transfer
6
“Stateful load balancing” TCP session level load balancing The best possible balancing Maximal overhead introduced Fixed health-check time slot Less overhead due to health checks reduction Operational difficulty in high traffic fluctuations Adjusted health-check time slot load balancing Health-check time slot follows the variations in web traffic Minimal overhead introduced Good protection from servers overloading situation
7
Health-check timeslot adaptation to traffic rate technique Health-check timeslot – Prediction based on traffic rate trend Traffic rate increase expected -> Health-check timeslot decrease Traffic rate decrease expected -> Health-check timeslot increase Low traffic rate for most of the time -> significant reduction of health-checks performed
8
Technique implementation T(i+1)=T(i)*
9
Load balancer traffic distribution
10
Web server load estimation (I) First method : Server load = Exchanged bytes in the last timeslots Traffic exchanged by each server is balanced Some other load metrics are “neglicted”
11
Web server load estimation (II) New method Protect servers from CPU overload Ponder combination of metrics Active sessions number (biggest weights) CPU utilization Load[i]=connectionCount(i)*(1+x) A better estimation of web server load
12
Simulations scenario HTTP v1.0 5 servers cluster Web servers health-check timeslot is 1 second Simulation duration 1000 seconds Random traffic rate < 20 requests/second
13
“Round-Robin” algorithm estimation Not a very good load balancing, especially in short periods of time
14
Estimation of first technique with fixed health-check timeslot Optimal web servers load balancing Very high number of health-checks performed
15
Estimation of new technique with fixed health-check timeslot Good load distribution among servers
16
Estimation of first technique with adjusted health-check timeslot Very good load balancing
17
Estimation of new technique with adjusted health-check timeslot (I) Small traffic rate decrease-> small health-check timeslot increase Only 226 health-checks performed-> around 800 health-checks reduced
18
Estimation of new technique with adjusted health-check timeslot (II) Very good load balancer performance Added complexity due to traffic forecasting
19
Conclusions Optimal web servers load balancing achieved with health- check performed for each incoming HTTP request, but much overhead introduced. We have proposed a load balancing technique with adaptive health-checks timeslot that in turn, reduces heavily heath- checks performed -> less work for load balancer The proposed load estimation method assures a good distribution of traffic while protecting servers from CPU overload NS2 simulations have shown a good load balancing with 3-4 times less health-checks performed Traffic forecasting increases the complexity
20
Future work We see, as a future work, the implementation of this novel technique in real web servers, where additional metrics such as response time, CPU utilization can be used. Load estimation algorithm optimization to reach the soonest possible the load balance among servers of the cluster
21
THANK YOU !
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.