Impact of server dynamic allocation on the response time for energy-efficient virtualized web clusters energy-efficient virtualized web clusters Carlos Oliveira, Vinicius Petrucci, Orlando Loques {cjunior, vpetrucci, loques} @ic.uff.br Universidade Federal Fuminense (UFF) Niteroi, Rio de Janeiro, Brazil 1
Agenda • Introduction • Our work • Virtualized cluster architecture • Experiments • Experiments – Live migration – Replication • Conclusion and ongoing work 2
Introduction • Energy consumption – Current context – Data center, web servers – High demands for energy – Global warming • What is server virtualization? • What is server virtualization? – Server associated to VMs, not to physical machines • Why adopt server virtualization? – On-demand allocation – Smaller number of servers – Increase resource utilization – Reduce the use of computer resources and the associated (electrical) power consumption 3
Our work • Issues – Service disruption on the course of migration – How to allocate VMs in a cluster • How to solve or reduce them? – We aim to investigate replication to reduce this disruption • This work aims to – Analyze the disruptive impact during dynamic server alocation – Compare migration’s (cold and live) disruption to replication disruption • We performed a set of experiments to achieve our goal 4
Virtualized cluster architecture The VM Manager is implemented using the OpenNebula toolkit which enables the management of the VMs in the cluster, such as deployment and monitoring deployment and monitoring 5
Virtualized cluster architecture The VM Manager is implemented using the OpenNebula toolkit which enables the management of the VMs in the cluster, such as deployment and monitoring deployment and monitoring The Load Balancer implements a weighted round-robin scheduler strategy provided by Apache 6
Virtualized cluster architecture The VM Manager is implemented using the OpenNebula toolkit which enables the management of the VMs in the cluster, such as deployment and monitoring deployment and monitoring The Load Balancer implements a weighted round-robin scheduler strategy provided by Apache The Optimizer is designed to monitor and configure the virtualized cluster 7
Virtualized cluster architecture We used Xen as the virtual machine hypervisor and Apache servers for running the web applications 8
Response time monitoring • When CPU utilization of an app is low, the response time is also low • When utilization is high, the response time abruptly goes up 9
Live Migration • Goal: To move a VM to another physical machine – But maintaining an acceptable QoS (app response time) during this movement • The disruptions, observed during dynamic changes in live migration, last a short time and are basically unavoidable because cache contents are not migrated 10
Alternative: Replication • Idea: To create and deploy a new VM replica for the app on the destination server • When the new replica is ready (already booted) for processing the client requests, we start redirecting the requests to this the client requests, we start redirecting the requests to this new replica • The goal is to evaluate the response time impact compared to the live migration 11
Replication execution • Replication shows improvements compared to the live migration We noticed that if all the current requests were abruptly redirected to the new VM • replica it would take a long time to get both, throughput and response time, stable 12
Live migration vs Replication Live Migration Replication 13
Live migration vs Replication Minor disruption compared to migration 22 ms 300 ms Live Migration Replication 14
Conclusion • Our goal was to carry out experiments to evaluate the performance impact in terms of response time and throughput of applications during the course of VM migration and replication • Our results showed that by using replication we can reduce performance disruptions incurred during migration • This evaluation work will help us to implement our overall optimization approach for virtualized clusters (Petrucci et al. 2010) 15
Ongoing work • Include energy savings and response time control in our experiments • Experiments with state-aware applications (database tier) – RuBis (benchmark application – simulates an eBay site) – RuBis (benchmark application – simulates an eBay site) • Improvements in the dynamic allocation (replication) scheme – Keep the application on the source server too for load balancing (and fault tolerance) proposes – Identify what part of the application workload would be allocated both in the source and target physical machines 16
Tempo Lab UFF Thank you! Our webpage: www.tempo.uff.br The contemporary Art museum in Niteroi, Rio de Janeiro 17
Experiments • We have measured the maximum number of requests per second that our physical machines can handle • We used State-less requests 18
Response time measurement • The response time is defined by the difference between the time a response is generated and the moment the server has accepted the associated request • To obtain the response time for the web applications we have implemented a new Apache module that collects the time implemented a new Apache module that collects the time information between these two moments 19
Cold Migration • Cannot be used in Real Time Systems because it stops • We have made an experiment in this scenario only to show it only to show it 20
Replication ( Details ) • (1) booting a new VM replica • (2) redirecting the requests to the new replica • The time needed to boot a VM is in between 25 and 40 seconds seconds • We may also boot the new replica on the target machine a few seconds earlier to have it running and ready at the moment necessary for using the replicated VM • In this experiment, we redirected 10% of the requests each time, until 100% of the requests were redirected to the VM replica 21
Testbed Processor Memory Camburi Intel Pentium 4 2.80GHz 1GB RAM Cumulus Intel Pentium 4 2.80GHz 1GB RAM Henry AMD 3GB RAM Athlon 64 3500+ Maxwell Intel Core i5 2.67GHz 8GB RAM Edison Intel Core i7 CPU 8GB RAM 22 2.67GHz
Server virtualization • Server virtualization allows for on-demand allocation (using either migration or replication) of virtual machines (VMs), which run the web applications and services, to physical servers • We have measured and analyzed the disruptive impact on the • We have measured and analyzed the disruptive impact on the QoS (quality-of-service) provided by the applications, in terms of server-side response time and throughput, during dynamic allocation of virtual machines in a server cluster • We have used Xen as the virtual machine manager and Apache servers for running the web applications 23
Xen hypervisor • The Xen hypervisor offers two kinds of migration: cold and live migration. The difference between them is that on cold migration the VM stops running during migration • The live migration stages are listed below: • Cold migration doesn’t have stage 2 presented in the figure 24
Recommend
More recommend
