![Dipartimento di Elettronica e Informazione
26Micro-benchmarking Web Service Application Experiments
§ Number of VMs varied between 2 and 4
§ For system identification purposes request arrival rates vary stepwise every
1 minute
§ Each request consumes si
k CPU time varied between 0.06 s and 1.1 s.
§ 1,440 intervals (24 hours)
§ Parametrization [sk
i ρk
i]](https://image.slidesharecdn.com/parvis-130713122034-phpapp01/85/D-Ardagna-PARVIS-Performance-mAnagement-of-VIrtualized-Systems-26-320.jpg)
![Dipartimento di Elettronica e Informazione
28SPECweb2005 Experiment
§ Two VMs running the banking and e-commerce loads
§ The number of users Ni
k accessing each of the two VMs varied stepwise every 1
minute, with values between 10 and 220
§ Proportional assignment scheme:
§ 1,440 intervals (24 hours)
§ Parametrization [Nk
i ρk]](https://image.slidesharecdn.com/parvis-130713122034-phpapp01/85/D-Ardagna-PARVIS-Performance-mAnagement-of-VIrtualized-Systems-28-320.jpg)
D. Ardagna - PARVIS - Performance mAnagement of VIrtualized Systems
- 1. PARVIS - Performance mAnagement of VIrtualized Systems Danilo Ardagna joint work with Mara Tanelli and Marco Lovera Dipartimento di Elettronica e Informazione, Politecnico di Milano ardagna@elet.polimi.it Milan, November 23 2010
- 2. Dipartimento di Elettronica e Informazione 2 Data Centers, Virtualization, and Cloud Computing § Nowadays, large Data Centers provide computational capacity on demand to many customers by sharing a pool of IT resources: • IBM Cloud • Amazon EC2 • Windows Azure • … § Issues: • Workload variability and QoS guarantees • Energy consumption
- 3. Dipartimento di Elettronica e Informazione Workload variability § Requests rates may change by order of magnitudes with a business day 3
- 4. Dipartimento di Elettronica e Informazione Workload variability § Requests rates may change by order of magnitudes with a business day § Traffic surges 4
- 5. Dipartimento di Elettronica e Informazione Workload variability § Requests rates may change by order of magnitudes with a business day § Traffic surges § Sport events 5
- 6. Dipartimento di Elettronica e Informazione Data Center energy consumption: An environmental problem... About 0.5% of global electric power consumption is due to DC In developed country: § UK: 2.2-3.3% § USA: 1.5% From the environmental point of view: • 2% of global CO2 emissions Source: EU Commission 0.0% 5.0% 10.0% 15.0% 20.0% 25.0% 2005 2020 % European IT Energy Consumption Cellular phone Network Telecom Network Server & Data Center
- 7. Dipartimento di Elettronica e Informazione Data Center energy consumption: …but first an economic one New Servers costsEnergy and cooling costs IT costs Number of Servers (M units) Company Server Electric Power (TWh) Cost eBay 16K 0.06 $3.7M Microsoft >200K 0.6 >$36M Google >500K 6 >$38M 25% 28% 12% 35% Source: EU Commission Source: Microsoft Research
- 8. Dipartimento di Elettronica e Informazione DC Inefficiencies Courtesy of IBM
- 9. Dipartimento di Elettronica e Informazione 9Virtualization of Physical Resources § Virtualization, proposed in early ’70s, is driving again the interest both of industry and academia § Enabling technology for server consolidation and cloud computing § Advantages: • Physical resources are partitioned among competing running VMs, improved security and reliability, performance isolation • Resource allocation parameters can be updated by in few milliseconds without introducing any system overhead
- 10. Dipartimento di Elettronica e Informazione 10 § Hardware resources (CPU, RAM, ecc...) are partitioned and shared among multiple virtual machines (VMs) § The virtual machine monitor (VMM) governs the access to the physical resources among running VMs Virtualization of Physical Resources
- 11. Dipartimento di Elettronica e Informazione 11Virtualization of Physical Resources: Research Challenges § Performance modelling of virtualized environments is challenging § Traditional queueing network models are inadequate to model virtualized systems performance at a very fine- grained time scale
- 12. Dipartimento di Elettronica e Informazione PARVIS goals § Develop novel resource allocation policies virtualized cloud infrastructures via an interdisciplinary approach: § Performance evaluation and optimization methods for the long-term management of the physical infrastructure § System identification and control engineering methods to derive load-dependent black-box models of virtualized systems and to design short-term control systems 12
- 13. Dipartimento di Elettronica e Informazione Internet Application2 Application1 Application3 Free Server Pool PARVIS Data Center: Autonomic resource management
- 14. Dipartimento di Elettronica e Informazione Virtual Machine Monitor S.O. App1 S.O. App2 S.O. Appn … VM1 VM2 VMn Infrastructure controller System Controller Performance metrics Performance goals DFS CPU weights Admission control Local controller PARVIS Reference framework Local controller Short term time horizon Dynamic models à Control theory Time scale: minute/seconds Fine grained performance and energy consumption goals Workload partitioning Performance goals of individual servers Infrastructure controller Long term time horizon Queuing network models à Non linear optimization Time scale: ten minutes/hour ... ... ...
- 15. Dipartimento di Elettronica e Informazione PARVIS Reference framework § Infrastructure controller: • Mixed Integer Non Linear Problem • Local Search § Local controller: • Linear Parameter Varying Models • Model Predictive Controllers 15 § D. Ardagna, B. Panicucci, M. Trubian, L. Zhang Energy-Aware Autonomic Resource Allocation in Multi-tier Virtualized Environments. IEEE Transactions on Services Computing. To Appear. § M. Tanelli, D. Ardagna, M. Lovera. Identification of LPV state space models for Autonomic Web service systems. IEEE Transactions on Control Systems Technology. To Appear. § B. Addis, D. Ardagna, B. Panicucci , L. Zhang. Autonomic Management of Cloud Service Centers with Availability Guarantees. In Cloud 2010 Proceedings. § D. Ardagna, M. Tanelli, M. Lovera, L. Zhang. Black-box Performance Models for Virtualized Web Service Applications. Proceedings of the 1st Joint WOSP/SIPEW International Conference on Performance Engineering. WOSP/SIPEW2010. ACM DL.
- 16. Dipartimento di Elettronica e Informazione 16 Revenues are a function of average response times Average response time soft-constraint SLA – Service Level Agreement Average response time Revenues
- 17. Dipartimento di Elettronica e Informazione 17 • Open queueing network model: heterogeneous service centers and a delay center • VMM modelling: GPS (Generalized Processor Sharing) scheduling Hexogenous arrival rate Session modelling Service centers model physical servers which support VMs execution A class k request becomes a request k’ with probability pk,k’ or terminates Service Center Performance Model
- 18. Dipartimento di Elettronica e Informazione 18Optimization Problem • Objective: maximize SLA revenues minus energy costs • Decision variables: § xi server i ON/OFF (binary variable) § λk i,j server i arrival rate for the VM operating at tier j of request class k § φk i,j server i CPU capacity fraction devoted to the VM operating at tier j of request class k § zk i,j assignment of the VM operating at tier j of request class k to server i (binary variable) § fi,h server i operating frequency (binary variable)
- 19. Dipartimento di Elettronica e Informazione 19 • Heuristic solution based on problem decomposition: § Initial solution: Assign VMs to physical servers (problem equivalent to a special case of a CFLP, Capacitated Facility Location Problem) § Optimum load balancing and capacity allocation (fixed point iteration) § The solution is then enhanced by a local search: • Switch servers ON and OFF • Change VMs placement • Change servers’ CPU frequency Local search
- 20. Dipartimento di Elettronica e Informazione 2020 Linear Parameter Varying (LPV) systems are a class of time-varying systems In discrete-time state space form: “Time varying systems, the dynamics of which are functions of a measurable, time varying parameter vector p.” LPV State Space Models LPV System uk yk pk
- 21. Dipartimento di Elettronica e Informazione 21LPV state-space models LPV System uk yk pk λk 2 Arrival rate Wk 1 , Wk 2 sk 1 , sk 2 Rk 1, Rk 2 Arrival rate λk 1 φk 1 φk 2 • Virtualized system identification: • Scheduling parameters: arrival rates, requests service times • Ouput variables: requests response times • Control variables: VMM parameters
- 22. Dipartimento di Elettronica e Informazione 22 • Real log traces (Politecnico di Milano Web site), 10 requests classes • Comparison with IBM Tivoli resource allocation policies Scenario 1: users come from the same time zone Scenario 2: users come from different time zones Infrastructure controller: IBM Tivoli comparison
- 23. Dipartimento di Elettronica e Informazione 23IBM Tivoli comparison – scenario 1 Our solution IBM Tivoli Our solution IBM Tivoli
- 24. Dipartimento di Elettronica e Informazione 24IBM Tivoli comparison – scenario 2 Our solution IBM Tivoli Our solution IBM Tivoli
- 25. Dipartimento di Elettronica e Informazione 25System Identification - Experimental setting § Two reference scenarios: • A Micro benchmarking instrumented Web application • SPECweb2005 industrial benchmark § VMM monitor: Xen 3.0 and Xen 3.3 § Validation: Synthetic workload inspired by a real-world. Log trace from a large financial system
- 26. Dipartimento di Elettronica e Informazione 26Micro-benchmarking Web Service Application Experiments § Number of VMs varied between 2 and 4 § For system identification purposes request arrival rates vary stepwise every 1 minute § Each request consumes si k CPU time varied between 0.06 s and 1.1 s. § 1,440 intervals (24 hours) § Parametrization [sk i ρk i]
- 27. Dipartimento di Elettronica e Informazione 27 Micro-benchmarking Web Service Application Experiments
- 28. Dipartimento di Elettronica e Informazione 28SPECweb2005 Experiment § Two VMs running the banking and e-commerce loads § The number of users Ni k accessing each of the two VMs varied stepwise every 1 minute, with values between 10 and 220 § Proportional assignment scheme: § 1,440 intervals (24 hours) § Parametrization [Nk i ρk]
- 29. Dipartimento di Elettronica e Informazione 29 SPECweb2005 Experiment
- 30. Dipartimento di Elettronica e Informazione 30PARVIS future work § Analysis of real applications § Local controller design § Integration of the two approaches
- 31. Dipartimento di Elettronica e Informazione § Mara Tanelli, Marco Lovera, Barbara Panicucci, Marco Bergamasco, Alessandro Barenghi, Alessandro Colleoni, Bernardetta Addis, Giuliana Carello, Antonio Capone, Politecnico di Milano § Folco Bombardieri, Danilo Ghirardelli, Gianluca Pisati, Giovanni Pirotta, Marco Caldirola, Mauro Speroni, Paolo Sala, Marco Casiero, Stefano Vettor, Massimo Bergami, Politecnico di Milano students § Marco Trubian, Università degli Studi di Milano § Li Zhang, IBM Research § Massimo Leoni, Carla Milani, IBM Italy Akwnoledgements