Title: Fair workload distribution for multi-server systems with pulling strategies
Location: Meeting room, Building Zeta
Type: Research Result
Speaker: Andrea Marin
In this talk we present the paper that has received the best paper award at the conference Valuetools 2016.
We consider systems with a single queue and multiple parallel servers. Each server fetches a job from the queue immediately after completing its current work. We propose a pulling strategy that aims at achieving a fair distribution of the number of processed jobs among the servers. We show that if the service times are exponentially distributed then our strategy ensures that in the long run the expected difference among the processed jobs at each server is finite while maintaining a reasonable throughput.
We give the analytical expressions for the stationary distribution and the relevant stationary performance indices like the throughput and the system’s balance.
Interestingly, the proposed strategy can be used to control the join-queue length in fork-join queues and the analytical model gives the closed form expression of the performance indices in saturation.
Title: Optimisation of servers with different quality of services
Location: Meeting room
Type: Research Result
Speaker: Ivan Stojic
In many applications involving client/server or equivalent architecture, it is possible to vary the quality of services provided by the servers. Dynamic control of the quality of services is an effective mechanism for control of the performance of the system in conditions of varying system load. One of the basic problems involved in applying this mechanism is optimisation of the controlling policy that determines the level of quality of services based on the state of the system. In this seminar, after a brief introduction to Markov chains and basic queueing models, a Markov chain based model of a multiserver queue with processor-sharing discipline and different quality of services is presented. Calculation of the performance indices from the model and the prohibitive complexity of the exact solution are discussed next, motivating the introduction of a simpler model that heuristically approximates the original model and allows for more efficient model solution and optimisation of the controlling policy.
Title: Tuning of large computer clusters to maximise performances
Location: Meeting room
Speaker: Michele Mazzucco
The lecture is part of the course of “Performance and reliability of computer systems” and “Foundations of programming languages”. The speaker works for Demonware and Activision Blizzard and is part of the team that is in charge of the capacity planning for the company’s data centres. In this lecture, Dr. Mazzucco will describe the data centre infrastructure and its requirements for acceptable online gaming experiences. Then, he will show how the capacity planning of these large and powerful infrastructures can be carried out by using formal models
such as queueing systems.
Title: Revenue Maximization Problems in Commercial Data Centers
Time: 1:00 pm
Location: Sala riunioni
Type: industrial application
Speaker: Michele Mazzucco (Demonware, Dublin)
As IT systems are becoming more and more important, one of the main concerns is that users may face major breakdowns and eventually incur major costs if computing systems do not meet the expected performance requirements: customers expect reliability and performance guarantees, while under-performing systems loose revenues. For example, it has been reported that Amazon tried delaying the page generation by 100 ms and found out that even very small delays would result in substantial and costly drops in revenue (1% sales drop for 100 ms delay). In this talk I will discuss some performance models aiming at optimizing the revenue earned by IT providers running ‘jobs’ subject to Quality of Service (QoS) constraints. The presentation is divided into two parts. In the first part I will analyze a business model where the QoS guarantees are formally defined through Service Level Agreements (SLAs), and thus the provider is liable to pay a penalty every time the promised performance level is not met. Experimental results show that revenues can be dramatically improved by imposing suitable conditions for accepting incoming traffic, and that the proposed policies perform well under different traffic conditions.
In the second part of the presentation I will discuss two queueing models for power and performance. The main difference compared to the first part of the talk is that now the QoS is implicit, and thus customers simply leave the system (or wait) if it under-performs, while the provider also takes into account the energy consumed by servers when deciding how many servers to allocate.
Michele Mazzucco graduated in Computer Science at the University of Bologna and obtained his PhD at the University of Newcastle under the supervision of prof. Mitrani. His main reserach interests include models for the performance evaluation and optimization of data centers. He has published in major conferences and journals on topics such as cloud computing and green computing. Since 2012, he works for Demonware.
DemonWare is an Irish software development company and a subsidiary of Activision Blizzard. DemonWare’s products enable games publishers to outsource their networking requirements, allowing them to concentrate on playability. The organisation has offices in Dublin, Ireland; and Vancouver, Canada.Primary products developed by DemonWare include the “DemonWare State Engine” and “Matchmaking+”. The State Engine is a high-performance state synchronization C++ programming framework that eliminates the need to reinvent netcode multiplayer games. Matchmaking+ provides services for multiplayer games such as matchmaking, user profiling, and gaming statistics. DemonWare’s main product has been used to support the development of several online games of success, among which Call of Duty.
Hi there, the ACADIA website will be up and running soon … stay tuned!