Memory resource balancing for virtualized computing

Virtualization has become a common abstraction layer in modern data centers. By multiplexing hardware resources into multiple virtual machines (VMs) and thus enabling several operating systems to run on the same physical platform simultaneously, it can effectively reduce power consumption and building size or improve security by isolating VMs. In a virtualized system, memory resource management plays a critical role in achieving high resource utilization and performance. Insufficient memory allocation to a VM will degrade its performance dramatically. On the contrary, over-allocation causes waste of memory resources. Meanwhile, a VM’s memory demand may vary significantly. As a result, effective memory resource management calls for a dynamic memory balancer, which, ideally, can adjust memory allocation in a timely manner for each VM based on their current memory demand and thus achieve the best memory utilization and the optimal overall performance. In order to estimate the memory demand of each VM and to arbitrate possible memory resource contention, a widely proposed approach is to construct an LRU-based miss ratio curve (MRC), which provides not only the current working set size (WSS) but also the correlation between performance and the target memory allocation size. Unfortunately, the cost of constructing an MRC is nontrivial. In this dissertation, we first present a low overhead LRU-based memory demand tracking scheme, which includes three orthogonal optimizations: AVL-based LRU organization, dynamic hot set sizing and intermittent memory tracking. Our evaluation results show that, for the whole SPEC CPU 2006 benchmark suite, after applying the three optimizing techniques, the mean overhead of MRC construction is lowered from 173% to only 2%. Based on current WSS, we then predict its trend in the near future and take different strategies for different prediction results. When there is a sufficient amount of physical memory on the host, it locally balances its memory resource for the VMs. Once the local memory resource is insufficient and the memory pressure is predicted to sustain for a sufficiently long time, a relatively expensive solution, VM live migration, is used to move one or more VMs from the hot host to other host(s). Finally, for transient memory pressure, a remote cache is used to alleviate the temporary performance penalty. Our experimental results show that this design achieves 49% center-wide speedup.

Managing server energy and reducing operational cost for online service providers

The past decade has seen the energy consumption in servers and Internet Data Centers (IDCs) skyrocket. A recent survey estimated that the worldwide spending on servers and cooling have risen to above $30 billion and is likely to exceed spending on the new server hardware . The rapid rise in energy consumption has posted a serious threat to both energy resources and the environment, which makes green computing not only worthwhile but also necessary. This dissertation intends to tackle the challenges of both reducing the energy consumption of server systems and by reducing the cost for Online Service Providers (OSPs). Two distinct subsystems account for most of IDC’s power: the server system, which accounts for 56% of the total power consumption of an IDC, and the cooling and humidifcation systems, which accounts for about 30% of the total power consumption. The server system dominates the energy consumption of an IDC, and its power draw can vary drastically with data center utilization. In this dissertation, we propose three models to achieve energy effciency in web server clusters: an energy proportional model, an optimal server allocation and frequency adjustment strategy, and a constrained Markov model. The proposed models have combined Dynamic Voltage/Frequency Scaling (DV/FS) and Vary-On, Vary-off (VOVF) mechanisms that work together for more energy savings. Meanwhile, corresponding strategies are proposed to deal with the transition overheads. We further extend server energy management to the IDC’s costs management, helping the OSPs to conserve, manage their own electricity cost, and lower the carbon emissions. We have developed an optimal energy-aware load dispatching strategy that periodically maps more requests to the locations with lower electricity prices. A carbon emission limit is placed, and the volatility of the carbon offset market is also considered. Two energy effcient strategies are applied to the server system and the cooling system respectively. With the rapid development of cloud services, we also carry out research to reduce the server energy in cloud computing environments. In this work, we propose a new live virtual machine (VM) placement scheme that can effectively map VMs to Physical Machines (PMs) with substantial energy savings in a heterogeneous server cluster. A VM/PM mapping probability matrix is constructed, in which each VM request is assigned with a probability running on PMs. The VM/PM mapping probability matrix takes into account resource limitations, VM operation overheads, server reliability as well as energy effciency. The evolution of Internet Data Centers and the increasing demands of web services raise great challenges to improve the energy effciency of IDCs. We also express several potential areas for future research in each chapter.

Potential for hardware-based techniques for reuse distance analysis

Reuse distance analysis, the prediction of how many distinct memory addresses will be accessed between two accesses to a given address, has been established as a useful technique in profile-based compiler optimization, but the cost of collecting the memory reuse profile has been prohibitive for some applications. In this report, we propose using the hardware monitoring facilities available in existing CPUs to gather an approximate reuse distance profile. The difficulties associated with this monitoring technique are discussed, most importantly that there is no obvious link between the reuse profile produced by hardware monitoring and the actual reuse behavior. Potential applications which would be made viable by a reliable hardware-based reuse distance analysis are identified.

Development of a high performance parallel computing platform and its use in the study of nanostructures : clusters, sheets and tubes .

Small clusters of gallium oxide, technologically important high temperature ceramic, together with interaction of nucleic acid bases with graphene and small-diameter carbon nanotube are focus of first principles calculations in this work. A high performance parallel computing platform is also developed to perform these calculations at Michigan Tech. First principles calculations are based on density functional theory employing either local density or gradient-corrected approximation together with plane wave and gaussian basis sets. The bulk Ga2O3 is known to be a very good candidate for fabricating electronic devices that operate at high temperatures. To explore the properties of Ga2O3 at nonoscale, we have performed a systematic theoretical study on the small polyatomic gallium oxide clusters. The calculated results find that all lowest energy isomers of GamOn clusters are dominated by the Ga-O bonds over the metal-metal or the oxygen-oxygen bonds. Analysis of atomic charges suggest the clusters to be highly ionic similar to the case of bulk Ga2O3. In the study of sequential oxidation of these slusters starting from Ga2O, it is found that the most stable isomers display up to four different backbones of constituent atoms. Furthermore, the predicted configuration of the ground state of Ga2O is recently confirmed by the experimental result of Neumark's group. Guided by the results of calculations the study of gallium oxide clusters, performance related challenge of computational simulations, of producing high performance computers/platforms, has been addressed. Several engineering aspects were thoroughly studied during the design, development and implementation of the high performance parallel computing platform, rama, at Michigan Tech. In an attempt to stay true to the principles of Beowulf revolutioni, the rama cluster was extensively customized to make it easy to understand, and use - for administrators as well as end-users. Following the results of benchmark calculations and to keep up with the complexity of systems under study, rama has been expanded to a total of sixty four processors. Interest in the non-covalent intereaction of DNA with carbon nanotubes has steadily increased during past several years. This hybrid system, at the junction of the biological regime and the nanomaterials world, possesses features which make it very attractive for a wide range of applicatioins. Using the in-house computational power available, we have studied details of the interaction between nucleic acid bases with graphene sheet as well as high-curvature small-diameter carbon nanotube. The calculated trend in the binding energies strongly suggests that the polarizability of the base molecules determines the interaction strength of the nucleic acid bases with graphene. When comparing the results obtained here for physisorption on the small diameter nanotube considered with those from the study on graphene, it is observed that the interaction strength of nucleic acid bases is smaller for the tube. Thus, these results show that the effect of introducing curvature is to reduce the binding energy. The binding energies for the two extreme cases of negligible curvature (i.e. flat graphene sheet) and of very high curvature (i.e. small diameter nanotube) may be considered as upper and lower bounds. This finding represents an important step towards a better understanding of experimentally observed sequence-dependent interaction of DNA with Carbon nanotubes.

Development of Hardware-In-the-Loop Simulation System for Hybrid Electric Vehicle Study

The development of embedded control systems for a Hybrid Electric Vehicle (HEV) is a challenging task due to the multidisciplinary nature of HEV powertrain and its complex structures. Hardware-In-the-Loop (HIL) simulation provides an open and convenient environment for the modeling, prototyping, testing and analyzing HEV control systems. This thesis focuses on the development of such a HIL system for the hybrid electric vehicle study. The hardware architecture of the HIL system, including dSPACE eDrive HIL simulator, MicroAutoBox II and MotoTron Engine Control Module (ECM), is introduced. Software used in the system includes dSPACE Real-Time Interface (RTI) blockset, Automotive Simulation Models (ASM), Matlab/Simulink/Stateflow, Real-time Workshop, ControlDesk Next Generation, ModelDesk and MotoHawk/MotoTune. A case study of the development of control systems for a single shaft parallel hybrid electric vehicle is presented to summarize the functionality of this HIL system.