Examples and definitions are given for SaaS, PaaS, and IaaS. Computational models must be designed with the problems and effective resources in mind. A demonstration of cloud use for Bioinformatics on the FutureGrid educational testbed shows how clouds offer advantages of provisioning and virtual cluster support. Overhead and performance issues are touched upon through charts showing the use of three different virtual clusters.
Todo
slide is outdated
The work of previous B649 students is presented, covering a variety of topics and software: HBase, DryadLINQ, virtualization, commercial cloud storage, and IU’s Twister iterative MapReduce.
This lecture offers further examples of prior student projects in areas like cloud storage, infrastructure and platforms, as well as high level languages like Apache Pig. Other topics cover MapReduce in Single Program Multiple Data mode, genetic algorithms in MapReduce, Hadoop supporting recommender systems, K-means clustering, etc. Following this the benefits/vulnerabilities of clouds and some of their most well-known programs are showcased in other research assignments.
Todo
slide is outdated
Importance of virtualization is explored, including cross-platform applications. Virtualization has seen rapid growth in recent years in terms of use and services offered. Virtual machines differ from traditional computers in that software virtualization layer (hypervisor) runs on hardware, allowing guest OS to run on top of host OS. VMs can run independent of hardware specifications. Four different types of VM architecture, defined by the layer which the virtual machine monitor (VMM) runs on. VM is identical to physical machines and can be saved and stored, as well as migrated across hardware.
Virtualization can be implemented on five levels: application, library, OS, hardware, and instruction. Their benefits are compared in terms of performance, flexibility, complexity, and isolation. A layout is provided for the Linux virtualization layer, OpenVZ (OS level), which creates virtual private servers. CUDA is a high performance computing library, not designed for VMs; vCUDA is a virtual layer that allows interaction between CUDA and VMs, creating a virtual CUDA library.
A list of major hypervisors is given. Type 1 hypervisor resides on the bare metal computer, while Type 2 runs over the host OS. XEN is an open source hardware level hypervisor: consists of hypervisor, kernel, and application. Domain0 in XEN is a VM that manages other VMs. Two types of hardware virtualization: full virtualization and host-based virtualization. Para-virtualization does not need to modify the guest OS like full virtualization and works through hypercalls. An example is the ESX server from VMware.
A hybrid approach to virtualization involves offloading some tasks to the hardware to reduce overhead. This can be combined with para-virtualization for even greater effects. In a guest OS, the VMM provides shadow page tables to transfer virtual memory to machine memory. An example is shown in the Intel Extended Page Table. A virtualization layer for an I/O device is possible, allowing it to act like a physical device and manage host and guest addresses, shown in a detailed VMware example.
Characteristics of VM clusters are listed, including the ability to run multiple VMs on the same node and size alteration. Physical clusters are linked through nodes, while virtual clusters can be linked through physical or virtual nodes and can be replicated in virtual servers. Prepackaged OS can be installed in a virtual cluster. Should a VM fail for any reason, its image can be migrated to a new host so work is not lost. An example of this is demonstrated with XEN.
Whole data centers can be virtualized, enabling for the construction of private clouds. Some tools for Infrastructure as a Service clouds are Nimbus, Eucalyptus, OpenNebula, and vSphere. Eucalyptus is shown in greater detail. Trust issues in cloud security are answered in virtual machines. Suggested reading material is provided at the end.
Clouds run as servers for data storage and sharing on the Internet in an on-demand capacity. Cloud services are scalable depending on the client’s needs, allowing for a seemingly limitless source of computing power that can expand or shrink to meet financial demands. Some examples of cloud services are LinkedIn, Amazon S3, and Google App Engine. Different variations of clouds like IaaS and PaaS are offered by both open source and commercial providers. Cloud systems are composed of separate elements like Eucalyptus, Xen and VMware.
Todo
slide is missing
The capabilities of several IaaS cloud structures like Amazon EC2 or PaaS like Microsoft Azure are listed. Public and private clouds share certain features; the main difference is public clouds are owned by service providers while private clouds are offered by individual corporations. Certain enabling technologies are required for clouds to provide quick and scalable computing. These include virtual cluster provisioning and multi-tenant environments. PaaS demands the capability to process huge amounts of data as in the case of web searches. Some challenges faced by cloud computing include vendor lock-in owing to lack of standard APIs and metrics; for scientists, there is uncertainty about whether experiments can be reproduced effectively in different cloud environments. However there are distinct advantages clouds potentially have to offer: standardized APIs can eliminate lock-in, and encryption offers data confidentiality.
Todo
slide is missing
Huge data centers enable cloud computing, containing up to a million servers. Large data centers charge less for their services than small ones. A diagram illustrates the typical setup of a cloud; rack space on the bottom, on top of which are load balancers, then excess routers and border routers. The next figure compares cost effectiveness in a traditional IT model to a cloud. Other figures display small server clusters and a typical data center arrangement, including emergency power supply and cooling system. A chart shows the power consumption based on CPU, disk, etc. Disks in warehouse servers may be onsite or attached to outside connections like InfiniBand. Switches can form an array of racks. The distribution of memory across a local, rack, or array server in warehouse server setup is listed.
Todo
slide is missing
Power utilization effectiveness (PUE) for a warehouse is determined by comparing it to IT power usage. Racks can contain 40 servers, shipping containers can have up to 1,000 servers; a data center could take 2 years to construct. Warehouse scale computing has greater economy of scale than data centers by reducing network and administrative costs. Individual users can interact with clouds in the SaaS model, while organizations use PaaS. Clouds generally use VMs to recover from system failures. It is predicted that the cloud job market and demand for clouds will experience great growth in the future. Clouds have become ubiquitous in all aspects of the private and public sector. In the future clouds must take into account user privacy, data security and copyright protection.
Todo
slide is missing