On Tuesday, June 25, 2002, IBM will have an exhibit at ICS'02, with a focus on Autonomic Computing topics. Here are the demos that will be part of the IBM exhibit:
Application service providers (ASPs),
storage service providers (SSPs), and similar hosting companies outsource
applications, and storage capacity. Customer can negotiate a service level
agreement that defines access to various combinations of hardware and software.
Common models for implementing computing utility infrastructures are the co-location
and fully managed models. In these
models, dedicated resources are statically assigned to and configured for
specialized tasks. Additional resources are manually added. In [FKT01], Ian
Foster et. al. refers to the ASP and
SSP technology as an important low-level building block of (GRID-based) virtual
organizations, and to dynamic reconfiguration and resource management within a
single ASP as a challenge rarely attempted. The Oceano system that we are going to present is a
cluster resource management system the addresses exactly this challenge.
Océano [AFG01] provides management software that enables an application hosting service to allocate servers to several customers’ applications in response to usage. Océano maintains a pool of servers on a common network infrastructure. Mechanisms are provided that automate the allocation of theses servers to hosted customers’ applications. When a server is allocated, it is primed with a fresh copy of an operating system, customer applications and data, and the network is reconfigured for placing the server into the customer’s service domain. Performance monitoring data from the server is correlated with data from the other servers resulting in an overall view of the performance of the application. The overall performance will then drive subsequent cycles of resource allocation or de-allocation based on policy. A failure detection service sends notifications on servers that are down; the servers are automatically replaced by fresh servers from the free-pool.
In this demonstration we will present the Oceano system manages a cluster of servers that is used by 3 fictitious application domains (customers). A dedicated set of servers is allocated to each of the customers. The servers that are not assigned are kept in a free pool. A set of metrics (response time, CPU load, etc.) is monitored continuously. The Oceano’s GUI shows the current server assignment, the monitored metrics, and the pre-defined threshold values (for dynamic allocation) for every customer. We will demonstrate 3 different functions of the Oceano system:
Failure detection: A down server is detected as such and automatically replaced.
Automatic Allocation: Upon a request (through a simple menu in our GUI) a server is allocated to one of the customers.
Dynamic Allocation based on load: When a response time threshold value is exceeded, a server is automatically allocated to a customer.
[AFG+01], K. Appleby, L. Fong, , G. Goldszmidt, S. Krishnakumar, S. Fakhouri, D. Pazel, J. Pershing, and B. Rochwerger,. “Océano - SLA-based management of a computing utility”, Proc. of the 7th IFIP/IEEE International Symposium on Integrated Network Management (IM2001), May, 2001
[FKT01] I. Foster, C. Kesselman, S. Tuecke, “The Anatomy of the Grid: Enabling Scalable Virtual Organizations”, Intl. J. Supercomputer Applications, 15(3), 2001.
Managing the performance of commerce sites is quite challenging. Site content changes frequently as do customer interests and business plans. All of this contributes to dynamically varying workloads. To maintain good performance, system administrators must tune their IT environment on an on-going basis. Unfortunately, doing so requires considerable expertise. This talk presents the use of AutoTune agents to automate much of this task. Our work exploits intelligent agents that employ techniques from control theory for automatic tuning. The Apache web server is investigated in detail. Two tuning parameters (i.e., MaxClients and KeepAlive) of the Apache web server are considered and the performance metrics are the CPU and MEM utilizations. We apply the linear quadratic regulation (LQR) technique to design the model-based feedback controller to handle the dynamic and interrelated dependencies between the tuning knobs and performance metrics. Using the Agent Building and Learning Environment (ABLE), we construct AutoTune agents to automate the procedure of model building, controller design, and run-time feedback control. The effectiveness of the AutoTune agents is demonstrated through a live demo regarding to variations in workload.