Driving up Server Farm CPU Utilization

Nowadays leading technology is in the business of producing intellectual properties such as software, designs, movies, etc. on large number of computers. It takes more than 50 million dollars to build a server farm with five thousand CPUs. Each percentage of CPU utilization costs at least 500 thousand dollars.

What is the average CPU utilization in your server farm? You may be surprised by the low number if you collect the data and figure it out. If you have job slots configured for your farm, a more meaningful performance indicator of your farm is job slot utilization.

As R&D IT budget goes down year after year for many companies, you can unleash tremendous compute power from your server farm without much new investment if you can raise your CPU utilization and job slot utilization consistently while maintaining your service levels. The first step toward a more efficient server farm is to select some CPU utilization metrics for your farm and collect the data. Once you establish baseline utilization, you are ready to run a more efficient farm.

The single most important component of your server farm is LSF master. It works as a pace make for the heart of your business. LSF master a) collect the load information of all LSF servers in the cluster, b) accept jobs from users, c) schedule jobs to the open slots on these servers, and d) control the execution of the jobs. LSF master's hardware capacity needs to be over-provisioned. If a bottleneck exists in any of above areas, the utilization of your farm will go down. As workloads and usage patterns differ from companies to companies and from time to time in the same company, it is difficult to know when you may have a bottleneck unless you keep a record of your master performance for a long period of time. You need to have a chronicle "medical history" of your unique LSF master in order to ensure there is enough head room on your master.

Furthermore, your file system has to be fast enough to avoid a disk I/O bottleneck when you have close to half a million of jobs and each job is recorded as a file in LSF /work directory in addition to an lsb.event file with half a million lines. Your LSF installation directories should be customized and they should locate on different types of file systems. One of biggest reason of server farm low utilization is down time, ranging from "daemon not responding" to LSF master failures. You need to design your cluster with high availability in mind, minimize interruption to the cluster by b* commands including bjobs. Backfill your jobs if there is any resources reservation ahead. Job exception management enables you to make sure job are scheduled and executed correctly. Storage aware scheduling ensures file systems are ready to read and write files.

In summary, it takes all parts of IT infrastructure working coherently to run high performance and highly efficient farm.

HP EDA Computing Technologies and Case Studies

EDA (Electronic Design Automation) silicon and systems design complexity and computing performance demands continue to increase rapidly.  Coupled with this is a broad range of new emerging computing technologies such as multi-core processors, high speed computing node interconnects, hardware accelerators and high speed large scale storage.  In this presentation we will explore several customer case studies from leading chip design companies and see how they are taking advantage of computing technology and software management infrastructure tools to drive increased design productivity and datacenter power and cooling efficiencies.

There are many new and emerging technologies and trends on the horizon that have the potential to impact our customers’ Grid and HPC environments. This session is intended as an opportunity for customers and partners to discuss in an open and interactive format, their experiences and perspectives on a range of emerging trends and technologies, providing insight and opinions on the impacts of these on their business.

Platform will provide some introductory thoughts to get the discussion rolling, and facilitate the interactive discussion, but please come prepared to provide your thoughts, comments and opinions on the following topic areas. If there are other areas of interest beyond these three, we are open to extending the discussion based on real time feedback and input.

This presentation will examine the business challenges at a large financial service firm, their technical and business objectives, and how Platform's Symphony, when combined with Windows Compute Cluster Edition, provided an HPC solution that exceeded customer expectations. The underlying HPC requirements for this firm and other financial service companies will be examined. In addition, Platform/Microsoft solutions that will incorporate Windows HPC Server 2008 for hedge fund analysis and other applications will be discussed.

Cheryl Doninger, Director of the Enterprise Computing Infrastructure, SAS Bio

John van Uden, Director, FICC Infrastructure & Shared Services Technology, Citi
Bio

Are you dealing with exponential increases in data volumes, over-utilized or obsolete SMP servers? Do you have the need to architect a computing platform that will grow as your business grows while reducing corporate costs at the same time?

SAS and Platform have partnered together to leverage all the benefits of a distributed HPC infrastructure to provide software solutions for many of the challenges facing enterprises today. Learn how SAS Business Intelligence and Analytics solutions can easily leverage a grid infrastructure and benefit from the intelligent resource allocation, load balancing and management capabilities provided by the Platform Computing components.
Customer implementations will be discussed to provide real world validation of the business value of running SAS solutions on a grid.

Analytics infrastructures need to scale as never before in order to compete in todays volatile market place. Many firms have begun to experiment with acceleration technologies in order to meet these needs. However, most do not have the performance and enterprise deployability of the IBM BladeCenter® QS22 with the third generation Cell Broadband Engine™ (Cell/B.E.™). Designed for ultra low-latency, high performance, and power efficiency in a standard blade form factor. When coupled with Platform Symphony the IBM BladeCenter® QS22 has the potential to substantially accelerate grid enabled analytics to help improve response time and ultimately a firm’s competitive position. See for yourself – IBM will share early customer experiences in running calculations on the QS22 blade utilizing the power of Platform Symphony, to enable the most powerfull, energy efficient grid imaginable.