IBM Systems Magazine, Mainframe - September 2014 - (Page 46)

ADMINISTRATOR Improve SCALABILITY of Hardware Threads IBM tests show finding the optimal number of n-way processors helps maximize performance T he view of system behavior for z/OS* symmetric multiprocessing with a large number of processors (n-way) has changed substantially. Traditional views of system performance at high n-way are outdated. Mark Wisniewski is a senior technical staff member for the System z platform. He has spent his entire career at IBM working on systems performance. At IBM, we have done some recent work to dramatically improve the scalability of some key internal System z* workloads using a number of middleware products on z/OS. While Parallel Sysplex*, with its datasharing capabilities, remains a vital z/OS environment, the focus for this work was single-system z/OS image scaling with a large number of processors in one LPAR. Number of n-way Processors Matters It has always been a challenge to ensure software components can efficiently exploit large numbers of hardware execution threads or processors within a single OS image. We have been working for decades to improve how well systems scale to achieve acceptable performance as more processors are added to a single system image. 46 // SEPTEMBER/OCTOBER 2014 Depending on a variety of systems constraints and performance issues that arise, the performance gain realized from adding each additional processor can quickly approach zero or, even worse, go negative. The key to understanding how the OS and related middleware products will behave as the number of processors in a single system image grows is the shape of the n-way curve. We must scale up the desired workload environments to sufficiently high n-way values to stress various system components across the OS, middleware and hardware; instrument these components using hardware and software performance monitors to provide insights into bottlenecks that may be arising; and prototype and make product improvements available to address these issues. Several aspects of large systems stack performance and scale bear on this discussion including: ĀÃ +DUGZDUHÃWKUHDGÃVSHHGÃZKLFKÃ can be driven by hardware technology and software code optimization ĀÃ 0HPRU\ÃVFDODELOLW\ÃZKHUHÃ software can play a big role by removing large memory exploitation inhibitors ĀÃ 3DUDOOHOÃ6\VSOH[ÃZKHUHÃ software can allow for horizontal growth by removing inhibitors and constraints ĀÃ +DUGZDUHÃWKUHDGOHYHOÃ parallelism, enabled by multiple cores or processors, large system software constraints removal, and n-way scalability of the operating system and middleware This last point, of scaling large numbers of parallel hardware threads, is this article's focus.

Table of Contents for the Digital Edition of IBM Systems Magazine, Mainframe - September 2014

Table of Contents
Editor's Desk: Moving in the mobile world
IBM Perspective: Mobile money thanks to the mainframe
Partner PoV: A clean IT audit does not guarantee network or computer security
Focus on Storage: How Elastic Storage helps remove bottlenecks and lower costs
Currents: IBM Smarter Storage puts businesses in control of their data
Trends: Choosing the best proprietary or open-source applications
Case Study: Cooperative Consolidation: The Credit Union System of Brazil saves millions of dollars by virtualizing on the mainframe
Cover Story: The Mobile Mainframe: System z is a perfect fit for the growing trend
Feature: Connecting Applications and Data: WebSphere Liberty z/OS Connect makes business assets accessible to mobile devices
Tech Showcase: How Linux on System z has become a go-to solution for businesses
Administrator: IBM tests show finding the optimal number of n-way processors helps maximize performance
Tech Corner: The new z/OSMF Workflows feature optimizes programming tasks
Solutions: OMC-FLASH 4.7.0; z86VM; Storage Director 4.2
Stop Run: IBMer's passion for engineering puts his house to work
Reference Point - Global Events, Education, Resources for Mainframe
2015 Mainframe Solutions Edition

IBM Systems Magazine, Mainframe - September 2014