Downloadable copies of our papers and presentations are available to registered site visitors. In addition, videos of our most recent presentations may be viewed from this page.
I've Looked at Clouds from Both Sides Now: Measuring VaporsDr. H. Pat Artis - Published in 2011While the value proposition of cloud computing is seductive to the management of organizations ranging in size from small and medium businesses to global enterprises, the measurement of the performance, reliability, and availability of the cloud-based applications needs to be carefully defined to serve as the basis for a contractual agreement between the service provider and the client. |
||
Evaluating the Impact of a Fusion-io ioDrive on a Complex Server ApplicationDr. H. Pat Artis, John Gatch, and Mike Hammock - Published in 2011Bus-attached flash storage devices offer a cost effective approach to addressing the I/O bottlenecks for server based applications. This paper evaluates the impact of a Fusion-io ioDrive on the performance of a complex Linux-based application compared to traditional SAS storage. The paper provides a methodology to evaluate the potential impact of bus-attached flash storage for server applications. |
||
z/OS Data De-De-De-De-duplicationDr. H. Pat Artis - Published in 2010The introduction of de-duplication technologies for virtual tape processing in the z/OS marketplace in early 2010 presents an entirely new set of opportunities and questions for enterprise users. This paper provides a high level overview of how de-duplication works, discusses the characteristics of z/OS workloads that lend themselves to de-duplication solutions, and proposes performance metrics for de-duplication appliances. |
||
Evaluating the Performance Characteristics and Resiliency of IBM DS8000 Global MirrorVic Peltz and Dr. H. Pat Artis - Published in 2009This paper presents the results of a study that evaluated the resiliency of IBM's DS8000 Global Mirror asynchronous remote replication solution. The study modeled workloads using operational data gathered from several Fortune 500 financial services companies. The study examined the behavior of Global Mirror under workloads found in typical steady-state operating conditions and additionally the behavior under transient conditions associated with workload spikes. The methodology for creating the workload model and selected results from the study are discussed. One workload, namely the opening 15-minutes of trading on the New York Stock Exchange, is discussed in detail. |
||
The Dawn of a New Paradigm: Understanding the Impact of Flash Devices on z/OS StorageDr. H. Pat Artis - Published in 2009While the introduction of storage subsystems comprised of commodity disk devices and large caches in the 1990s provided substantial improvements for both response time and cache-hit percentage, cache resistant workloads still present significant performance problems. Specifically, the vast majority of cache-misses are attributable to a small fraction of a subsystem’s storage capacity. This presentation will:
The presentation concludes with a discussion of the long term cost trends of flash storage and the storage subsystem architectures it will facilitate in the future. The video is approximately 1 hour in length. |
||
Product Tutorial: Using the PAI/O Driver for z/OS Workload Profiling Component to Perform Remote Copy AnalysisDr. H. Pat Artis - Published in 2009The planning and implementation of synchronous and asynchronous z/OS remote copy solutions present a myriad of workload characterization problems as well as risks to the enterprise. This product tutorial discusses:
This presentation is intended to provide a how-to tutorial for existing users and to provide an in-depth discussion of the PAI/O Driver for z/OS capabilities for installations that are considering the acquisition of the product. Each part of the video is approximately 45 minutes in length. |
||
The z/OS World is Flat: A Reexamination of the Storage PyramidDr. H. Pat Artis - Published in 2008Perhaps the most persistent mainframe storage-marketing paradigm of the last four decades is the storage pyramid. That is, data should be placed at the least expensive level of storage that meets its response and initial access time requirements. Since the cost of 2314 class storage in the early days of the System/360 was approximately $1,800 per MB, DASD was reserved for only the most critical data with the majority of an enterprise's data being relegated to tape. In fact, tape sort-work datasets persisted through the early 1970s due to the high cost of DASD storage. However, four decades of technological advances have reduced the cost per megabyte by a factor of more than 20,000:1. As a result, the slope (i.e., differential cost) of the storage pyramid's sides has significantly decreased. Coincidentally, the fully loaded salary expense of storage administrators and the quantities of storage for which they are responsible have vastly increased. Hence, it is time to ask the question, "Is the z/OS storage world flat?" That is, when does the life cycle cost of managing a storage hierarchy exceed the savings offered by a differential storage acquisition strategy? The video is approximately 1 hour in length. |
||
A Reexamination of Z/OS Storage TaxonomiesDr. H. Pat Artis - Published in 2008Since the inception of OS/MVT in the mid-1960s, volumes have been the corner stone of storage management. While the introduction of This paper will discuss the exploitation of HyperPAV managed aliases and Extended Address Volumes (EAVs) to define a new storage taxonomy for z/OS. This taxonomy applies the n-way ported memory model to storage performance and casts off the last vestiges of volume centric storage taxonomies. The video is approximately 1 hour in length. |
||
Understanding the Performance Implications of HyperPAVsDr. H. Pat Artis - Published in 2006HyperPAVs are intended to address the logical volume size and performance limitations that users have encountered for traditional static and WLM managed aliases. This paper provides a review of traditional PAVs and an overview of IBM's HyperPAV implementation. In addition, it presents initial performance results for HyperPAVs and suggests algorithms for estimating the number of HyperPAV managed aliases required to meet the performance requirements of a workload. The video is approximately 1 hour in length. |
||
Workload Characterization Algorithms for Remote Copy PlanningDr. H. Pat Artis - Published in 2006The planning and implementation of synchronous and asynchronous z/OS remote copy solutions present a myriad of workload characterization problems as well as risks to the enterprise. This presentation will discuss algorithms for calculating write data and I/O rates, identifying problem volumes, and present a methodology for testing and certifying a remote copy environment before committing your production workload. |
||
Understanding the Performance Implications of MIDAWsDr. H. Pat Artis - Published in 2006Modified Indirect Data Address Words (MIDAWs) are a new feature of IBM z9 processors, enabled by z/OS 1.7, that can substantially improve the performance of FICON connected storage subsystems. This paper provides an overview of MIDAWs and discusses how they can be employed to restructure complex channel programs. An experimental design is introduced for the evaluation of the performance implications of MIDAWs. Finally, the performance benefits of MIDAWs for 4K transfers are evaluated for FICON Express2 and FICON Express4 channels. |
||
Designing and Managing FICON Inter-Switch Link InfrastructuresSteve Guendert and Dr. H. Pat Artis - Published in 2006FICON Inter-Switch Links (ISLs) provide an elegant solution to the bandwidth requirements of complex parallel sysplex environments. This presentation will review basic ISL concepts, how they are defined in HCD, focus on design principals for fault tolerance, and examine the measurement data available for performance management and capacity planning. |
||
Managing Complex FICON ConfigurationsDr. H. Pat Artis - Published in 2005While FICON configurations provide a wealth of performance benefits, managing complex switched and cascaded director configurations can present a variety of challenges. This tutorial will examine the available performance metrics and discuss guidelines for the management of complex FICON configurations. |
||
Understanding the Differences Between z990 and z900 Service Time MeasurementsDr. H. Pat Artis - Published in 2004With the introduction of the z990, IBM increased the precision of the timing measurements provided by the I/O subsystem from 128 to 0.5 mu-secs. This paper will explore the benefits and impacts of this change and answer the question "did my service time really increase?" |
||
Running z/OS in a Little Itty-Bitty 64-Bit WorldDr. H. Pat Artis - Published in 2003With a 40 MIPS minimum entry-point, the 64-bit world provided by IBM's zSeries processors is beyond the reach of most small VM and VSE enterprises as well as z/OS developers. Pentium based z/Architecture emulation addresses this gap and provides a low-cost 64-bit solution for small environments. This paper will present an overview of the FLEX-ES emulator architecture as well as discuss Performance Associates' system programming, performance, and compatibility experiences with FLEX-ES on a zDev server as part of IBM's PartnerWorld for Developers program. |
||
Understanding FICON Channel Path MetricsDr. H. Pat Artis & Robert Ross - Published in 2003, Revised March 2004While FICON channels offer significant advantages in terms of channel bandwidth and performance, many users long for the old steam gauge utilization metrics provided for ESCON channels. While there are no simple rules of thumb for FICON channel management, this paper will explore the available path utilization metrics and suggest new metrics for FICON channels. |
||
An Experimental Design for the Evaluation of Switched and Cascaded FICON Director ConnectivityDr. H. Pat Artis & Robert Ross - Published in 2003This working document outlines an experimental design for the evaluation of switched and cascaded FICON director connectivity using the PAI/O Driver for z/OS. The experiments are designed to provide both throughput and quality of service metrics for switched and cascaded FICON topologies. |
||
FICON Configuration Design: Thinking Outside of z/BoxDr. H. Pat Artis & Robert Ross- Published in 2003Although a great deal has been written about the technological advantages of FICON connectivity for z/OS configurations, a fundamental question remains. Specifically, how will the introduction of FICON change configuration design, sizing, and management of storage configurations for large Sysplexes? This paper explores the configuration constraints imposed by ESCON and provides an overview of how FICON addresses them. |
||
Exploring the I/O Performance Characteristics of Intel Based FLEX-ES Servers for z/OSDr. H. Pat Artis - Published in 2002As a result of the continuing evolution of zSeries processors, an ever widening gap has developed between the minimum size of IBM's entry level processors and the MIPS requirements of small z/OS software developers as well as the thousands of VSE and VM installations. This paper explores the architecture of FLEX-ES product from Fundamental Software Incorporated (FSI) as delivered in the iDev member of Cornerstone System's iFrame product family. Specific attention will be paid to the evaluation of the performance provided by the FLEX-ES I/O subsystem. |
||
Understanding the Implications of Logical and Physical Volume Size on the Performance of z/OS Storage SubsystemsDr. H. Pat Artis - Published in 2001This paper explores the performance implications of increasing logical volume and SCSI device sizes on the performance of z/OS storage subsystems for random access workloads. Particular emphasis is placed on the historical performance implications of larger logical volume sizes on OS/MFT and its descendents, how parallel access volumes (PAVs) address the issue for a single logical volume, potential impacts of RAID schemas on cache-misses, cache sizing considerations, access density considerations, and the overall impact of SCSI device size. |
||
FICON I/O Measurements: Everything You Know is Wrong!Robert Ross - Published in 2001For more than 2 decades, I/O subsystem analysis has been based on device level PEND, DISC, and CONN times as well as channel utilization. While OEM and ESCON channels are circuit switched, FICON employs a packet switched protocol for the transmission of channel programs and data. The use of a packet switched protocol fundamentally redefines the critical device and channel level measurements on which I/O subsystem analysis is based. This paper explores how FICON redefines the RMF-measured service time components as well as new metrics for FICON directors and channels. |
||
Architectural Directions for Server I/O SubsystemsDr. H. Pat Artis - Published in 2000From the perspective of an experienced performance analyst, perhaps the most frustrating aspects of server I/O are the lack of hardware measurement data and the architectural limitations imposed by the traditional bus structures. This paper is intended to provide a tutorial on IBA, the InfiniBand Architecture, and will discuss what it will mean for the design of enterprise class servers, the generic layered driver model, as well as storage area networks in the future. Specific emphasis is placed on the primitives for I/O measurement that are incorporated in the architecture. |
||
Understanding the Performance Implications of PAVs and Multiple Allegiances for Storage SubsystemsDr. H. Pat Artis - Published in 1999Abstract: This paper presents the results of paired-sample evaluation of the benefits of parallel access volumes (PAVs) and multiple allegiances for reducing logical volume IOSQing delays. A simple back-of-the-envelope technique is provided for estimating the potential performance benefits of moving an existing workload to an IBM 2015 Enterprise Storage System or other storage subsystem which implements these features. |
||
DIBs, Data Buffers, and Distance: Understanding the Performance Characteristics of ESCON LinksDr. H. Pat Artis - Published in 1998Abstract: This paper will review the character set and frame structure employed to transmit channel programs, data, and status information over ESCON links. In addition, the performance aspects of the components of an ESCON adapter card will be discussed. Based on these conceptual discussions, a back-of-the-envelope model of ESCON frame traffic flow will be introduced which will be used to relate the performance aspects of DIB size, data buffer size, and ESCON transmission distance. |
||
An Introduction to Fibre Channel ConnectivityWayne Munson - Published in 1998With the current development of Storage Area Networks, clustering technologies, and gigabit networks, is it time for you to gain an understanding of Fibre Channel technology? This paper will present the basics of Fibre Channel technology including topologies, functional levels, physical media, and packets and frames. The paper has been updated to incorporate FICON. |
||
The Evolution of Performance Warranties in the DASD Acquisition Process at DSTPat Jones and Miller Dixon - Published in 1998Abstract: This paper discusses how DST Systems, Inc. uses the PAI/O Driver for z/OS to ensure that our new DASD subsystems are powerful enough and properly configured to meet our performance requirements. It describes a process that has been developed over a period of 2-3 years and has been designed to assist our management in making intelligent DASD acquisition decisions. |
||
Understanding the Performance Characteristics of Synchronous Remote CopyDr. H. Pat Artis - Published in 1997Abstract: With the availability of three different synchronous remote copy schemes (SRDF, PPRC, and HRC), many installations are starting to consider or install remote copy subsystems to ensure business continuance. Perhaps the most common question is will it work? This paper will explore the performance relationships between I/O rate, write fraction, distance, blocksize, number of copy links, and the number of communication lines between the primary and remote DASD subsystems. Emphasis will be placed on the experimental design and measurement issues associated with the analysis of synchronous copy schemes over common carrier lines. |
||
The Impact of Pipelining Protocols on Synchronous Remote Copy PerformanceDr. H. Pat Artis - Published in1997Abstract: The performance of the first generation of synchronous remote copy schemes (SRDF, PPRC, and HRC) is limited by two factors: distance and synchronous protocol employed to move the data over the communications line. This paper will explore the conceptual benefits of pipeline protocols, review experimental results for EMC's pipelining implementation that will become generally available in late 1997, and contrast the results with those for synchronous protocol schemes. |
||
Understanding Disk I/O Performance Metrics in Windows NTWayne Munson - Published in1997Abstract: As deployment of Microsoft Windows NT becomes more prevalent in the enterprise there is a requirement to understand the performance metrics associated with the operating system. This paper will present the metrics available for monitoring and characterizing the disk I/O subsystems attached to the Windows NT operating system. |
||
Toasters Need Tuning Too!Dr. H. Pat Artis & Wayne Munson - Published in 1997Abstract: In an era which is characterized by the explosion of applications of imbedded technologies, perhaps it is time to reexamine the old axiom that tuning activities should be focused on expensive and complex systems which typically reside on raised floors. Rather, tuning might be a desirable for even low cost imbedded systems that have profound productivity impacts on user communities that they support. |
||
MVS: Mainframe Virtual SecurityBrian Currah - Published in 1996 - Revised in 2002Abstract: Enterprises using IBM's premier operating system for System/390 mainframes, MVS/ESA, may have a false sense of confidence regarding the vulnerability of corporate assets. The protection offered by widely used security products can often be circumvented as a result of loopholes in other add-on software. Many such systems are frequently not nearly as resistant to security penetration as corporate management have been led to believe. New inspection, certification, and verification techniques must be implemented before MVS-based systems can be employed as a safe haven for the conduct of electronic commerce and a secure repository for corporate and national assets. |
||
Sibling PEND: Like a Wheel within a WheelDr. H. Pat Artis - Published in 1996Published in the Proceedings of CMG '96 Abstract: After more than 2 decades of refining performance measurement and tuning strategies for traditional DASD devices and the MVS I/O subsystem, performance analysts are now presented with the vagaries of measurement and tuning for RAID-5 based subsystems. While a variety of conceptual models might be proposed to explain the performance characteristics of these subsystems, this paper will focus on the author's paradigm of a physical I/O subsystem within the logical MVS I/O subsystem and the sibling PEND which results from collisions within the physical disk subsystem. |
||
Understanding the Performance Characteristics of EMIF Channel SubsystemsDr. H. Pat Artis - Published in 1995Abstract: A paired test methodology is employed to compare the performance characteristics of DASD subsystems supported by dedicated ESCON channels with shared EMIF channels on a partitioned system. The results of the paired tests provide comparisons of device service times, channel utilizations, PEND times, and other critical I/O subsystem metrics. |
||
Record Level Caching: Theory and PracticeDr. H. Pat Artis - Published in 1994Abstract: In this paper, we will examine experimental data collected for a 3990-6 responding to an I/O workload comprised of random read/write pairs. Read/write pair workloads are typical of business critical online transaction processing (OLTP) workloads such as credit card or ATM applications. The experimental data will be used as a basis to examine past theoretical explanations as well as to comment on the efficacy of IBM's new record level cache one (RLC I) facility. |
||
A Graphical Paradigm for the Design and Analysis of Distributed SystemsDr. H. Pat Artis - Published in 1994Abstract: A graphical paradigm for the design, representation, and analysis of distributed systems is discussed. For the purposes of this paper, we use the term distributed system to define client/server, peer-to-peer, multi-tier and other non-traditional hardware/software solutions that are typified by applications which employ the resources of multiple platforms to perform business processes rather than being traditional mainframe solutions. The primary objective of the paper is to introduce a schema for the graphical representation of distributed systems and to demonstrate how existing techniques such as degradation analysis can be embodied in the proposed paradigm. |
||
ECKD CAID and RAID: When is the Right Time to Write?Dr. H. Pat Artis - Published in 1994Abstract: Traditional dual copy, CAID, and RAID DASD subsystems can offer improved data reliability in cases of actuator and/or media failures. However, these schemes impose a write penalty for the extra I/Os required to maintain image copies or parity information for data contained in the subsystem. In this paper, we will employ a review of the 3990-3/6 read and write data flows as a basis for discussing write penalties and storage overheads associated with these extended count key data (ECKD) CAID and RAID DASD subsystems. |
||
Comparing Processor Performance: How to Specify What You Want and Determine if You Got ItDr. H. Pat Artis - Published in 1993Abstract: A methodology is presented which can be used to quantify the relative performance of two processors running a production workload. It is based on Major's CPU-I/O equation and his notion of relative I/O content. Using this methodology, a customer can determine whether he got what a vendor promised or what he thought the vendor promised during a processor upgrade. More important, a customer can incorporate the methodology in an acquisition contract to provide measurable performance criteria. |
||
DASD Subsystems: Evaluating the Performance EnvelopeDr. H. Pat Artis - Published in 1993Abstract: As the complexity and size of new high performance subsystems has increased, it has become more difficult for users to evaluate their performance characteristics. In this paper, we will introduce an I/O driver that can exercise a DASD subsystem over a wide range of I/O rate, skew, locality of reference, and read/write ratio values. The evaluation of an EMC Symmetrix 5500 using the I/O driver will be employed as a case study to introduce the testing methodology. |
||
Quantifying MultiProcessor OverheadsDr. H. Pat Artis - Published in 1991Abstract: As a general rule, most performance analysts are confident in fundamental relationships like one plus one equals two. Unfortunately, this time proven relationship is a gross over estimation of the actual capacity delivered by dual or multiprocessor configurations. In this paper, we will present a simple conceptual model of multiprocessor performance and provide a generalized first order result for evaluating the capabilities of alternative multiprocessor designs. |
||
Workload Characterization Algorithms for DASD Storage SubsystemsDr. H. Pat Artis - Published in 1990Abstract: During the implementation of DFSMS, the user must develop a variety of constructs and policies that broadly describe physical data set characteristics as well as archive and migration policies for their management. This tutorial will present an algorithm for analyzing an existing DASD storage subsystem to identify potential data classes to support the development of data classes for converting an existing storage subsystem to DFSMS control. |
||
Optimal Storage Allocation for Partitionable ProcessorsDr. H. Pat Artis - Published in 1990Abstract: Processor partitioning mechanisms such as MDF, PR/SM, and MLPF allow a user to partition the resources of a physical processor to create several logical systems that meet the needs of the user's workloads. While the distribution of the processor's CPU is primarily a political question and the allocation of the processor's channels are dictated by the characteristics of the I/O configuration, the allocation of the processor's real and expanded storage is a complex question. This paper will present an algorithm for allocating real and expanded storage based on the hypothesis that an optimal allocation of the processor's storage resources has been achieved when the percentage of the processor's CPU resource used for memory management is minimized. |
||
Data Set Allocation and Performance Management Under System Managed StorageDr. H. Pat Artis - Published in 1990Abstract: System managed storage (SMS) redefines data set allocation for the MVS operating system and provides new data set level performance management capabilities that exploit advanced cached control unit functionality. Unfortunately, the algorithms and measurement data sources that support these DFSMS functions are poorly defined in the available documentation. This tutorial will provide a detailed description of the execution of ACS (Automatic Class Selection) routines, SRM (System Resources Manager) interface, device selection, volume free space management, inter-system communication, data set level caching, and interfaces with the 3990-3 control unit. |
||