Virtualizing ERP Infrastructure Using SEM Extended n-Tier Workloads

Many applications are combinations of multiple workloads — Gartner calls this a "technical pattern." ERP is a classic example. A typical Oracle or SAP implementation will include at least one OLTP DBMS serving workload, plus multiple Web and application serving workloads. These various modules and functional capabilities work in collaboration to create a highly customized implementation. This pattern of at least three individual workloads — which become related and interdependent on each other as part of the user's experience — is what we call in SEM terminology an "n-tier pattern." Figure 1 is based on the output of Gartner's enterprise architecture (EA) research, and shows how an n-tier pattern would include a minimum of one Web serving, one application serving and one DBMS serving workload. Typically, however, patterns extend to multiple instances of each workload, and increasingly include optional integration with business intelligence (BI) workloads — which creates the concept of the extended n-tier pattern.

Because the strengths of each workload favor different hardware and OS combinations, the traditional homogeneous nature of n-tier applications is gravitating toward heterogeneity in terms of servers. Where they are not built exclusively on x86 and Windows, new ERP projects routinely include some sort of Unix/Linux or other mixed combination, and this trend will accelerate as new application infrastructure developments, such as NetWeaver and Oracle Real Application Clusters (RAC) adoption, challenge legacy buying behavior.

The n-tier trend is further driven by new "unified" server architectures of servers, storage, networks and solutions, from the likes of Cisco and HP, plus availability of verticalized ERP appliances, like Fujitsu's FlexFrame for SAP.

The database tier is generally perceived as the most business-critical component of an n-tier pattern, and it is consequently a workload where a more risk-averse approach to server and OS selection is commonplace. This has created a historic preference for RISC-, Itanium- and mainframe-class servers, and x86 platforms tend to score lower. However, with every new generation of the SEM, the gap is narrowing. A typical n-tier pattern may host hundreds of Web servers, tens of application servers and probably only a handful of DBMS servers. Couple this with the largest code paths residing in the DBMS server tier, and we have the majority of performance and availability requirements at mission-critical levels and superior I/O requirements. Thus, OLTP DBMS serving represents one of the most conservative server selections that organizations make. SEM leadership favors proven, high-end Unix and mainframe platforms that are functionally rich and usually expensive to buy. But these servers deliver reliable performance, good functionality and well-honed platform management that helps balance the investment over time.

New dynamics are redefining the market as less-expensive x86 servers inexorably establish their OLTP DBMS serving credentials. Vendors like Teradata have already established the viability of x86 as the foundation for DW DBMS serving. The recent financial crisis has driven more users toward alternative platforms to save costs, further driving adoption. Short-term buying decisions still favor the Unix/mainframe approach, but x86 servers running managed Windows (typically with SQL Server, but potentially MySQL) or Linux (typically with DB2 or Oracle) have started to penetrate the Top 10 list of OLTP DBMS serving choices (see Figure 2). Further pressure to adopt a node-based approach instead of the big symmetric multiprocessing (SMP) option will come from rapidly increasing trust in Oracle's RAC strategy for DBMS scale-out clustering. The potential for x86 will grow further as other database vendors promote their own parallelized strategies to compete with Oracle. The latest developments in x86 server design will even benefit users who prefer the more traditional scale-up approach to DBMS serving. The latest-generation x86 servers are now offering scaling to 64 cores and more, and up to 3TB of memory; these capacities will address the majority of situations where the perceived scaling superiority of RISC, Itanium or mainframes has been a factor.

Figure 2 is a simplified chart showing the 10 highest "best of breed" results for OLTP DBMS serving. The SEM evaluates multiple server platforms from nine different vendors. For each platform, there will usually be more than one OS choice, and always a variety of software stack choices. For some platforms, there are also variations in processor and form factor choice. For each individual platform, the SEM selects the strongest score for each of these variations, and then stack-ranks the platform alongside its peers. Where there are preferences for certain hardware/software/stack combinations (or prejudices against them), the SEM is available to qualified clients via gartner.com either as an online tool (which provides a programmable dashboard that allows options to be included and/or excluded at will) or as a downloadable Excel-based toolkit.

Unlike the Web and application serving tiers, the three strongest OLTP DBMS serving results are all achieved by RISC, Itanium and mainframe platforms; but x86-based servers are steadily improving their penetration. The top-ranking x86 server (HP Proliant running SQL Server) is now in fourth place, up two places from last year. There are now five x86-based servers in the top 10, up from three in 2009. But the top three platforms still demonstrate a measurably stronger score that echoes continued market confidence in RISC, Itanium and mainframes for the most demanding and business-critical OLTP DBMS workloads.

Figure 3 shows the full set of OLTP DBMS serving results, again working on a best-of-breed approach. This figure is a direct extract from the SEM model that can be downloaded as an Excel spreadsheet or accessed as an online tool or related Toolkits.

Figure 3 confirms that the strongest choice for OLTP DBMS serving is the IBM Power System based on AIX and Oracle RAC. The data was finalized in late 2009, so this latest SEM assessment is based on Power6, and not Power7, and the competitive positioning among IBM, HP and Oracle-Sun Microsystems was based on the market situation in 4Q09. If anything, IBM has strengthened its Unix market hold since 4Q09, and the probability is high that the next SEM update will show IBM Power System as the strongest choice, potentially by an even wider margin.

In recent years, BI and analytics have become more strategic to most IT organizations, driving a desire to minimize the latency between live OLTP DBMS data and the data stored in DWs. As a result, DW DBMS serving is now more closely aligned to line-of-business operational workloads. Figure 4 shows the full set of DW DBMS-serving results, working on a best-of-breed approach.

As with OLTP DBMS serving, the DW DBMS-serving workload has historically favored Unix, mainframe and other large-form-factor designs that are well-proven, DBMS-serving foundations. However, because of the previously stand-alone nature of the workload, a market has developed for more-specialized platforms, including Teradata, Netezza and HP NonStop solutions, that use massively parallel processing (MPP) technology to achieve nearly limitless theoretical scaling. As DW DBMS serving becomes increasingly business-critical to smaller organizations, large, expensive enterprise DWs become more difficult to justify. This has spawned more-focused appliances that address specific submarkets, such as data marts, DWs and business analytics workloads, which can require extreme scaling or extreme performance.

The hybrid nature of DW DBMS serving creates a more complex mixture than we see for other workloads. Teradata holds the strongest position (as it has throughout the life of the SEM) with its x86-based MPP platform. However, Teradata's strongest scores are now achieved with Linux, and not Unix, as the OS. RISC and Itanium Unix servers from IBM, HP and Oracle-Sun are also well-represented in the top 10 scores. However, a small majority of the top-10 results are x86-based platforms, showing that x86 penetration of the DW DBMS market is more advanced than it is in the OLTP DBMS serving market.

The choice of DW DBMS then becomes even more complex in an ERP environment, as DBMS solutions like Teradata and Netezza are targeted only at BI workloads, while DBMS solutions from IBM, Microsoft and Oracle are typically used as the OLTP DBMS foundation. Furthermore, these other DBMS products are more generic, and therefore capable of being deployed for both OLTP and DW workloads. This creates a challenge of best of breed versus "fit for purpose" in many new ERP deployments.

Users must remember that OLTP DBMS serving is just one of four workload types in the extended n-tier pattern. Within an n-tier workload, many architectural similarities exist between Web server and application server workloads, where small form factors, such as blades or rack-mounted devices, are dominant in most new businesses. Web and application servers are generally single-function devices that are deployed in an "appliance" manner (that is, standard, repeatable configuration) for prescriptive workloads. The software stack is based on standardized software components, which makes independent software vendor (ISV) enthusiasm and the size of the available software portfolio key requirements. The hardware demands are generic, which enables a wide range of vendors and platforms to be viable choices. Users will value vendors that can execute strongly in local markets and demonstrate good ISV partnerships.

Web and application serving workloads are usually more highly standardized and deployed in a replicated manner, so that failures on one server do not bring down an entire network of servers. Therefore, individual servers will rarely be regarded as mission-critical servers, although the overall "farm" of servers will fulfill this requirement. This explains why more-commoditized x86 servers running Windows or Linux are generally suitable (and preferred) for most new deployments, and the usage of other architectures continues to decline because of concerns over costs and standards. Unix servers are still a viable choice for this workload, particularly where installed-base users favor a homogeneous approach to n-tier workload deployment. But Web and application serving (and other workloads that are favoring smaller commodity form factors) will drive ever-increasing erosion of Unix's share in the coming years. As an example, Figure 5 shows the full set of Web serving results, working on a best-of-breed approach.

Although the IBM Power System server (and other RISC or Itanium Unix servers from Fujitsu, HP and Oracle) has viable positioning as a Web or application server, the natural leadership is lost for these workloads as market momentum for new business increasingly favors x86. RISC and Itanium architectures (and, by association, the Unix OSs) offer excellent vertical scaling, manageability and uptime advantages against x86 servers, but four market forces are making these differentiators increasingly latent:

The degree of Unix leadership has been diminishing steadily across all workloads, as x86 architectures and OSs constantly improve and "close the gap."
The evolving requirements of most Web and application serving workloads do not need these attributes in order to deliver lasting business value.
x86/Linux represents the development environment for ERP vendors, such as SAP and Oracle, and x86/Linux is the fastest-growing installed base.
Windows is left as the primary port with the biggest installed base, with Unix/RISC as a secondary port, leaving traditional mainframe and other options as niche platforms — with fringe support of the ERP applications (although these platforms may be worthy DBMS platforms in mixed environments).

Either way, RISC/Itanium Unix remains a technically competent option for Web and application serving, but becomes increasingly overengineered for most requirements. It is, therefore, no surprise that we now see that the end-to-end OS platforms growing the installed bases are all Windows and all Linux among ISVs and ERP vendors.

While the SEM is a valuable tool to analyze the strongest results by workload, it delivers its greatest value to organizations that can specify their software stack preferences or prejudices. The ERP pattern is an excellent example of how the SEM results will vary greatly according to stack strategy. Once a strategic choice of DBMS and Web/application serving stack have been made, the choice of processor and OS is at least influenced — and often predicated — by this action. Figure 6 shows how the SEM6 application serving results are skewed when SAP NetWeaver is "forced" as the only application serving choice.

While SAP NetWeaver is certified for multiple OS environments, the most proven deployments generally favor Windows and x86, with VMware as the most favored virtualization layer, where appropriate. Contrast this with Figure 7, which is an identical analysis of the SEM database — differing only in the choice of Oracle Fusion, rather than SAP NetWeaver.

While the hardware choice changes very little, the choice of the Oracle Fusion stack mitigates against Windows and leads to a strong Linux bias for most platforms, again favoring VMware where virtualization is appropriate. Remember that Oracle's whole Fusion strategy has evolved greatly since the SEM6 data was finalized in 4Q09, with the integration of BEA Software's software suite and now Oracle-Sun's software assets.

The SEM is updated once a year, and the data that supports the latest version (SEM6) was finalized in 4Q09. As a result, the platform comparisons pre-date several more-recent events, such as:

Power7 introduction in 2H09
Oracle acquisition of Sun in 1Q10
New-generation HP Integrity launch in 2Q10

A new SEM version will be released in 4Q10 or 1Q11. This will supersede the data in the current version, and we will incorporate this and other market events into the new version.

Many organizations have been deploying ERP systems for a decade and more, with many users dating deployment back to the days of the material requirements planning (MRP) era; as a result, ERP modernization projects can be the second or even third major system redesign for many users.

Until only a few years ago, most vendors promoted as much homogeneity as possible for new ERP deployments. Rationalizing the ERP environment to the most centralized and standardized environment simplified the task of implementation and support for hardware and ERP vendors alike. There are many benefits to a largely homogeneous deployment, such as simpler support infrastructure, fewer vendor relationships, easier integration of different modules, etc. While x86 platforms delivered visibly inferior scaling and/or OS functionality, it was relatively easy for the market to maintain commitment to Unix (and even mainframes) for a business-critical workload like ERP. But as we have already demonstrated, the justification for RISC/Itanium Unix homogeneity for new ERP workloads is becoming more and more difficult to maintain.

Indeed, where there is a justifiable business case for a single architecture and OS across a whole ERP deployment, this is more likely to favor x86 (and more often Windows than Linux) today. Windows has been the top-selling OS choice for SAP users for at least six years, and the vertical scaling of SQL Server on new-generation Intel 7500-based servers is likely to meet the majority of OLTP DBMS serving needs. So Windows end to end is good for all Microsoft organizations that are rationalizing to Microsoft platforms, frameworks and technologies. With the large ERP/ISV installed base, this enables new application content to be adopted promptly. However, this one-stop-shop approach has the disadvantage of "all the eggs in one basket" — a cliché, but one that may restrict IT organizations' negotiation power and conflict with the long-term direction of the ISV/ERP vendor. This ISV loyalty will be tested on the introduction of Longhorn, the rate of adoption by vendors and users, and any delivery of Microsoft's virtualization promises.

Although most Windows momentum favors the use of SQL Server, we see continued deployments of Oracle in select situations. While most Oracle deployments will favor smaller designs, we also see vendors such as Unisys successfully deploying Oracle on much-larger Windows platforms.

Linux DBMS servers can be considered for ERP and ISV DBMS workloads that enable the sharing of x86 platforms for multiple workloads. However, for high-end workloads, Oracle RAC or Unix RISC in a mixed environment may have to be considered tactically for three to five years. The ISV and ERP vendor may also promote open-source software (OSS) as a reduced-cost alternative to commercial DBMSs, but this may be appropriate only for small workloads and small or midsize business (SMB) targets. The range of x86 ERP advantages and challenges is outlined in Figure 8.

Alternately, RISC and Itanium Unix remain excellent choices for the DBMS tier of an SAP or Oracle E-Business Suite (EBS) n-tier pattern. They are then viable choices (but not the optimum choice) for the Web and application server tiers. The choice, therefore, comes down to whether the organization prefers a homogeneous or heterogeneous (hybrid) solution.

An all-Unix solution will be generally simpler and is best-suited to maintaining legacy system trust and familiarity. Acquisition (capital) costs are likely to be more expensive than a hybrid solution, but lifetime (operating) costs may be cheaper. The resulting hardware infrastructure should be easier to manage. A hybrid solution (Unix at the back end, with Windows or Linux on x86 at the front end) is undoubtedly more complex, but provides access to inexpensive x86 hardware for part of the solution, and is better-suited to the evolution of the market. The options are presented in Figure 9.

While few new ERP deployments would justify the role of a mainframe, in reality IBM System z would be well-suited to the same deployment options where we see Unix being viable. It would also share most of the same challenges, such as being overengineered for the Web and application tiers. But we will follow the progress of IBM's new zEnterprise strategy closely. Through its ability to host native Linux or AIX workloads on x86 or Power-based blades, the latest generation of the System z is capable of becoming a complete n-tier pattern in its own right, even with the ability to support hybrid workload combinations of z/OS, Linux (either CMOS or x86-based) and AIX. Architecturally, such a platform would run an n-tier pattern like ERP in a very holistic way. But with x86 blade support not due before 1Q11, we will not be able to observe this behavior until 2011, at the earliest. This will also inhibit the potential of Z/Linux application hosting as ERP vendors look for volume-mixed environments with mainframe DBMS and Linux/x86 application servers.

A new variable has emerged that organizations must consider when planning the deployment of ERP, or any other workload: the potential to deploy the workload on a physical server, a logical combination of consolidated servers or a virtual machine. The fifth generation of the SEM evaluated virtualization and consolidation as separate workloads, but this methodology did not enable us to demonstrate how some platforms will behave better for one deployment type than another. In SEM6, we introduced three deployment types — virtual, logical and physical — that can be assessed across all workloads.

As a result, instead of treating virtualization and consolidation as a form of hybrid workload, we have assigned weighting factors to all six workloads that enable us to recognize how each workload is best deployed. Three deployment types are supported: virtual, logical and physical.

Virtual refers to any situation in which the workload runs on a single virtual server that is abstracted from the physical platform(s) on which it resides. In 2009, there were significant increases in the adoption of virtualization across all server platforms, and this has continued in 2010 to the extent that virtualization is often regarded as appropriate for every workload. Although Gartner believes that product enhancements and maturing business practices are constantly increasing the proportion of workloads that can be run in a virtual environment, many workloads are still best addressed by a physical or logical approach. Hence, data centers face a key challenge to "right size" workloads for different deployment types by understanding application behavior. The latest SEM generation can assist users in this effort.

Because ERP is not one workload, but rather an example of a pattern of workloads, it becomes even more important to right size any new project to reflect multiple market forces that are not always complimentary with each other.

As a global trend, we see more and more users who are updating legacy ERP environments favoring a more hybrid approach — retaining RISC/Itanium Unix (or even a mainframe) at the database tier, but deploying x86 elsewhere. There will, however, be exceptions to this rule. For instance, we see proportionally stronger enthusiasm for RISC and Itanium Unix throughout the Pacific Rim countries. Where organizations have more in-house expertise in RISC or Itanium Unix, and are concerned about the support aspects of a hybrid solution, it can still make better sense to go with a homogeneous RISC or Itanium solution.

Justification of a homogeneous approach becomes harder and harder to justify. Where homogeneity does make sense, it will more likely be at the level of x86 and Windows or Linux, rather than an all-Unix solution. Market belief in the hardware cost of ownership advantages of x86 servers is unshakeable, and this has helped drive consideration of a hybrid approach into the most risk-averse of data centers.

ERP buyers now face new pressures that also force an architectural rethink. Where an organization is embracing new-generation Web/application serving software, like SAP NetWeaver or Oracle Fusion, rather than just refreshing an existing ERP implementation with minimal change, there is a new need to prove the viability of the new stack components — particularly on older, legacy OS environments where the stack might be certified, but there is little new business.

All users planning an ERP refresh must consider the role and potential of virtualization, even though it is rare for an entire ERP pattern to run under virtual machine (VM) control. As with all n-tier environments, virtualization has its most obvious opportunity in the Web and application serving tiers. Virtualization of OLTP DBMS serving is at a more nascent stage, and virtualization of DW DBMS serving is positively embryonic. But a growing number of ERP users will consider virtualized databases, particularly in multi-DBMS scenarios where some databases will be smaller and less business-critical than others.

© 2010 Gartner, Inc. and/or its affiliates. All rights reserved. Gartner is a registered trademark of Gartner, Inc. or its affiliates. Reproduction and distribution of this publication in any form without prior written permission is forbidden. The information contained herein has been obtained from sources believed to be reliable. Gartner disclaims all warranties as to the accuracy, completeness or adequacy of such information. Although Gartner's research may discuss legal issues related to the information technology business, Gartner does not provide legal advice or services and its research should not be construed or used as such. Gartner shall have no liability for errors, omissions or inadequacies in the information contained herein or for interpretations thereof. The opinions expressed herein are subject to change without notice.