Blade Computing.Doc 1 ©2003-04 IMEX Research the Rise of Blade Computing

An Industry Brief

Modular Computing Brief Series 2004

[coverpage]

the rise of

Blade

Computing

By Anil Vasudeva, Principal Analyst & Founder, IMEX Research.com This article appears in BladeLetter by permission of IMEX Research. Excerpted from “Blade Servers Industry Report 2003” ©2003 IMEX Research. All rights reserved. The rise of Blade Computing...... 3 Deficiencies of Current Architectures...... 3 CISC vs. RISC...... 4 HW vs. SW Provisioning ...... 4 The evolution of Blade Architectures ...... 4 Fabric Architecture for Blades...... 5 Shared I/O architectures ...... 5 Growth Drivers for Blade Servers ...... 6 Towards de facto standardization...... 6 Market Segmentation ...... 7 Market Segmentation by Tiers ...... 7 Tier-1 Edge Computing...... 7 Tier 2 Application Servers...... 7 Tier-3 Back End Data Base Servers ...... 7 Market Requirements by Tiers-1, -2 & -3 Computing...... 8 Market Requirements in Tiered Computing ...... 8 Sizing of I/O Traffic in Tiered Computing ...... 8 Market Segmentation by Applications ...... 9 Market Segments by Technology...... 10 References...... 11 Acknowledgements ...... 11 Key Blade Server Technologies...... 12 Key technologies in Computing & Communications Blades...... 12 State of Progression of Key Technologies...... 12 IP SAN...... 12 Status of iSCSI...... 12 Industry Initiatives ...... 13 Cisco, Network Appliance, qLogic ...... 13 Microsoft Initiative - Chimney...... 14 TCP Chimney...... 14 RDMA Chimney ...... 14 Technology Challenges...... 15 RDMA ...... 15 Market Acceptance of IP SANs ...... 16 Players in Blade Servers by Market Segment...... 16 Vendors by Technology Segment ...... 16 The Future of Data Centers ...... 18 Emergent Technologies driving Data Centers of the Future ...... 18 Server provisioning software for Blades...... 18 Investment opportunities in Servers...... 19 Definitions...... 20 References...... 20

Deficiencies of Current Architectures Pressured by CEOs to “do more with less” in this challenging economic environment, interest in server consolidation has soared unabated. And with average CPU utilization in many large companies running around 25% while stranded direct-attached storage assets being utilized at 50%, improving the productivity of its assets has become a top mandate for CIOs in many companies. This low resource utilization has been, in large part, due to the performance and flexibility deficiencies of current server architectures wherein most servers are I/O rather than compute bound.

IT Managers have always drooled at Clients (End Users) the cost effectiveness of Windows and Linux based servers but the lack Applications of support for multiple workloads concurrently within the operating System Software system have kept them away from deploying them in mission critical OS Discovery, Configuration environments. In contrast, the Monitoring,Visualization mt

Mainframe and Unix based systems g Provisioning that they already own have

Services Security sophisticated software tools providing hardware level Billing provisioning. This enables them to Clustering System Utilities Network Services System Services Resource M deploy several independent Process Fail-Over workloads in each operating system instance, with each workload Infrastructure Elements executing in a separate virtual server environment.

CIOs are willing to invest in new I/O Power Storage Memory Network IF - Fabric technologies so long as deploying Processors them provides them the benefits inherent in cost-effective industry standard, interoperable IT gear. But with ad hoc proliferation of servers within the organizations and each server requiring a multitude of connections such as several power cords, multiplicity of adapters for primary and backup storage, networking, keyboard, video, and mouse outlets, the process of provisioning server resources has become complex, human resource intensive, error prone and very disruptive.

The ultimate cost reduction opportunity for servers, however, lies in the elimination of monolithic, proprietary server environments altogether and instead augment the industry-standard server operating systems from Microsoft and Linux community with advanced access, availability, and administration functions emulated or replicated from proprietary mainframe and Unix environments and implement them on scalable and modular computers.

The ability of new generation blade servers to be scalable (incrementally added), to manage shifting peak workloads through software provisioning versus the inflexibility of large UNIX based SMP servers dedicated to certain types of workloads puts them at a distinct advantage for adoption in various environments from small businesses to large enterprises as well as system integrators and service providers.

-012- Rise of Blade Computing.doc 3 ©2003-04 IMEX Research CISC vs. RISC Mainframe and proprietary, RISC-based Unix environments typically allow provisioning of processors, memory, and I/O resources across a large number of concurrently executing processes. As a result, IT staff can consolidate multiple applications and services onto a single server, thereby dramatically lower operating complexity.

The price of these high-end Unix systems, is typically influenced by several additional value- added silicon, system, and software features that offer dramatically improved server access, availability, and administration features relative to commodity Intel servers running industry- standard operating systems from Microsoft or the Linux community.

However the economic significance of Intel’s push into 64-bit computing should not be underestimated. HP and IBM have committed to migrating their RISC-based server environments to IA-64. Sun Microsystems ultimately will follow the other vendors in adopting the IA-64 architecture. Sun Microsystems may be the last holdout but the ongoing price/performance improvements of Intel’s designs relative to the SPARC platform (despite their rumored outsourcing of high end Servers to Fujitsu Japan) will leave Sun with no visible alternative but to embrace it.

Intel’s new 64-bit architecture with advanced features rivaling RISC platforms promises to relegate the proprietary RISC platforms to the annals of enterprise computing history, over time. The price/performance position of Intel’s 64-bit computing should find a receptive ear from the current “do more with less” mentality of IT personnel.

HW vs. SW Provisioning One of the key contributions of value-added core logic solutions relates to hardware enforced provisioning. Hardware enforced provisioning involves the segmentation of server resources at the processor, multiprocessors, or board level. Each of these segmented hardware resources then may be allocated to one of multiple operating system instances. This form of provisioning provides a high degree of fault tolerance, simplifies the hardware upgrade process, and may allow hardware at different performance levels running diverse applications to co-exist in a single system or chassis. This is the basis of blade architecture.

Intel server designs will end up replicating the provisioning capabilities of mainframes and proprietary Unix platforms to drive greater adoption in large application and database systems. In the future expect Fujitsu, IBM, Hitachi, HP, NEC and Unisys to apply their system design expertise to create hardware level provisioning to create differentiated platforms on industry- standard processors.

The evolution of Blade Architectures

The lack of sophisticated provisioning capabilities has so far been a key inhibitor of adoption for Intel servers in business critical environments. Now emerging “blade server” platforms offer some further relief for the Intel ecosystem.

Blades facilitate convergence of computing (channels) and Networks (nodal interconnectivity) for delivery of transactions and content between modular system nodes sharing resources to achieve low cost computing. This is accomplished by use of standard high volume modular servers while mitigating scalability and low reliability bottlenecks of existing monolithic von-Neumann computer architectures of today.

-012- Rise of Blade Computing.doc 4 ©2003-04 IMEX Research Servers are packaged as one-, two-, or four-way Intel CPU “blades”, and these blades may be added or removed from the system through a single connection to a highly integrated back plane, (called midplane in blades parlance) dramatically simplifying cable management.

Major vendors offering such configurations include Dell, Egenera, Fujitsu-Siemens, Hewlett- Packard, IBM, and RLX Technologies and Sun Microsystems. Earlier blade offerings, such as those from HP, Dell, Fujisu and Sun, targeted the “stateless” workloads in the front-end web access and presentation tier. Mem Mem Mem On the other hand IBM, Intel, Egenera and second-generation blades from HP directly CPU+ CPU+ CPU+ targeted the heavy-duty enterprise CS CS CS workloads for application and associated database servers in backend tier as well.

I/O Disk I/O Disk I/O Disk

Fabric Architecture for Blades

The first generation blade servers rely on captive, bus-oriented connections for Ethernet Fabric storage and network I/O wherein the (Passive Backplane) operating system assumes complete control of the connected storage and network I/O resources at the expense of balancing server, network, and storage resource utilization or their scalability.

Data Center managers, unable to predict the performance from a given I/O capacity in place end up throwing new servers to scale aggregate performance when the network or storage I/O connection becomes saturated, despite the fact that server processors remain underutilized. Shared I/O architectures (a concept borrowed from the mainframe and Unix server markets) such as those promised by blade architectures eliminate these concerns by decoupling the I/O complex from individual servers. Multiple server blades share storage and network I/O resources and are provisioned with minimal effort, to individual or multiple blades as needed. Shared I/O architectures Future blade server options are likely to employ fabric back plane designs that support shared I/O configurations. Mem Mem Mem

Most Intel server platforms, including CPU+ CPU+ CPU+ first generation blade servers, rely on CS CS CS captive, bus-oriented connections for storage and network I/O. In these configurations, the operating system Modular Backplane assumes complete control of the connected storage and network I/O resources. This presents challenges when ial r

attempting to balance server, network, FC Enet and storage resource utilization. Shared Se I/O architectures eliminate these concerns by decoupling the I/O complex from individual servers.

-012- Rise of Blade Computing.doc 5 ©2003-04 IMEX Research Some of IBM’s latest Intel platforms offer a Remote I/O feature that allows multiple Intel servers to share I/O resources located in a separate I/O adapter rack. Egenera uses an internal SAN fabric as the back plane for its BladeFrame architecture for concurrent access to shared I/O resources.

Growth Drivers for Blade Servers The Data Centers migration is underway to leverage the volume driven economics of industry standard servers using Linux/Windows operating environment. The integrated modular architecture of blade systems providing Virtualization, Provisioning and Self- Driven Automation capabilities already are starting to strike a pleasant resonant chord with CIOs.

Some of the major growth drivers accelerating this phenomenon include: • UNIX to Linux Migrations to leverage cost effectiveness of open source Linux • HPC Linux Clusters have struck an explosive growth in academia, national laboratories for scientific computing and migrating to commercial world for Bioinformatics, decision support financials, visualization etc. • Server Consolidation onto fewer servers with virtualization of resources, provisioning and lights out automation capabilities • Blades Density to provide economic benefits in real estate and operational environmentals. • Standardization on fewer OS types for servers, storage and network protocols deployed.

Towards de facto standardization Improved hardware provisioning provides a compelling value proposition for blade servers, but lack of standardization is undermining this message. Each server vendor initially has attempted to create proprietary designs to protect margins and secure long-term footprint in data center configurations. This lack of standardization is impeding adoption, as customers want standard form factors supported by multiple vendors.

IBM has made Blade servers as a cornerstone of its x-series (Intel) product line and is heavily pushing its BladeCenter along with larger SMP computers. To mitigate concerns of non- standardization and to leapfrog other competitors IBM chose to parlay its relationship with Intel to create a partnership in which the two companies share technology and development costs for blade server designs and drive a de facto standard. IBM is hoping that the partnership will drive the commoditization of blade server hardware, allowing the company to win the day on its superior software and services. In exchange, Intel got a faster ramp in blade server designs and the blade market and, in turn, allows Intel to drive volume of its Xeon (and later Itanium) processors into higher performance, higher margin enterprise markets.

IBM has taken considerable share in the blade server market in recent months. It has now become the new leader in blade server revenues in 2Q2003, according to IMEX Research, a San Jose CA Technologic Research & Consultancy. (See Blade Servers Industry Report 2003 – IMEX Research Doc.BS-203 Sep2003)

Intel is now selling its BladeCenter “clones” to other OEMs. It has signed 12 OEMs and System Integrators worldwide including Bull Computer etc. Indications are that Dell and HP are giving strong consideration to a license agreement with Intel for the BladeCenter design in order to accelerate their blade server roadmaps.

Market Segmentation by Tiers Blade Servers are architected to be flexible for software provisioning to match the requirements of applications. AS such it is easy to address the various specialized requirements of Tier1, 2 and 3 applications in Blades.

3 Tier Computing Architecture

Web App SAN Storage Network Storage DB Network Storage Web Web App SAN Internet Web SAN Edg DB e Core Web Optical App Web NAS Caching, Proxy, FW Security, IDS, DNS Web DB Load Balancing … App NAS Web Servers Tier-1 Tier-2 Tier-3 Edge Computing Applications Data Center

Tier-1 Edge Computing Segment Front End & Web Servers Markets ISPs, Low End Bus Addressed Applications Web Servers, Infrastructure … Work Loads Caching, Directory Services, DNS, Proxy, Firewall, Security, Intrusion Detection Servers, Low End File/Print, Low End Messaging … Tier 2 Application Servers Segment Back End - Mid Range Applications Markets Small/Medium Businesses Addressed Applications Application Servers, Infrastructure, High Performance Computing, Streaming … Work Loads Midrange OLTP, ERP, SCM, CRM, File/Print, Messaging, HPC Linux Clusters … Tier-3 Back End Data Base Servers Segment Back End – Data Base Applications Markets Enterprise Businesses Addressed Applications OLTP/DB Servers, Numeric Intensive Computing, Decision Support … Work Loads High End Data Sets, OLTP, ERP, CRM, SCM, High End SMP/HPC Clusters

Market Requirements in Tiered Computing Tier-1 Tier-2 Tier-3 Applications Front End – Back End - Back End – Web Apps Mid Range Apps Data Base Apps Markets ISPs, Low End Bus Small/Medium Bus Enterprise Businesses Typical Applications Web Servers, Application Servers OLTP/DB Servers Infrastructure … Infrastructure Numeric Intensive High Perf. Computing Computing, Decision Streaming … Support … Work Loads -Caching, -Directory MidRange – OLTP, – High End Data Sets, - Services, -DNS, - Proxy, ERP, - SCM, -CRM, - ERP, CRM, SCM - Firewall, -Security, File/Print, - Messaging HE SMP/HPC Clusters - LE File/Print, - LE - HPC Linux Clusters … Messaging … Performance Connections Intensive Transactions Data Access & Movement Chassis* 1-2U 2-4 U 4-8U CPUs/Blade* 1 - 2 way 2 - 4 way 4 - 8+ way CPUs* LV Pentium/ Pentium/Athlon/ Xeon/Opteron/ Embedded U-Sparc2/IBM U-Sparc3/IBM Storage* HE RAID & Backup I/O Requirements* - Network Traffic Very High** Low Very Low - IPC Very Low High High - Storage High Very Low High Manageability OS based Add-On Tools Proprietary Tools Typical Examples Fujitsu-Siemens BX 300 HP BL 40p IBM Blade Center HS 20 Dell 1600 MC Egenera Intel HP BL40e Egenera *Typical Registered names and trademarks belong to respective companies **Legend Sizing of I/O Traffic in Tiered Computing I/O Traffic Very Low Low Medium High Very High Gbps < 0.1 Gbps 0.1-0.5 Gbps 0.5-1Gbps 1-5 Gbps > 5 Gbps

-012- Rise of Blade Computing.doc 8 ©2003-04 IMEX Research Market Segmentation by Applications Web Infrastructure – Caching, Firewall, DNS, Security (Secure Sockets, Data Compression, Encryption..), VPN, Load Balancing, Web Content Acceleration, Directory Services, Proxy Servers, … Data Center Infrastructure - File/Print, Data Backup & Restore, Resources Allocation & Administration, Data & Network Management, System Utilities … High Performance Computing – Clustering, Parallel Processing, Numeric Intensive Computing, Supercomputing … Online Transaction Processing – OLTP, SCP, CRM, ERP … Decision Support – Application Development, Business Logic, OLAP, Business Intelligence, Data Warehousing … Streaming – Data Acquisition, Medical/Bioinformatics, Video On Demand …

Applications in 3-Tier Computing

n-way

4-8+ , s e n i M l r R i P C A , … , T g e M n c i r C L k e S n m , a O P , , B m e R / o P c e A E C S n c L e g e O g n n S , li i a c l s n D i te u 2-4 i g n o F o I h L s re iz s a B e in W s ta u a B D n tio C g, za I rin ali N te su us Vi ing Cl g/ ut C lin p Messaging P de m H o rco M pe ific u nt S ie MP 1-2 Infrastructure Sc S Web ming trea ing, Print/File S n, Imag quisitio rencing ta Ac t/Confe Da oadcas tics, Br nd.. Informa nDema VideoO

Market Segments by Technology Fundamentally volume driven economics of industry standard servers is driving the scale out architectures for servers. Software Provisioning technologies used in single server are giving way to hardware provisioning of cluster of low cost nodes. Technologies such as Clustering, Back Plane Fabrics (e.g. using Infiniband or TCP Offloaded IP to achieve low latency, high bandwidth), Virtualization/Provisioning Software, High Density/ Power Management, High Availability architectures, etc.

The platform market can be segmented in three distinct areas based on Scale Up and Scale Out technologies deployed: o Large SMP Servers suing Software Provisioning for Parallel Processing o Cluster Architectures for High Availability or High Performance o High Density, Highly Integrated Blades with non-blocking Interconnect Fabrics

Server Market Segmentation - Platforms

)

Large SMP Parallel Sysplex Clusters •Scalable Nodes • Dynamic LPAR •SW Provisioning High Availability/ IBM: i890/p690/z990 SUN: SunFire 4800,12K Hi Perform.Clusters • Linux/AIX/Solaris Clusters • HPC Clusters DELL: PowerEdge 6650 ORACLE: 9iRAC/PE6650 (Consolidated SMP Servers

Blade Servers High Density Rack Mount • Highly Integrated Dense FormFactor • Rapid Deployment, FlexibleArchitect

Scale Up RLX: ServerBlade 300i,800i, HP:BL40p FUJITSU: BX300, DELL:MC1655, IBM: HS20, Sun:

Scale Out (Clustered Blade Servers)

-012- Rise of Blade Computing.doc 10 ©2003-04 IMEX Research References Report: Blade Servers Industry Report 2003 Document RPT-107 (Aug 2003) Brief: UNIX versus Windows based Server Architectures. Document BRF-112 (Feb 2003) Report: Competitive Analysis of Startups and Incumbents in Blade Computing & Communications Document RPT-115 (Oct 2003) Brief: Market Opportunities in Blade Computing – Scalable architecture from PCs to Supercomputers Document BRF-128 (Dec 2003)

Acknowledgements Royal Bank of Canada (RBC) Papers (Oct 2003) for incisive views on Fabric Computing Steven Beedle, ZNA Communications for continuous helpful support Dr. Lance Leventhal, Emulative Systems, for being the mentor

Key technologies in Computing & Communications Blades Some of the key technologies in progression now, to enhance the performance of Ethernet as the best price/performance core fabric include: 10 Gigabit Ethernet for increased bandwidth, TOE to reduce CPU utilization and increase throughput RDMA1 to reduce End-to-End latency and Buffer Copies (1Remote Direct Memory Access - technique introduced first in Infiniband) iSCSI to handle SAN traffic over Ethernet, iSER to maximize storage efficiency (combining the benefits of RDMA for latency reduction and iSCSI for SAN traffic).

State of Progression of Key Technologies IP SAN Today, the TCP/IP standards community is working to add block storage to the long list of data traffic supported by the Internet environment.

Since 2000 IETF, has formally hosted the standards effort of TCP/IP/Ethernet protocol stack.

IMEX Research estimates the iSCSI market is expected to grow to $1.089 billion in 2007 from $7.5 million in 2002. Primary inhibitors to date include the battles in the standards body, the technical challenges of acceptable performance, and the negative impact of the broader economic malaise on the R&D budgets of the major server and storage OEMs.

Meaningful iSCSI deployment may initiate in 2004 in low-performance environments. Traction as a challenger to Fibre Channel in high-performance situations appears much further out, probably no earlier than 2006.

Status of iSCSI iSCSI’s adoption is predicated on the assumed benefits of lower cost Ethernet network equipment and the ability to use a single network for the LAN and the SAN. Management benefits are expected to come from using LAN administrators as SAN administrators, thereby reducing the training and ongoing headcount expense required to manage SAN infrastructures. Today storage management software capabilities on Fibre Channel networks are much more sophisticated than those available on TCP/IP networks, direct-attach parallel SCSI installations in departmental and SME environments are the primary natural target for the first generation of iSCSI products particularly those applications requiring a networked configuration for greater manageability, but without high performance in terms of latency and bandwidth utilization. Today, the applications that meet this profile are usually candidates for Network Attached Storage (NAS) solutions.

Generally, we believe users will continue to find NAS attractive in many lower performance networked storage environments because of the additional file sharing and management benefits relative to a block storage approach. More users would turn to NAS solutions if these systems were freed from the performance issues created by current TCP/IP solutions. iSCSI as such is not a major threat to NAS technologies. Actually iSCSI must find a suitable subset of the NAS markets in which block level access to storage resources is a technical requirement. The most obvious beachhead applications for iSCSI fitting this profile reside in departmental and SME class Microsoft environments, where NAS is not well supported as a networked storage solution.

-012- Rise of Blade Computing.doc 12 ©2003-04 IMEX Research Microsoft Exchange, unlike Oracle’s database, is not engineered to interface with file systems, instead offering only”raw device” storage access mode.

Only NAS storage offerings that emulate a raw device to the database through an installed filter driver on the Microsoft host can be certified for Exchange environments. Vendors such as Network Appliance, Inc. are required to provide a proprietary iSCSI-like solution to gain Microsoft certification. Microsoft will continue to require raw device support for all instances of Microsoft Exchange..

The blade server market offers another potential catalyst for iSCSI in the data center. Many server vendors are working to deliver “diskless” blade server architectures, which rely on external, networked storage resources to boot up and execute. This offers compelling economics for users with SAN environments. Linux and most Unix operating systems will boot from NAS devices or block storage devices. Microsoft’s environment only boots from a block storage device due to the company’s software licensing model, which is not likely to change soon. To accommodate this constraint, diskless blade servers for Microsoft environments must provide SAN connectivity on each blade.

Several designs using Fibre Channel for this purpose are under development or in early development. iSCSI solutions may see significant deployment as well for lightly loaded Microsoft blades that use a single network connection for LAN and SAN connectivity needs, including boot up. Such configurations would provide strong “slot leverage” in highly integrated form factors where each peripheral adapter slot is highly coveted.

Beyond specific departmental and SME Microsoft environments, storage network extension is another possible beachhead for iSCSI to attack. Fibre Channel does not extend natively beyond 10 or 100 kilometers without the use of MAN or WAN transports. However, because Fibre Channel gained first-mover status for block oriented storage networks, alternative approaches to storage over IP networks, such as FC-IP or iFCP, may see greater near-term traction in these applications. Both of these approaches encapsulate Fibre Channel traffic in IP packets instead of native SCSI traffic. Unlike iSCSI, this preserves the addressing and protocol semantics required to link two Fibre Channel networks over distance.

Industry Initiatives Microsoft as the center of gravity for iSCSI deployment Microsoft now requires that customers deploy Microsoft certified hardware configurations to receive full support in Exchange environments. No server OEM will seriously commit to iSCSI in Microsoft environments until Microsoft is ready for certification testing. Microsoft has been determined to develop its own driver stack for iSCSI in order to exert greater control over the IP SAN ecosystem than was ever obtained in the Fibre Channel domain. Microsoft recently released the first version of the driver. We anticipate at least six months of certification testing, suggesting that Microsoft certified solutions might not be available until the end of 2003.

Cisco, Network Appliance, qLogic Cisco’s iSCSI router, which bridges from TCP/IP/Ethernet environments to Fibre Channel storage networks, has yet to gain meaningful traction despite an aggressive promotional campaign.

Network Appliance, Inc. recently announced iSCSI support, and many other major system vendors are expected to offer support by mid- 2004. Several startups also are addressing the opportunity, including EqualLogic, Intransa and LeftHand Networks, Inc.

TCP Buffer Buffer TCP

IP IP

NIC NIC

Network

Client or Server 2 Server 1

Microsoft Initiative - Chimney As such Microsoft is answering the call with an initiative codenamed “Chimney”, which it plans to roll out in two phases:

TCP Chimney Microsoft Chimney - TCP Offload SW Phase 1 - By mid-2004, a driver will offload most TCP/IP processing functions to a TCP/IP Application protocol engine. This approach may be well suited for most applications at 1- and 10-GBps Logical Switch speeds.

RDMA Chimney TCP Phase 2 - By 2005, a RDMA layer will be r Chimne placed on top of the TCP Chimney to enable full offload of TCP/IP processing. This design RDMA is expected to be necessary for applications Chimne that have very high frequency, small payload Data Transfe traffic patterns, such as storage, some file State Updates NDIS servers, clustered databases, and inter- processor communications. NDIS Miniport RDMA Enabled TCP/IP Adapter NIC HW

Ethernet LAN

-012- Rise of Blade Computing.doc 14 ©2003-04 IMEX Research Technology Challenges In early months, security and performance related issues were the major contributors to the malaise in finalizing the iSCSI standard.

On the security front, the concept of connecting storage assets to Internet technology harkens for many users a vision of hackers vying to convert business-critical data into digital mush. The iSCSI standards effort responded with a mandate for all iSCSI endpoints to incorporate security features, but this decision delayed the standards effort for months while adding cost to implementations.

As for performance, a number of issues were addressed, yet enhancements necessary to allow TCP/IP/Ethernet implementations to provide the sort of channel-oriented communications performance that storage I/O applications typically command.

Current TCP/IP/Ethernet implementations draw significantly upon the processor and memory resources of host systems and introduce significant latency in the communications link. Each message that an application sends using TCP/IP requires a copy of message buffers from the application’s memory space to the operating system kernel’s protected memory space (Exhibit 43). Then an operating system context switch from "user" mode to kernel mode is necessary to initiate the data transfer. The process works in reverse for each packet received. However, the receive process adds further processor and memory tasks prior to delivering a received packet to an application. Since TCP guarantees in-order delivery of a stream of bytes to a client application, it must manage the out-of-order receipt of transaction segments using additional buffers. Such out-of-order receipt is not unusual for both wide area and local area configurations.

TCP/IP/Ethernet proponents argue that the relentless increase in processor speeds obviates the need to address the aforementioned issues. Processor bandwidth is not the issue. Instead bandwidth between the processor and memory is the constraining factor,

RDMA DMA and RDMA techniques are commonly employed by channel interconnects. They allow data to be transferred between two endpoints at minimal latency by minimizing the use of host processor execution cycles and memory bandwidth.

Reception to an RDMA-enabled TCP/IP standard has been mixed. Server and storage system vendors were enthusiastic, recognizing the need for a generalized solution that could serve the purposes of the iSCSI effort for one-Gbps links and secure the broader opportunity for 10 Gbps Ethernet throughout the data center.

RDMA Consortium, the members of which include Adaptec, Inc., Broadcom, Cisco Systems, EMC, Hewlett-Packard, IBM, Intel, Microsoft, and Network Appliance, Inc.

IETF would use Stream Control Transmission Protocol (SCTP), in lieu of TCP. RDMA Consortium would use an enhanced or modified TCP implementation. Several vendors will develop TCP/IP/Ethernet adapters that support RDMA.

Commercialization is not likely before 2005, as a number of technical issues must first be resolved. The exact nature of TCP enhancements or modifications to support message-oriented delivery must be defined. Error detection algorithm currently used by TCP for insuring data integrity may require enhancements as well.

Several vendors are developing firmware and silicon that could enable RDMA over TCP/IP/Ethernet, including:

-012- Rise of Blade Computing.doc 15 ©2003-04 IMEX Research Adaptec Alacritech Banderacom Broadcom Cenata Networks Chelsio Communications Intel iReady Corp. QLogic S2io Technologies Siliquent Technologies SilverBack Technologies and Trebia Networks

Ethernet switches also contribute to the latencies associated with TCP/IP implementations by employing non-deterministic congestion algorithms that occasionally drop packets. TCP/IP/Ethernet switching solutions for storage applications must address several performance issues before competing head to head with Fibre Channel. Features such as out-of-order packet delivery, packet dropping due to congestion, and error recovery through random back-off algorithms simply do not mix well with SAN traffic.

Market Acceptance of IP SANs A new generation of enhanced equipment must emerge to service needs beyond the entry-level segments. More sophisticated SAN interconnects will begin to include TCP/IP connectivity in the 2003 timeframe, though typically in small quantities relative to the Fibre Channel ports on the same platforms. Software level support for RDMA-enabled TCP/IP, particularly implementation at the OS level is critical for broad commercialization.

Players in Blade Servers by Market Segment

Vendors by Technology Segment Virtual Machines5 Connectix/Microsoft,, Ensim, Swsoft Inc, VMWare Inc, Aurema Virtual OS Corosoft, Egenera, Synchon, VMWare Inc. Server Automation & Provisioning SW5 Altiris, Blade Logic, BMC Software,CA, Ejascent, HP, IBM, Jareva/Veritas, Microsoft, Moonlight Systems, Opsware, Platespin, Polyserve, Sphera, Terraspring/Sun, Think Dynamics/IBM, VM WareInc. Server Administration5 IBM, BMC, CA, HP, NetIQ, Micromuse, Veritas/Precise, Quest, Altiris, Amphus, Blade Logic, ConfigureSoft, Corosoft, Ejascent, Jareva/Veritas, Moonlight Systems, Novadigm, ON Technology, Opsware, Marimba, Mountain View Data, Platespin, Polyserve, Relicore, Swsoft, Sphera, Terraspring/Sun, Think Dynamics/IBM, VM Ware Inc HA Clustering5 MF -IBM , UNIX-Sun, HP, IBM , Windows – Microsoft, Legato, SteelEye, Veritas, Linux – Legato, SteelEye, Veritas, HA Technical Solutions, HP, Linux Networks, Mission Critical Linux, Polyserve, RedHAt, StoneSoft Clustered File Systems4 Polyserve, Sistina, Oracle, Veritas Interconnects5 Infinicon Systems, Scali, TopSpin, Voltaire KVM Switches5 Avocent, Belkin, Raritan, Rose Electronics 10 Gigabit Ethernet Intel, NEC, TOE

-012- Rise of Blade Computing.doc 16 ©2003-04 IMEX Research Alacritech, qlogic, Emulex, RDMA iSCSI - iSER Intel RDMA over TCP/IP Firmware and silicon Adaptec, Alacritech, Banderacom, Broadcom, Cenata Networks, Chelsio Communications, Intel, iReady Corp., Qlogic, S2io Technologies, Siliquent Technologies, SilverBack Technologies and Trebia Networks

Emergent Technologies driving Data Centers of the Future The evolution of Network Storage in Data Centers will evolve over the next 4 years from IP NAS and Fiber Channel SANs to IP SANs (iSCSI). In the interim multiple fabric types will co-exist in the Datacenter during this evolution.

Industry Standard High Volume Blade Servers will act as the core platforms to reduce system costs and complexity using a common fabric. Volume driven economics of Ethernet will prevail as it evolves to handle end-to-end traffic for LANs, IPC, SANs, Management Traffic as well as long-haul traffic in WANs/MANs for Remote Mirroring and Disaster Recovery.

Business Agility Link & Align IT with Business Processes Communicate, measure & deliver service Complete Data Center

Bus. Processes Virtualization Business Efficiency Provision Resources Link IT w Business Align IT based on dynamic needs with Business Processes Optimize utilization of Communicate, measure & business processes

Services Manage E2E business deliver services Business Stability Flow & Processes Plan, Provision, Monitor, Inventory, Model, Maintain, Control: Networks, Servers, Storage,

Resources Applications, Clients

Discrete Integrated Virtualized Partitioned Clustered

Server provisioning software for Blades Server consolidation initiatives drive adoption of server provisioning software for Intel based Windows and Linux platforms. Server consolidation is the top priority in Windows server environments, while lack of support for multiple workloads within the OS is the top functional issue. virtual machine software will see strong adoption on Intel platforms to support mixed workload scenarios. Meanwhile, several vendors are developing software architectures that provide centralized, automated server provisioning across all servers in the data center.

KVM over IP solutions add software tools in support of remote server provisioning solutions. KVM switching connects directly to the console signals on each server and to a centralized management station. The latest technology of KVM over IP offerings connect managed servers to centralized consoles by transmitting KVM signals over TCP/IP networks with minimal latency.

-012- Rise of Blade Computing.doc 18 ©2003-04 IMEX Research Investment opportunities in Servers So, the most attractive investment opportunities involve accelerating the penetration of industry standard platforms throughout the data center and replicating innovations from mainframe and UNIX platforms onto Wintel and Lintel platforms to improve server access, availability, and administration.

-012- Rise of Blade Computing.doc 19 ©2003-04 IMEX Research Definitions Virtual OS – enables aggregation of mutiple Windows or Linux OS instances or servers to form a single powerful Virtual OS like in mainframes or UNIX Servers Virtual Machine – To dramatically improve the physical and logical provisioning capabilities of Industry Standard Servers by emulating extrinsic functions on Wintel & Lintel platforms Server Automation & Provision SW – Tools to centralize and automate provisioning of distributed software spanning multiple servers such as configuration, monitoring, visualization, reporting and provisioning Server Administration SW - Tools to integrate and administer such as configuration, monitoring, visualization, reporting and provisioning

HA Clustering Clustered File Systems4 Interconnects5 – High bandwidth, low latency SW to interconnect servers RDMA over TCP/IP/Ethernet –

References Report: Blade Servers Industry Report 2003 Document RPT-107 (Aug 2003) Brief: UNIX versus Windows based Server Architectures. Document BRF-112 (Feb 2003) Report: Competitive Analysis of Startups and Incumbents in Blade Computing & Communications Document RPT-115 (Oct 2003) Brief: Market Opportunities in Blade Computing – Scalable architecture from PCs to Supercomputers Document BRF-128 (Dec 2003)