Moving x 86 Virtualization Forward with AMD s Quad-Core

Moving x 86 Virtualization Forward with AMD’s Quad-Core Technology Margaret Lewis Director, Commercial Solutions

AMD and Novell – Optimized Platform for 64 -bit Computing with Linux Delivering Industry-leading Innovation Novell SUSE Linux Enterprise Server was one of the first operating systems featuring 64 -bit support at AMD Opteron™ processor launch in April 2003 Novell and AMD have a long-standing collaboration focused on delivering AMD 64 optimizations into Linux Enabling the Open Enterprise The innovation of power-efficient AMD Opteron™ processor-based servers running industry leading Linux technology from Novell Addressing IT challenges Delivers performance that scales as demand increase while reducing power, cooling and software costs Virtualization capabilities available today that provide the underlying functionality to consolidate resources and maximize IT investments 2 Month ##, 200# Presentation Title

Transforming Data Center Infrastructure Dedicated Infrastructure Shared Infrastructure Virtual Resource Pool Internet • Significant unused capacity • Higher rates of utilization • Inflexible • Reconfigure dynamically • Many legacy technologies • Unified resources “managed as one” 3 Month ##, 200# Presentation Title

What Is A Virtual Machine? Application Guest OS Virtual Machine Hypervisor Real Machine 4 Month ##, 200# Technology that partitions a computer into several independent machines that can support different OS’s and applications concurrently Hypervisor is the underlying software which runs directly on the hardware and manages multiple operating systems Virtual machine is selfcontained operating environment that runs on top of the hypervisor and behaves as if it is a separate computer Presentation Title

Behind Every Virtual Machine Stands (a Fraction of) A Physical Machine One Physical Machine Can Typically Host 10 to 30 VMs Multi-Core CPUs Will Further Improve This Ratio Reduces “Server Sprawl” Better Use of Data Center Real Estate Improves Electrical Power Utilization 10 -20 Watts Per Idle VM (vs. 100 -200 W Per Idle PM) Simplifies Deployment of New Applications, Servers Source: Picasso, 1968 (Homme a la Pipe) 5 Month ##, 200# Presentation Title Facilitates Rapid Recovery From Hardware Failures

Virtualization In the x 86 World Application g in R Ri 3 ng 2 ng Ri Ri 1 ng 0 Operating System Device Drivers Only privileged code can define memory-mapping parameters or set up interrupt handlers Application 6 Month ##, 200# Presentation Title • In the x 86 environment, the OS is “selfish” and expects direct control of hardware resources • The OS runs in “Ring 0” and exercises control over software running at higher rings • Traditional x 86 processors can't simultaneously run both host and guest OS’s at Ring 0 • Virtualization vendors have been forced to overcome both hardware and software limitations to allocate and manage processor, memory, and I/O resources

Software-Enabled Virtualization Application g n Ri 3 ng g n Ri 2 Modified Guest OS n Ri g Ri 1 ng 0 Hypervisor Host OS Ri 3 ng Ri Unmodified Guest OS 2 ng Ri 1 g in 0 Hypervisor R Para-virtualization Full Virtualization • Hypervisor handles Host/Guest OS interactions • Hypervisor is also Host OS • Guest OS is recompiled to run at Ring 1 • Intercepts and emulates instructions and virtualizes devices • Host OS handles devices • Guest OS runs unmodified 7 Month ##, 200# Presentation Title

Hardware-Enabled Virtualization with AMD 64 Processors Application g in 3 R ng Ri • Hypervisor controls access to devices 2 g in R 1 g in • Creates a Guest Mode that allows Guest OS to run unmodified with Hypervisor Guest Mode 0 R Hypervisor • Architecture designed to speed the switches between hypervisor and guest modes • Instructions ease creation and management of virtual machines 8 Month ##, 200# Presentation Title

AMD Is Driving x 86 -based Virtualization AMD 64 Technology Road Map Direct Connect Architecture Increase performance with multi-core capabilities and fast memory access Virtualization Nested I/O Extensions Page Tables Virtualization (Today) (2007*) (2008*) Simplify virtualization software and helps reduce CPU overhead Reduce overhead of switching between virtual machines Virtualize devices to improve performance and security AMD is developing a robust virtualization ecosystem that provides users with innovation and choice *Planned features for future AMD Opteron™ processor-based systems 9 Month ##, 200# Presentation Title

AMD Opteron™ Processor The Platform for Virtualization 64 -bit and Dual-Core Performanceper-watt Enhances performance while offering the flexibility to support both 32 - and 64 -bit applications Assists data centers in controlling power consumption and heat output Virtualization Performance High performance in smaller form factor Grid Computing Perfect balance of performance, power, and cost 10 Month ##, 200# Presentation Title Enables separate, secure operating environments

Introducing “Barcelona”… Native quad-core upgrade for 2007 Platform Compatibility Native Quad-Core Processor Socket and thermal compatible with “Socket F”. To increase performanceper-watt efficiencies using the same Thermal Design Power. Direct Connect Architecture Advanced Process Technology • Integrated memory controller designed for reduced memory latency and increased performance • Memory directly connected 65 nm Silicon-on Insulator Process Fast transistors with low power leakage to reduce power and heat. • Provides fast CPU-to-CPU communication • CPUs directly connected • Glueless SMP up to 8 sockets 11 Month ##, 200# Presentation Title

Direct Connect Architecture CORE MCP CORE MCP USB CORE MCP PCI CORE MCP 8 GB/S PCI-E Bridge PCIe™ Bridge Memory Controller Hub I/O Hub CORE 8 GB/S XMB XMB 8 GB/S PCIe™ Bridge XMB PCIe™ Bridge 8 GB/S USB I/O Hub PCI Legacy x 86 Architecture • • 20 -year old front-side bus architecture CPUs, Memory, I/O all share a bus Major bottleneck to performance Faster CPUs or more cores ≠ performance 12 Month ##, 200# Presentation Title AMD 64 Technology with Direct Connect Architecture • Industry-standard AMD 64 technology • Eliminate bottlenecks of traditional front-side bus architecture • Hyper. Transport™ technology interconnect for high bandwidth and low latency

AMD Virtualization™ Today Enabling More Efficient Workload Consolidation • Instructions added to help reduce complexity of virtualization software Application Guest OS Virtual Machine • Reducing overhead by selectively intercepting information destined for guest Virtual Machine • Enables Guest OS’s to run unmodified for easier implementation and support Hypervisor Host OS Virtualization Instructions • Increased isolation to improve security of virtual machines • Offers architectural enhancements to improve efficiency of switching between hypervisor and the guest OS’s 13 Month ##, 200# Presentation Title

Virtualization is Memory Intensive VM VM VM VM VM VM CPU CPU Memory Controller Memory Controller Hub Shared memory can create bottlenecks • Shared front-side bus can decrease application performance within a virtual machine • Untagged TLB means less efficient switching between virtual machines • Software-based memory management and security (via external Memory Controller Hub) can reduce overall virtualization performance and efficiency 14 Month ##, 200# Presentation Title Dedicated memory for scalability • Direct Connect Architecture helps improve application performance within a virtual machine • Tagged TLB means more efficient switching between virtual machines • Hardware-based memory management and security (Integrated memory controller with DEV) can improve overall virtualization performance and efficiency

Virtualization Performance Today – 4 P Servers Advantages of AMD Opteron™ processor -based system for x 86 Virtualization 96% CPU Util. At 20 VMs: Over 2 x better web serving Over 50% better file & print serving 82% CPU Util. 15% less power draw 14% more CPU headroom At Same CPU Utilization: 96% CPU Util. 6 additional VMs hosted 2. 6 x better web serving 2. 1 x better file & print serving 883 W 740 W 902 W 763 W 796 W 12% less power draw Full 4 P study available at: http: //www. veritest. com/NR/rdonlyres/F 1 F 21288 -8 D 6 D-4769 -94 F 3252 FCB 02 D 51 F/272/AMD_Virtualization_Project_Final_Report. pdf 15 Month ##, 200# Presentation Title

Virtualization Performance – 2 P Servers Advantages of AMD Opteron™ processorbased system for x 86 Virtualization 126% Geomean At 12 VMs: 10% Advantage in Light Workload 26% Advantage in Heavy Workload 110% Geomean Approximately a 26% performance/watt advantage At similar CPU utilization rates: 7 -12% more SQL Requests per second 13 -51% more Web requests per second 2 -19% better file & print serving Full study available at: http: //www. veritest. com/ 16 Month ##, 200# Presentation Title

Nested Paging Enabling More Efficient Switching App Page Table Guest OS Page Table Application App Page Table Application Guest OS Page Table Shadow Page Tables App Page Table • Provides the guest OS with the illusion that it is managing memory • Page tables are actually kept up by the hypervisor in software • Requires more software intervention from the hypervisor 17 Guest OS Virtual Machine Hypervisor Nested Page Tables Virtual Machine Hypervisor App Page Table Guest OS Virtual Machine Application Month ##, 200# Presentation Title Host Page Table Guest OS Page Table • Each guest physically has their own world to manage • Requires less intervention • Memory look ups done in hardware which can be faster than software management

AMD’s “Barcelona” Processor Improves Virtualization Performance Nested Paging (NP) Guest and Host page tables both exist in memory – The processor walks both guest and host page tables Nested walk can have up to 24 memory accesses! – Hardware caching accelerates the walk “Wire to wire” translations are cached in TLBs Eliminates hypervisor cycles spent managing shadow pages – As much as 75% of hypervisor time Barcelona also reduces world-switch time by 25% World-switch time : round-trip to the hypervisor and back 18 Month ##, 200# Presentation Title Application Guest OS Virtual Machine Nested Page Tables Hypervisor Host OS AMD-V Host Page Table Guest OS Page Table

Cygwin compile with AMD Nested Paging VMware products will take advantage of hardware-based Guest Page Translation assist (such as AMD Nested Paging) in 2007 and beyond as hardware becomes available. Cygwin Compile Among Best Case Improvement for Nested Paging, which mainly helps memory-management intensive workloads; not representative of all workloads. Nested Paging reduces compile time by 43% Platform: Experimental AMD Processor with Nested Paging running experimental build of VMware Workstation. 19 Month ##, 200# Presentation Title

AMD: Consolidating Hardware In Austin, approximately 135 servers consolidated to 7 active ESX 3. 0 servers plus 2 swing servers In Sunnyvale, approximately 45 servers consolidated to 2 active ESX 3. 0 servers plus 1 swing server Overall consolidation ratio between the two sites is 15: 1 including swing servers Month ##, 200# Presentation Title Anticipate 79% reduction in power consumption (Austin Data 20

CDW: Wowing the Software Developer • One of the largest direct providers of technology solutions for business, government and education Virtualization Environment • • • Develops most applications in-house Virtualization Software: VMware ESX Servers: 2 P servers and 4 P HP Pro. Liant DL 585 s running AMD Opteron™ processors Guest OS’s: Windows® and Linux® Applications: Windows domain controllers, IIS, FTP Endeca, ISA Server 2004, File & Print, SQL Server, Exchange, Blackberry Enterprise Server, Biz. Talk Server, Cognos, People. Soft ü Can now easily provision development and test environments without having to build physical servers ü Reduced server deployment time from a day to a half hour ü $25, 000 a month savings from deploying virtual machines ü CPU utilization increased from 5 – 15% to an average of 50% http: //www. vmware. com/pdf/cdw. pdf 21 Month ##, 200# Presentation Title

The Data Center of the Future In the 1970’s and 1980’s, Large Systems Prevailed The Larger, the Better (Grosch’s Law) Over the Last Decade, Scale-Out Deployments Became Feasible The Smaller, the Better (Moore’s Law) Today, the Pendulum Will Start to Swing Back Toward Larger Systems Reduce Server Sprawl and Power Requirements Simplify System Management Improve Resource Utilization Virtualization Plays a Key Role in Enabling this Shift

Closing Considerations Virtualization Facilitates Resource Sharing But You’ve Got to Have Resources to Share Some workloads are not the best candidates for virtualization May Be of Little Value in Compute. Intensive HPTC environments Virtual Machines Require Real Software Licenses Paid for with Real Money Source: Picasso, 1947 Ulysses and the Sirens) 23 Month ##, 200# Presentation Title

Trademark Attribution AMD, the AMD Arrow logo and combinations thereof are trademarks of Advanced Micro Devices, Inc. in the United States and/or other jurisdictions. Other names used in this presentation are for identification purposes only and may be trademarks of their respective owners. © 2006 Advanced Micro Devices, Inc. All rights reserved. 24 Month ##, 200# Presentation Title