December 2007

Fabric Convergence

Attributes of the Ideal Unified Wire

• No changes to software
• No new switch infrastructure
  • Ability to work with legacy Ethernet switches
• Must be True Ethernet
  • No new requirements for IT training or personnel
  • Must be extendible
• Must be TCP/IP
  • Must be routable – no required multi-protocol gateways
  • Must be able to inter-operate with peer laptops with Win98 installations
• No compromise in performance
  • Performance must be as good as a dedicated fabric
• Must concurrently support all traffic types
  • network, cluster, and storage communications
• Ability to segregate the various traffic types
  • Ability to do traffic management
  • Ability to impose access controls
• Must not impose new demands on the host CPU
  • No additional power or cost tax
• Must be tested in enterprise environments before deployment

TCP/IP-based approach to convergence is ideal

Converged Fabric Comparisons

• Why not DCE?
  • Standard not ratified
  • Serious questions about protocol viability – PAUSE storm, hotspots
  • Requires infrastructure replacement – 85% of the cost of a typical SAN today is in the switch
  • QoS has been in play for 10+ years and yet not adopted fully yet, obviated each time by the next speed upgrade of Ethernet.
    40Gb Ethernet will be available in 2008, diminishing the need for DCE again. 100GbE is around the corner.
• Why are some vendors grey’ed out?
  • No Traffic Manager or Access Control Engine
  • No enterprise-deployed TOE and hence no non-demo iSCSI or iWARP. TOE has been demo/PR-only and unstable, poor
    performance in real-life networks. Therefore it is effectively a NIC only.
  • RDMA support is not in the kernel
• What about IOAT?
  • Limited performance benefits
  • No solution for clustering
  • Limited to and requires new Intel chipsets
  • No Traffic Manager or Access Control Engine

Unified Wire Vitals

• Tremendous ROI for users today
  • Less hardware to purchase
  • Less personnel to manage infrastructure
  • Savings in power, real estate
    • TOE at 10Gb saves $100/year in utilities alone
• Unified Wire on the motherboard
  • The decision for clustering and storage users will be whether to purchase additional HBA/HCA, switches, and gateways for their storage and clustering applications or to just run everything on the chip that is on the motherboard -> incremental cost = 0
• Ethernet & TCP/IP based approach
  • Works with legacy clients and switches
  • Well understood and tested for the past 20 years
  • Will future-proof all new data center installations

10GbE TCP/IP Network Convergence is the next transition

Chelsio Snapshot

• Established vendor
  • 32 OEM customers – several Tier-1 server & storage
  • 131 engagements
  • 50 current evaluations
• ISO certified for 3 years
  • Repeatable processes
• Revenues tripling per year since 2004
  • Profitable in 2008
• In the channel with Bell Micro
  • Tier-1 storage distributor
• Best in class product
  • Strong roadmap, 3rd generation product
• In the top 5 companies to address a $2B market TAM
  in 2010

Chelsio is a long term player in this market

Market Opportunity

Chelsio enables fabric convergence across all data center infrastructure segments

Current Products

Third generation solution Holder of all Internet Land Speed Records Clustering, storage, networking, appliance, Video, IPTV solutions

Single Chip Can Do All

• A True Unified Wire
  • Runs NIC, TOE, iSCSI, RDMA applications concurrently and at 10Gb full-duplex
    • No need for gateways
    • Use existing Ethernet switches
    • No additional IT support
  • Integrated traffic management for seamless sharing of infrastructure
  • High-speed protocol implementation with exception handling for robust performance in real-life networks
• Virtualization support
  • Highest performance VMware solution
• Rich software roadmap

Delivers the ROI of Fabric ConvergenceAddresses InfiniBand& FibreChannel & Networking markets

Highest Performance

• Highest performance VMware implementation
  • 10Gb on Receive and Transmit
• Highest performance iSCSI implementation
  • 900K IOPs, 2200 MB
  • More than twice the performance of 4G FC for Oracle & Exchange
• Highest performance NIC adapter
  • Holds all the IPv4 and IPv6 records
  • 7M packets/sec – ideally suited for Web 2.0 applications
  • Highest transaction rate NIC on the market
• Highest performance TOE adapter
  • 10GbE at < 5% CPU with standard frames for ONE connection
  • Ability to accelerate the legacy sockets applications
  • Fast error recovery delivers consistent and scalable performance
• High performance clustering solution
  • 5-7us MPI latency

Highest performance solution across all dimensions

Robust, Tested, Seasoned

• Enterprise Grade product
  • TOE has been shipping for 2+ years – as a TOE
  • 700 node TOE cluster in use for 1 year at Osaka University
    • In the top 50 super-computers in the world
    • Largest 10GbE HPC installation to date
  • 128 node RDMA OFED cluster at Sandia
    • 5x bandwidth of DDR IB for 14+ nodes at 1/4th the CPU%
    • Largest iWARP installation to date
• Supplier to Tier-1 Storage & Server OEMs
• Independently tested, verified, and deployed

The only stable 10GbE solution shipping as a TOE
The only stable 10GbE solution shipping as RDMA

Value Added Features

• Integrated Traffic Manager
  • Ability to shape and pace transmission
  • Ideally suited for IPTV, VOD & Storage Virtualization
  • A required component when low latency clustering traffic has to coexist with bulk storage traffic
• Integrated Classification Engine
  • 16K simultaneous filter rules using bitwise masks and wild cards
  • Packet tracing functions
  • A required component for virtualization, WAN, & sniffer applications
• Broad product offering
  • PCI-Express or PCI-X 2.0
  • 1GbE or 10GbE
  • Fiber or CX4 or 10GBase-T
  • 1-port, 2-port, Server Blade Mezzanine form factors
  • Same software port will drive all variants

Much much more than a NIC

T3 Silicon Overview

Competition

• Chelsio competes primarily with companies that develop Ethernet silicon
  • Indirect competition from other fabrics such as InfiniBand and Fibre Channel
• Chelsio delivers a best-in-class, full-featured and seasoned solution

• Why not Multi-RISC?
  • Higher power consumption
  • Firmware stability problems compounded with need to rewrite when fixing bugs or adding features
  • Per connection performance is limited by RISC core – 10Gb performance is not for one connection
  • Performance doesn’t scale – 10Gb performance is not for large number of connections
  • Required store and forward results in a high latency solution

Chelsio is well positioned in the products category
The architecture is the enduring differentiator

Glossary

• OFED – Open Fabrics Enterprise Distribution
  • An API that is shared by InfiniBand and iWARP and promoted by www.openfabrics.org. This used to be called OpenIB. This API has been accepted into the Linux kernel, and is in-boxed by RedHat & Novell. Chelsio’s T3 is the first and only 10GbE Ethernet card to have been accepted. This enables running of the unmodified IB applications written to this API (about 99% of the applications), over Ethernet. Support for OFED is part of the strategy of every major server OEM.
• TOE – TCP Offload Engine
  • TCP is a 25 year old transport protocol for the internet. To implement or offload TCP in hardware has the benefits of offloading the host CPU, allowing for predictable processing latencies, saving CPU power (i.e. about $100/year in utilities), and in general converting any legacy L2 switch fabric into a lossless medium (similar to the currently proposed CEE protocol).
• iSCSI
  • This is a protocol that allows running applications written to an SCSI API (i.e. FibreChannel applications) to be run over the IP protocol. iSCSI protocol requires protocol data units that can span many TCP segments and also requires a CRC-32 computation for header and data. For this reason, hardware assistance (in the form of a TOE) is required to implement this protocol with high performance.
• RDMA – Remote Direct Memory Access
  • This is a protocol via which a server can place the data directly into a peers memory. The benefits are lower latency and CPU utilization and memory bandwidth savings. Open Fabrics is one form of RDMA. RNIC-PI API proposed by some OEMs is another.
• WSD – WinSock Direct
  • Microsoft’s API for direct communication between user space applications. This API bypasses the OS and allows for a lower latency interface. It is primarily used for clustering and HPC applications.
• Chimney
  • The marketing name of Microsoft’s version of TOE. Chimney uses a special form of TOE implementation called Partial Offload. This refers to a TCP connection that starts life in the host and then once the connection is established, it migrates down to the card. It is only useful for long lived connections but allows for simpler hardware, at the expense of considerably more complex software. Currently Chelsio’s T3 is the only 10Gb hardware that has successfully received a WHQL Logo from Microsoft for Chimney: A requirement for the server OEMs to productize this feature. Support for Chimney is part of the strategy of every major server OEM.
• Uwire – Unified Wire
• This refers to an Ethernet controller that can run OFED, iSCSI, TOE, NIC traffic simultaneously. When put on the motherboard, this solution can run FibreChannel and InfiniBand applications unmodified. Thus the decision for the customers becomes: why should they have to purchase another switch fabric, another HCA, train a new IT person for IB or FC, when they can run those legacy applications on their unified wire Ethernet. Today, Chelsio is the only vendor that has such a product.
• IKE, PKE – Internet Key Exchange, Public Key Exchange
• The handshake protocol required in implementing the IPSec protocol. IPSec protocol (part of Microsoft’s LongHorn), requires these handshake protocols and an in-line cypher engine to encrypt/decrypt the data. The complexity of the IKE, PKE code is significant and while Windows provides that, for other OS’s, it must be developed.
• Traffic Manager
• A hardware module that can shape and pace the TCP traffic on a per connection basis. This enables mixing of the low latency clustering and the bulk storage traffic onto the same converged legacy Ethernet switch fabric. It also for examples allows providing a specific 5.5Mb of MPEG traffic to a particular customer who paid for that much service, or to delivery exactly as much bandwidth to a storage disk as its spindle can handle, etc.
• Virtualization
• The ability to appear as many different entities at the same time. Server virtualization refers to the ability of the server to appear as many servers for example and run multiple guest OS’es simulatenously. NIC virtualization refers to appearing as many different NICs at the same time, etc.
• HA – High Availibility
• Refers to the ability to always have access to the data: a required function, especially when many different OS’es are virtualized. For example, when a branch office is running VMWare virtualization software, it needs at least two physical servers to allow high availibility from a hardware redundancy point of view.
• SAN – Storage Area Network
• An implementation of storage in a network fashion and external to the server. The vmware virtualization applications for example, require this mode of storage for high availibility and remote manageability reasons. For small branch offices though, this is an expensive value proposition. It is possible to take advantage of the high disk densities today and implement a SAN on the storage native to a server using StorageWare: i.e. StorageWare can make a direct attached disk look like a shared SAN, which is the API that the vmware software requires.