1 Networks: The State of the Onion Tom Lyon For Brocade Communications 9/10/2003

2 What is a Network?  Many possible definitions  As simple as a single wire  As complex as the Internet  For today’s talk, use is Ethernet & IP centric: A potentially large set of communicating computers

3 Big Trends  Life after the Bubble  Wireless Exuberance  Security  Voice over IP  The Price/Performance Trap

4 Network Markets  Telco/Carrier/Service Provider  ISPs, ASPs, etc.  Enterprise – G2000 vs SME  Home/SOHO

5 Post-Bubble  Depression, Disaster, Fraud, Dereliction & Bankruptcy  Telcos are hurting  ISPs are dead, but the Internet is booming  Enterprises are looking for fewer external dependencies – building private networks  Home networks are booming!

6 Private Networks  Wireless Pt-Pt: 1M – 1G  Dark Fiber: Anything  TDM: T1, T3, Sonet: 1.5M – 2.5G  Packet: Frame Relay, Private IP, MPLS, …  Metro Ethernet  Internet VPNs – IPSec

7 Wireless  3G Hangover – many telcos paid billions for 3G licenses, now 3G is being written off  WISPs – ISPs for Wireless Hotspots  Enterprise Wireless vs Security  Home wireless  No proof of any new $$ from wireless LAN

8 Security

9 Security Chaos  Firewalls don’t catch internal hackers  VPNs don’t prevent infected clients  Intrusion Detection systems – false alarms  Virus updates don’t get applied  Each new protocol (XML) breeds new security problems

10 Network vs Desktop Security  Central mgmt in network device can be more secure, but performance is a problem, e.g., anti-virus  When laptops move among networks, how can network security apply?

11 Security Policy  Customers demand flexible policy control in security systems  When vendors deliver, they discover no actual policies exist  Good/bad is too inconvenient; shades of gray rule in social systems  Vendor vs Admin vs User control

12 VoIP  Up to 10% of voice calls are now on IP  Mostly in trans-oceanic  Regulators are noticing  Enterprises can finally ‘converge’ voice and data  IP phones and 802.11 driving power over Ethernet

13 State of the Internet  Traffic growth 74% in 2003, up from 38%  Wholesale price: $100/Mb/mo, down 80% since 2000  Global backbone rev $1.7Bn in 2002; implies about 1.5Tbps of bandwidth  Source: TeleGeography

14 1: Physical Layer  Gigabit Ethernet on CAT5 – 1000BaseT  Wireless – 802.11a/b/g, Bluetooth, …  10 Gigabit – 10gBase-CX4  Optical – WDM, PON  Power over Ethernet: 802.3af  HomePNA – Ethernet on Phone Lines  HomePlug – Ethernet over Power lines

15 2: MAC layer  Ethernet über Alles  ATM no longer cool  Lots of different wireless MACs  Layer 2 switching:  Complete 8x1000BaseT switch - $125  Single chip 12x10G switch - Fujitsu

16 Ethernet History  1973: Metcalfe et al – 2.94Mb  1980: DEC, Intel, Xerox Blue Book 10Mb  1983: IEEE 802.3  1990: 10Base-T  1995: 100Mbps  1998: Gigabit Ethernet  2002: 10G Ethernet

17 Switch History  70s: “Ethernet – A Distributed Switch”  Early 80s: LANs take off  Mid 80s: Bridging between LANs  Late 80s: 10Base-T & hubs  1990: Kalpana EtherSwitch  Mid 90s: ASICs + performance explosion  Late 90s: “Layer 3” switches

18 3: Network Layer  IP, of course  IPv6 – waiting in the wings  MPLS – carriers only  IPSec – mature solution to the wrong problem  Big router: Procket 48x10Gbps  IPSec: Cavium 10Gbps chip

19 IP History  1972: Kahn proposes ‘Internetting’  1977: 16 network numbers  1978: Cerf proposes ‘Catenet’ model  1981: IPv4: RFC 791, 43 networks assigned  1983: ARPANET transition to IP & TCP  1995: Windows 95 released with TCP/IP  2002: 200M hosts, 700M users

20 Router History  70s: unique proxy code for each network pair  Early 80s: IP forwarding in UNIX  Late 80s: Cisco & router “appliances”  Early 90s: real router hardware  Late 90s: routing ASICs, performance explosion

21 Switch vs Router  What is the difference between L3 switch and a router?  Switches started with performance and added functionality  Routers started with functionality and added performance  Switches skimp on buffers; use flow control  Routers have big buffers; can’t flow control TCP  Switches in LAN, routers in WAN

22 4: Transport  TCP marches on  Terrible match with HTTP  No good in high bandwidth*delay environments  Sensitive to IP address loss/change – single homed  SCTP – new & cool  Multi-homing, other features  Started with SS7 over IP  TOE – TCP Offload Engines  Avoid OS/TCP overhead  Driven by iSCSI

23 RDMA  Network requires send and receive  Processor & OS desire write and read  RDMA layers write/read semantics onto TCP/SCTP  RDMA requires TOE to offload OS  What will win-  Change network to accommodate OS, or  Change OS to accommodate network?

24 SSL/TLS  Transport level security  Broad browser/OS support  TCP,SSL,TLS traverse NAT boundaries  Easier to attach policy to users & apps than at IP level  SSL VPNs – easier adoption than IPSec  SSL hw: key computation vs transport

25 5: Session  SCTP  HTTP – 1.0 vs 1.1  1.1 allows multiple pages per TCP connection  Good for throughput, sometimes bad for latency  HTTP Compression  Very compute intensive at server  Great for bandwidth reduction

26 6: Presentation  XML über Alles  XML is just a syntax standard  Evolution of HTML from man-machine to machine-machine; still uses HTTP/SSL/TCP  SOAP provides “datagrams” for XML  XML spawns new security problems

27

28 7: Application  Transparent use of network (NFS) is old- world  Explicit use of network: Client/Server, Clustered Apps, Peer-to-Peer, ….

29 9 Network Fallacies  The network is reliable  Latency is zero  Bandwidth is infinite  The network is secure  Topology is fixed  There is one administrator  Usage is free  The network is homogeneous  Scaling is easy

30 Other Worlds  Infiniband – is it a network? Is it a bus? Only the non-existent software knows for sure.  Supercomputing interconnects (Quadrics, Myrinet) – tightly couple high level APIs with low level hardware  Bluetooth – IRDA on steroids

31 The Price/Performance Trap  Moore’s Law:  Double transistor density every 18 months  When solutions move to single chip, huge performance increase is possible  If your product is measured primarily by price/performance you’ve entered a black hole from which there’s no return  Brand, positioning, etc can help but not cure the problem

32