Computer Networks Lecture #11 PDF

Computer Networks Lecture #11 In the last lecture LANs Addressing, Address Resolution Protocol (ARP) Ethernet Switches VLANs Today Link virtualization - MPLS Data center networking Link virtualization : MPLS  MultiProtocol Label Switching (MPLS) Works on the networ...

Computer Networks Lecture #11 In the last lecture LANs Addressing, Address Resolution Protocol (ARP) Ethernet Switches VLANs Today Link virtualization - MPLS Data center networking Link virtualization : MPLS  MultiProtocol Label Switching (MPLS) Works on the network of MPLS-capable routers Speed up IP forwarding Faster lookup using xed length identi er Borrowing ideas from Virtual Circuit (VC) approach But IP datagram still keeps IP address Flexible routing Tra c engineering is allowed Fast recovery from the link failure BoS (S) - can have more than one MPLS label - bottom of label stack: 1, otherwise 0 ffi fi fi  MPLS capable routers a.k.a label-switched router Forward packets to outgoing interface based only on label value Do not inspect IP address MPLS forwarding table is distinct from IP forwarding tables MPLS forwarding decision can di er from those of IP Use destination and source addresses to route ows to the same destination di erently (tra c engineering) Pre-computed backup paths can be directly used to re-route ows if links fail ffi ff fl fl ff MPLS vs. IP paths IP routing Path to destination is determined by destination address alone Same destination → same path MPLS vs. IP paths MPLS routing Path to destination can be determined based on source and destination address Flavor of generalized forwarding (MPLS 10 years earlier) Fast reroute: pre-compute backup routes in the case of link failure  MPLS signaling Modify OSPF, IS-IS link state ooding protocols to carry info used by MPLS routing e.g., link bandwidth, amount of “reserved” link bandwidth Entry MPLS router (R4 in the gure below) uses Resource Reservation Protocol- Tra c Engineering (RSVP-TE) signaling protocol to set up MPLS forwarding at downstream routers ffi fi fl MPLS forwarding tables Data center networking (DCN) Cloud computing Elastic resources Expand and contract resources Pay-per-use, infrastructure on demand Multi-tenancy Multiple independent users → resource isolation needed Amortize the cost of the shared infrastructure Flexible service management Resilience: isolate failure of servers and storages Workload migration: move task to other location Large-scale data centers Google data centers world wide Google’s Iowa data center How does a Data center look inside? Data center networks How to connect these servers in a data center? With a giant switch that connect all servers? What problem can you think of with such a design? Data center networks How to connect these servers in a data center? With a giant switch that connect all servers? What problem can you think of with such a design? Scalability issue (limited port density) Broadcast storm Isolation … Data center networks A dedicated network for the data center Tradeo between connectivity and complexity ff Tree-based DCN Example DCN topology (META) META F16 data center network topology Application-layer routing Load balancer Application-layer routing Receives external client requests Directs workload within the data center Return results to the external client (hiding data center internals from clients) Multipath Rich interconnection among switches, racks Increased throughput between racks Increased reliability via redundancy e.g., two disjoint paths highlighted between rack 1 and rack 11 Network performance metric Bisection width The minimum number of links cut to divide the network into two halves Bisection bandwidth The minimum bandwidth of the links that divide the network into two halves Full bisection bandwidth One half of nodes can communicate simultaneously with the other half of nodes  Oversubscription De nition Ratio of worst-case required aggregate bandwidth among end-hosts to total bisection bandwidth of the network topology Examples Typically, data center 1:1 → all hosts can use full uplink capacity subscription ratio is 2.5:1 ~ 8:1 5:1 → only 20% of host (uplink) bandwidth may be available fi Oversubscription 192 (16 × 12) nodes fat-tree built with Two 96-way switches Twelve 24-way switches What is the oversubscription ratio of the topology? Each 24-way switches has 8 uplink connections to the root 16 downlink connections to the hosts Oversubscription 192 (16 × 12) nodes fat-tree built with Two 96-way switches Twelve 24-way switches What is the oversubscription ratio of the topology? Each 24-way switches has 8 uplink connections to the root 16 downlink connections to the hosts Oversubscription 192 (16 × 12) nodes fat-tree built with Two 96-way switches Twelve 24-way switches oversubscription ratio = (16 × 12) / 48 Each 24-way switches has 8 uplink connections to the root 16 downlink connections to the hosts Factors behind DCN designs Commoditization in the data center Inexpensive, commodity servers and storage devices But the network is still highly specialized Data center network is not a “smaller Internet”. One admin domain, not adversarial, limited policy routing, etc. Bandwidth is often the bottleneck Data-intensive workloads (big data, graph processing, machine learning) Protocol innovation in DCN Link layer RoCE: remote DMA (RDMA) over Converged Ethernet Transport layer Explicit Congestion Noti cation (ECN) is used in transport-layer congestion control (DCTCP, DCQCN) Experimentation with hop-by-hop (back pressure) congestion control Routing, management SDN is widely used within/among organizations’ datacenter Place related services, data as close as possible (e.g., in the same rack or nearby rack to minimize tier-2, tier-1 communication) fi Questions?