Floodless in SEATTLE A Scalable Ethernet Architecture for Large Enterprises By Changhoon Kim, Ma/hew Caesar, and Jennifer Rexford

Transcription

1 Floodless in SEATTLE A Scalable Ethernet Architecture for Large Enterprises By Changhoon Kim, Ma/hew Caesar, and Jennifer Rexford Presented by: Charndeep Grewal Department of Electrical Engineering

2 MoCvaCon IP Networks require significant effort to configure and maintain. Ethernet is very simple, but not does scale well past local networks. SEATTLE is an alternacve network architecture that provides the scalability of IP and the simplicity of Ethernet.

3 Background Ethernet Bridging Each host has a MAC address that is used to send packets. Bridges use a forwarding table to store this informacon, if address not found, flooded on network. Problem Forwarding tables grow in size proporconal to the number of hosts. Control informacon starts comprising a large overhead in network traffic. Malicious hosts can also cause network wide floods.

4 Background Spanning Trees To prevent broadcast storms due to flooding, administrators create a root bridge, then other bridges colleccve compute a spanning tree based on distance to root to route packets. Problem Spanning trees cannot grow large in size. It forces all traffic to traverse single spanning tree, which makes forwarding error prone. Load is also higher on links near the root bridge.

5 Background Bootstrapping Protocols Bootstrapping protocols such as ARP and DHCP rely on broadcascng. DHCP discovery messages are broadcast when host thinks the network a/achment point is changed. Broadcast APR are generated whenever host needs to know MAC address. Problem The problem is broadcast messages must be processed by end hosts, owen not handled by network card but CPU. For portable devices ARP can consume significant fraccon of bandwidth for just handling ARP messages. They also introduce security issues.

6 MoCvaCon The current solucon is to bring IP roucng into Ethernet, but this breaks many of the desired properces. Network administrators now have to subdivide their address space to assign IP prefixes across topologies and subnet. This leads to wasted address space. Large overhead in configura=on, some es=mates put 70% of an enterprises network opera=on cost as maintenance and configura=on. Ensuring service concnuity across locacon changes like virtual machine migracon, which is becoming more prevalent, is challenging.

7 SEATTLE Is it possible to build a protocol that maintains the same configuracon free properces as Ethernet bridging, yet scales to large networks?

8 SEATTLE Open hop, link state, network layer DHT Use a network layer distributed hash tables (DHT) to form a flexible directory service to provide address resolucon. Traffic driven loca=on resolu=on and caching Switches cache responses to queries because typically communicacon is between a small number of hosts. Scalable, prompt cache update protocol Update process is triggered by network layer changes, not Cmeouts or broadcasts.

9 One hop DHT Mapping between MAC/IP address is done through a distributed directory service. A network level approach is taken where mappings are stored at switches to make network recovery awer failure fast and efficient as well as avoid control overhead.

10 One hop DHT SEATTLE enables shortest path forwarding by running a link state protocol. SEATTLE s link state protocol maintains only the switch level topology. SEATTLE runs a discovery protocol to determine which links are a/ached to hosts and which are a/ached to switches. DisCnguishing between different kind of links is done by sending control messages that Ethernet hosts do not respond to.

11 Hashing key values into switches To publish a (k,v) pair, a switch first uses a hash funccon F to find out which switch it will map to k to, using a switch idencfier F(k) = r k. (k,v) F(k) r k It then tells switch r k to store the mapping (k,v). r k becomes the resolver for k.

12 Hashing key values into switches A different switch may look up the value of k by using the same hash funccon to idencfy which switch is the resolver for a different k; because all switches know all the other switches idencfiers via link state a advercsement.

13 Topology Changes When a switch fails, the link state map changes. That means the old resolver r k old is different than r k new. To deal with this the original switch (k,v) that published k, monitors the liveliness of the resolver. When s k determines r k old and r k new are different it republishes to r k new and the old (k,v) is removed form r k old awer a Cmeout.

14 MulC level, one hop, DHT The above approach doesn t work if you want to provide strong fault isolacon or divide up administracve control. A network is divide into several regions and each region is connected to the backbone via it s own border switch like OSPF.

15 Host Address ResoluCon In convenconal Ethernet, hosts broadcast ARP request to lookup a MAC address. In SEATTLE switches used F(ip). Host A arrives at switch s a. Switch learns ip a, the IP address of host a. Computes F(ip a ) = v a, where v a is the idencfier of another switch. s a then tells v a of (ip a, mac a ) then v a becomes the IP resolver for a. Not that the hosts locacon resolve F(mac a ) might be different than the locacon resolver F(ip a ).

16 Host LocaCon ResoluCon

17 Handling Host Dynamics Host may change physical loca=on (wireless) Host h moves from s h old to s h new. s h new does a insert MACto locacon entry. MAC may change (NIC replacement) h s access switch s h inserts IP to MAC entry contain h s new MAC address. Change of IP (DHCP lease over) s h deletes old IP to MAC mapping and inserts new one.

18 Backward CompaCbility From end hosts perspeccve, the way hosts interact with the network to bootstrap themselves must be the same as Ethernet.

19 Bootstrapping Hosts To discover a new host MAC the same learning mechanism discussed earlier is used but it is only enabled on ports connected to end hosts. SEATTLE switches snoop gratuitous ARPs. If no gratuitous APR was issued, it an scll get IP address by snooping DHCP messages. When a DHCP discovery message is received, it delivers the answer to the DHCP server via unicast.

20 SimulaCons Network administrators won t give up informacon. Used real traces from Lawrence Berkeley NaConal Lab as well as synthecc traces made on assumpcons on workload characterisccs. 4 sets of trace were collected over a period of minutes. To and from ~9,000 end hosts distributed over 22 subnets.

21 4 Topologies for SimulaCon 1. Campus: ~40k students, 517 routers 2. AP Small: Small access provider, 87 routers 3. AP large: Large network, 315 routers 4. DC: Data Center of 4 full meshed core routers each connect to a mesh of 21 switches.

22 SensiCvity to cache eviccon Cmeout If a switch removes host informacon before locally a/ached host does the switch needs to perform locacon lookup to forward data packets. Therefore the simulacon was to evict unused entries awer a Cmeout then look at how many packets needed locacon resolucon for AP large topology. The simulacon showed that even for small Cmeout values of 60 seconds, 99.98% of packets were forwarded without a lookup. Timeouts larger than 600 seconds needed 0 lookups.

23 SensiCvity to cache eviccon Cmeout If a switch removes host informacon before locally a/ached host does the switch needs to perform locacon lookup to forward data packets. Therefore the simulacon was to evict unused entries awer a Cmeout then look at how many packets needed locacon resolucon for APlarge topology. The simulacon showed that even for small Cmeout values of 60 seconds, 99.98% of packets were forwarded without a lookup. Timeouts larger than 600 seconds needed 0 lookups.

24 Forwarding Table Size Event though cache increases table size by 1.5, it scll gives a 16x smaller table than Ethernet and no caching results in a 22x smaller table.

25 Control Overhead SEATTLE is much be/er here since it uses unicast and does not use network wide flooding. Caching degrades performance again, but using caching only increases control overhead from 0.1 to 1 packet per second. Even with that hit, it s 1000x be/er than Ethernet.

26 Effects of Network Charges In this simulacon they caused switches to fail randomly. SEATTLE wins here because it can use all the links in the topology to forward packets. While Ethernet can only forward over a spanning tree. Also awer switch failure the tree must be recomputed.

27 Effects of host mobility For high mobility, Ethernet is worse. This happens because Ethernet takes Cme to evict stale locacon informacon and learn new host informacon. Some hosts broadcast gratuitous ARP but this increases overhead.

28 ExperimentaCon Results SEATTLE requires less overall packet processing Cme* on paths longer than 3.03 switches level hops. A typical campus network is 4 hops for vast majority of host pairs. * Packet processing Cme is from the Cme the packet arrives to the switches inbound queue and to the Cme its ready to be moved to the outbound queue.

29 Forwarding Table Size Measured over a one second interval, Sea/le scll has a smaller forwarding table size (1.5x).

30 Control Overhead Spikes are due to scanning a/acks. Ethernet uses network wide floods. SEATTLE uses 1 unicast lookup to resolver then discards packet.

31 Returning to the quescon posed Is it possible to build a protocol that maintains the same configuracon free properces as Ethernet bridging, yet scales to large networks?