1. Link Aggregation - IEEE802.3ad Dov Litmanovitz (IBM) Tali Atzmon (Cisco) Hagai Avrahami (Elbit) Beni Zanguri (Attune-networks) Guided by: Dono van-Mierop (IBM) Itai Dabran (Technion)

2. General System Description Constraints Open Issues & and suggested ideas Simulation Description Summery Contents

3. Link aggregation, or trunking, is a method of combining multiple physical network links into a single logical link for increased performance. IEEE standard currently implemented or evaluated by: 3Com, Bay Networks, Cisco Systems, Extreme Networks,Hewlett-Packard, and Sun, and more... General

4. General View

5. Increased Bandwidth Optimal Load Sharing ( in order to achieve the desired bandwidth with the minimum resources ). Flexibility ( point-to-point carrying capacity of a link can be scaled according the bandwidth required ). Transparency ( Virtual-Mac, At Layer 2, the 802.3ad standard shows a single logical MAC address to other protocols in the stack ). Goals

6. standardization ( provide the benefits of vendor conformance and uniformity, enabling customers to build multi-vendor networks using robust link aggregation technology ). Generic ( minimum constraints on upper & lower layers, sometimes, tradeoff with performance ). Fast Deployment ( There is no need to modify network topologies or alter the infrastructure ). Goals (cont)

7. Plug & Play ( automatic configuration, and automatic “counter switch” identifying ). By Product advantages (avoid higher routing protocols from seeing loops,...) Goals (cont)

8. Like many standards, 802.3ad allows latitude in vendor implementation. The standard gives definition to the devices and functions to be performed, but much of the how is left to the vendors, revealing vendor competencies. 2 main open issues: –Aggregation algorithm / method. –Distributor algorithm. Open Issues

9. Operates at Layer 2. Actually a new “Proxy” sub-Layer: –Higher layers, Calls here “MAC Client” (like IP...) “think” they are talking with the Ethernet Protocol and operates on it, the regular requests. –Practically, the link aggregation layer talks with the MAC layer. Orientation - Protocol Stack

10. Protocol Stack (cont)

11. Conversation - Identifies a particular host- to-host communication. Generic goal constraint => frames belongs the same conversation must NOT bypass each other. (can not depend on higher layer re-ordering). Generic goal constraint => no knowledge of higher layer srtucture Conversation

12. Conversation Definition - (src MAC-Addr, dest MAC-Addr) better performance could be achieved if a higher lever assumed (for example TCP/IP): (src IP, src port, dest IP, dest port). Performance - generic trade-off. Conversation (cont)

13. System Block Architecture 3 main parts (As Everything in life..) : LAC - Link Aggregation Control (control and configure the whole Link-Agg sub-layer, using LACP protocol). Upper Tier ( maps physical ports to an aggregate port, where users are grouped by multiplexing ). Lower Tier ( deals with the physical ports on a switch ).

15. Aggregator: (standard 802.3 interface) –Frame Distributor –Frame Collector ( must be simple - no reordering ) –Marker Protocol: optional Generator/Receiver compulsory Responder –Aggregator Parser/Multiplexer Control parser/Multiplexer LACP System Block Architecture(cont)

16. A standard IEEE802.3 MAC address is assigned to –The System –Each Aggregator –Each port. The MAC client “see” the aggregator MAC address. Addresses

17. Maintain configuration information: –established by management, and –dynamic state (fail, working,...) Exchange configuration information - with other systems to allocate the link aggregation group. Creates Aggregation Groups (each group contains links terminates at the same pair of systems, with the same characteristics). Link Aggregation Control (LAC)

18. Attach/Detach a Port to a group. Enables/Disables the collector/Distributor Link Aggregation Control (cont)

19. KEY - The ability of one port to aggregate with another is summarized by a simple integer parameter. May reflect: –port physical characteristic –Static administrator configuration constraints. –Other.. Links may be aggregated iff their keys are identical (= group KEY). LAC - link aggregation

20. the protocol is actually more generic then practically required. Open Issue: link aggregation method: –statically: by switch administrator. –Dynamically: by LACP - different algorithms. –Hybrid: system administrator determine constraints, and LACP dynamically aggregates no more then x ports per group a specific port can not be aggregated … LAC - link aggregation (cont)

21. LACP - provide a standard mean for exchanging information between partners in order to reach an agreement about each - and mutual configuration. Transmits: Information and state ( rather then commands ). LACPDU - special frames exchanged by sides. LACPDU contains 128 octets of: –actor information (my state), –partner information (what do I “think” that my partner state is). LACP

22. Active / Passive mode. –Active - Periodic transmission. –Passive - only when configuration changes (ntt - Need To Tell). State machine design. LACP (cont)

23. Objective: optimal Load-Sharing-balancing: the sum of the links bandwidth equals the required bandwidth. Map dynamically each conversation to a port. Constraints: –frames belongs the same conversation must NOT be mis-ordered. –The algorithm shell not cause duplication of frames. Open Issue. Distributor

24. First Problem: measuring dynamically the load of each conversation. Possible Solution: –Aging functions (like routing table..) Distributor (cont) Quant 1 Quant 2 Quant 3 Quant 4 Quant 5 :Accumulative

25. Second Problem: Assuming we know each conversation load (and the links max load). How hard is to map each conversation to a port for optimal result ? Optimal solution is Hard. How much hard ? –NP-hard. Distributor (cont)

26. Reduction to the backpacks problem: –N backpacks (N physical link), with different capacities (bandwidth). –X blocks (conversations), with different weight (bandwidth). –Problem: distribute the blocks optimally between backpacks, for maximum accumulated weight. For optimal solution: exponential time. For 100 hosts each LAN… too much time. Distributor (cont)

27. => look for estimation heuristics. First solution: –simple hash function: –(src+dst) MODULO (no of links in aggregation) –advantages : simplicity, no memory need. –Disadvantages : Depends on MAC address numbers. Problem in a link failure/addition. Distributor (cont)

28. Second solution: –constant hash function + result map to port: – k = (src+dst) MODULO (K) –K - constant no. > no of links in aggregator. –Map each result (K possibilities) to a port. –Evaluate k bandwidth (instead of conversation). –Dynamically change, no failure/addition dependency, less memory needed then with conversations. Distributor (cont)

29. third solution: –Evaluate each conversation bandwidth. –Map each conversation to a link (with estimation function). –Computational hard, –lot of memory. –Best accuracy. Distributor (cont)

30. Problem: How to move a conversation from link to another, for load-balancing ? –Wait a while (maximum time of queue length). –Use the marker protocol (Optional impl.) send a marker PDU, with unique Transaction ID. do not send a frame belongs to a conversation until the marker PDU arrived. Move conversation frames to another link. Distributor (cont)