Session Directories and Multicast Address Allocation Mark Handley USC/ISI mjh@isi.edu
A Little History The first multicast session directory was sd from Van Jacobson and Steve McCanne at LBNL in 1992. I wrote sdr in 1994, and eventually it came to replace sd. Both tools multicast periodic messages to a well known group to announce multicast sessions. They also use the information in these announcement messages to perform multicast address allocation. Van Jacobson briefly described an address allocation scheme for sd in his Sigcomm 1994 tutorial, and this work stems from that original idea.
Basic Address Allocation Idea Sdr announces which multicast addresses are in use. To allocate an address, simply listen to know which addresses are used, choose randomly from those not in use, and announce the address you choose. Problem: Not all sessions you might clash with can be seen and so avoided. If I allocate a global scope address, it can clash with local TTL-scoped sessions elsewhere.
TTL Scoping Sender sets IP "time-to-live" field in packets. Each router decrements TTL. When TTL reaches zero, packet is dropped. TTL constrains how far a multicast packet can travel. Configured TTL thresholds in routers supplement the basic mechanism to divide the Mbone into appropriate scope-zones such as "Europe" or "site-local". When a session is created, the TTL to decide its scope is chosen. Announcement packets go to the same scope as the session they announce.
Informed Partitioned Random Multicast Address Allocation (IPRMA) To avoid clashes, split the address space into different ranges for each scope: of Allocation Probability ttl range ttl range ttl range ttl range ttl range ttl range 1-15 15-31 32-47 47-63 64-127 127-255 Address Range Prob. of Allocation of an ttl range address 64-127 for a session with a ttl in the range 64-127 224.2.80.0 224.2.100.255 Address Range
The Problem with IPRMA Not all locations choose the same TTL thresholds: Within Europe, TTL 48 divides countries. In North America, TTL 48 thresholds are not used. How then do you partition the address space? Too many ranges result in many empty ranges and poor utilization. Too few ranges result in failures of the informed mechanism.
Informed Mechanism Failure Multicast Addresses Allocated before Clash R ds1 R ds2 R ds3 1000 IPR 7 R ds4 IR ds1 IR ds2 IR ds3 IR ds4 Allocations before clash IPR 3-band ds1 IPR 3-band ds2 IPR 3-band ds3 IPR 3-band ds4 IPR 7-band ds1 IPR 3 IPR 7-band ds2 IPR 7-band ds3 IPR 7-band ds4 IR 100 R 100 1000 Address Space Size
Adaptive Address Space Partitioning If we don’ t know in advance the TTL values used worldwide and the distribution of sessions between them, it makes sense for the partitioning to be adaptive: Grow a partition when many addresses are assigned from it. Shrink unused partitions. Van first proposed this, but the difficult part is doing it without having a partition at one site grow to overlap a different scope partition at another site. When this happens the "informed" part of the mechanism fails.
Deterministic Adaptive Partitioning To perform address allocation so that heterogeneity of distribution of sessions between partitions does not cause partition overlap: Allocate numerous partitions at the high-TTL end of the space Have occupied partitions "push" the lower-TTL partitions down the space. Have inter-partition gaps between occupied partitions to cope with short-term variations of allocation.
Deterministic Adaptive Partitioning probability allocation of address range probability allocation of address range probability allocation of Bands with TTL Band with TTL address range less than t greater than or equal to t
Simulations of Adaptive IPRMA Adaptive IPRMA is designed to work well where the number of addresses allocated from a partition is to some extent stable. We can’ t simulate this by allocating until a clash occurs. New criterion: An address allocation scheme is acceptable if during the mean lifetime of a session, the probability of an address clash anywhere in the world is less than 50%.
Simulations Rough Algorithm 1. Allocate n sessions; reallocate until no clashes. 2. Delete one session. 3. Add one session. 4. Repeat from 2 until n sessions have been replaced. Simulations without communities: New session is unrelated to old session. No stability in number of "local" groups Simulations with totally stable communities: New session has same TTL and originator as old session. Group distribution seen from each site fixed.
Simulations: No local communities AIPR-1 (20% gap) AIPR-2 (50% gap) AIPR-3 (60% gap) AIPR-4 (70% gap) Allocations before clash probability > 50% AIPR-H (hybrid) IPR 3-band IPR 7-band 1000 100 100 1000 Address Space Size
Simulations: Stable communities Allocations before clash probability > 50% AIPR-1 (20% gap) AIPR-2 (50% gap) IPR 3-band 1000 IPR 7-band 100 100 1000 Address Space Size
Reality? How can we simulate the locality of future use? We can’t we could only speculate. The two simulation sets give lower and upper bounds for performance. Both scale O(n). Static allocation slightly outperforms adaptive allocation if we know the number of active partitions and their boundaries and their are no communities. In practice we don’ t know this. Thus the adaptive algorithm works well given the constraints.
Packet Loss and Propagation Delay The previous simulation ignores propagation delay and packet loss. Assume: mean session length: 2 hours mean advance announcement: 2 hours mean end-to-end delay: 200ms mean packet loss rate: 2% each announcement resent every 10 mins Mean end-to-end delay ~(0.98*0.2)+(0.02*600) = 12 secs Therefore approximately 0.1% (12 secs/4 hours) of sessions are not visible at any time.
Packet Loss and Propagation Delay Addresses allocated before probability of a clash is 0.5 1e+06 Upper Bound (y=x) Lower Bound (y=x^0.5) i=0.01m i=0.001m 100000 i=0.0001m i=0.00001m Addresses Allocated 10000 1000 100 10 100 1000 10000 100000 1e+06 Address Space Size
Packet Loss Instead of periodically announcing a session with a fixed base rate, start with a high rate (e.g. 1 per second) and exponentially back off down to a fixed base rate. With global propagation delay, this brings us to approximately i =0.0005 m Scales well to around 10000 addresses allocated in a partition. This is not good enough!
Two Solutions Introduce hierarchy to address allocation. Do distributed address-range allocation between domains on a long timescales to mitigate propagation delays and loss. Do distributed individual address allocation within a domain on short timescales from the address-ranges. Add scalable mechanisms to detect and correct clashes when they do occur.
Detecting Clashes We’ d like any address allocator that notices a clash to be able to report it if the originator doesn’ t. There may be millions of allocators. How do we avoid clash report implosion? SRM (Sigcomm ’ 94) uses random timers chosen uniformly from a range [ d1 : d2 ]. d1 and d2 are chosen based on measurements of propagation delay. One multicast response suppresses others. What can we do with no estimate of the population and no delay matrix?
Exponential Random Timers Uniform random timers are too dependent on knowing the number of nodes trying to respond. If we use exponentially distributed random timers we can largely remove this dependence.
Simulation: Uniform Random Timers A: Shortest path tree, delay=~distance B: Shared tree, delay=~distance C: Shortest path tree, delay=distance+random D: Shared tree, delay=distance+random No of responses 1000 100 10 51200 25600 12800 6400 3200 1 No of sites 1600 200 800 800 3200 12800 51200 2.0e5 8.2e5 3.3e6 1.3e7 400 200 D2 (ms)
Simulation: Exponential Random Timers Mean no of responses No of responses 1000 100 10 51200 25600 12800 6400 3200 1 1600 No of sites 200 800 400 800 1600 3200 640012800 400 200 25600 51200 D2 (ms)
Summary We’ ve designed a scalable approach for adaptive partitioning of the address space to deal with allocating addresses to TTL scoped sessions. The real problem is propagation delay. Solutions: Increase speed of light. Use a hierarchy with the upper levels operating on different timescales. Move away from TTL scoping. Within a lower level allocation domain: Use exponential backoff between announcements Use exponentially distributed random timers to send clash detection messages.
Recommend
More recommend
