On Multicast

Download Report

Transcript On Multicast

On Multicast
?
CS614 - March 7, 2000
Tibor Jánosi
What to Expect
•
•
•
•
•
•
•
Motivation / Why is it difficult?
IP Multicast Routing
Reliable Multicast Transport Protocol
Scalable Reliable Multicast
Light-Weight Multicast Services
Pragmatic General Multicast
PGM/LMS Comparison
Motivation
• Multimedia streams: live and not.
• Financial data distribution.
• Distributed fault tolerance.
All these: Same basic communication pattern,
but have widely different requirements.
Why Difficult?
•
•
•
•
•
•
Very different application needs.
Limited bandwidth and processing power.
Poorly known/changing network topology.
Hard to deploy changes in routers.
Large/unequal/changing propagation delays.
Unclear what “best” policy means in
various contexts.
• Etc...
Ideal Multicast
• Senders (S) and Receivers (R) not aware of
each other’s position in the network.
• Scalable.
• Low latency (join, data propagation).
• Low bandwidth and processing overhead.
• “Reliable”, if this is cheap (“end-to-end”?)
• Easy to join/leave.
IP Multicast: Basic Idea
• Multicast groups: abstract “rendez-vous”
points.
• Set up optimal spanning tree spanning
participants for each group.
• Make it cheap by not providing strong
guarantees: send out packets and hope for
the best. Not that bad, in fact.
Big question: Who gets which packets?
Send everything to everybody. You get …
Invent Multicast Routing, (try to) forward
only what’s needed, when needed!
LAN Multicast: IGMP
• Queries/Replies
• Random delay before reply.
• Don’t report multicast groups already
reported.
• Router will know groups with members on
its LAN.
Reverse Path Forwarding
From shortest path to S
From other path
router
router
How do we determine the shortest path to the source?
One possibility: routers exchange distance info.
Also, duplicate packets possible.
Idea: Pruning
prune
receiver
sender
Core Based Trees
CBT(2)
•
•
•
•
All senders could be sending to the Core.
Single point of failure.
Core address must be known; fallbacks also.
Each router has to know only which
interfaces to send packets on. Cheap.
• Join/leave explicit. No need to wait.
• No pruning.
Protocol Independent Mcast
• Two styles: sparse and dense.
• Dense: flood and pruning.
• Sparse: much like CBT: join a “rendezvous” point. Receiver’s routers can identify
“shortcuts”.
• No need for data to pass through rd point.
• Rd points send “alive” packets. Receivers
will switch to alternative, if rd’s dead.
PIM (2)
Reliable Mcast Transport Protocol
Smart “session manager”
elects DR’s and sets
parameters. How? Just
like that...
-S, R use windows
-Designated Receivers
eliminate ACK implosion
-ACK’s sent to DR’s
-DR’s and S cache data
and retransmit it when
needed.
RMTP(2)
• After set up S starts sending data. Receivers
send periodic ACK’s after first packet
received.
• If no ACK’s for a long time, connection
terminates.
• DR’s or S retransmit info using unicast or
multicast, depending on number of errors.
• Immediate TX request sent to DR’s, for
receivers that join the session.
RMTP (3)
• Sender window advance determined by
slowest receiver.
• ACK’s must not be repeated too often.
Measure RTT to AP.
• S adjusts (decreases) send window to 1 if many
errors; then increases linearly.
• DR’s are fixed, but each R chooses its DR.
(DR sends SND_ACK_TOME with TTL fixed
to a known value).
Scalable Reliable Multicast
• ALF: explicitly include app’s semantic in
protocol design. No solution will work for all.
• Data identified by unique, persistent names.
• Source id’s are persistent.
• IP Multicast is available.
• Data conventionally grouped in “pages.”
• No distinction between receivers and senders.
• These assumptions fit wb’s semantics.
SRM (2)
“Session messages” (SM) multicast periodically.
Used to:
• advertise sequence number of active page for
active sources (data grouped in “pages” to
limit history)
• determining set of participants
• estimate one-way distance between nodes
SRM (3)
• Loss signalled by multicast NACK’s with
persistent, unique name.
• NACK preceded by randomized wait.
• If waiting for data: Wait timer reset w/ time
doubled if NACK for same data received or
timer expired.
• If have data when NACK is received:
Randomized wait, then multicast repair data,
unless somebody else did during this wait.
SRM (4)
• Wait periods drawn from uniform distributions
on intervals w/ length dependent on distances
between hosts and (almost) arbitrary constants.
• These constants depend on topology and
network conditions. They should be adaptive.
• Leaving the group is indistinguishable from
being in inaccessible partition.
• No partial/total ordering provided; but these
could be built on top of SRM.
SRM(5)
Performance much better if local recovery is
possible (no need to multicast to everybody).
Solutions:
•TTL-based scoping (one step and two step);
•Separate multicast group for recovery;
•Administrative scoping.
SRM(6): Extreme Topologies
Deterministic supression:
•exactly one NACK;
•exactly one repair.
Probabilistic supression:
•at most g-1 requests, always
expect more than one;
•the longer the interval the fewer
the requests, but latency bigger.
Light-Weight Multicast Service
LMS modifies standard router forwarding. Achieves
minimal overhead, no pathological behavior,
minimal latency.
Three new functions are needed in router:
•Select a replier for each subtree.
•Send request to replier corresponding to subtree.
•Multicast replier’s repairs to loss subtree (subcast).
LMS (2)
Routers pick a “replier link”
from list of available links.
Router state added:
•upstream link;
•list of downstream links;
•replier link.
Problem: Same replier could
be picked many times.
LMS (3)
•All receivers detecting
loss send repair requests.
•Routers forward requests
to replier.
•If replier does not have
data, it sends a request
also.
•Replier’s request is
forwarded uplink.
LMS (4)
LMS (5)
Duplicate sends are
possible: select replier
with least amount of
loss.
Repliers can fail: receivers
will time out waiting for
the replier. They will ask
for a new replier to be
elected.
LMS (6)
Duplicate requests are possible (if everybody picks the same
replier). But we can protect against this.
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
R
LMS (7): Bad Replier Choice
Pragmatic General Multicast
• Somewhat similar to LMS.
• All retransmissions originate from the source.
Designated Local Retransmitters might help,
but they must be on the path.
• Receivers send NACK’s back to source; and
repeat it until they get a confirmation
(NCF).
• NCF’s inhibit NACK’s from other
receivers.
PGM (2)
• Only one NACK per (source, packet) is
propagated upward.
• S (or DLR) send RDATA when they get
NACK. RDATA retraces path of NACK’s.
• In routers: no NACK, no pass!
• “Dangling NAK state”: RDATA lost, first
NACK in router, subsequent NACK’s rejected.
• A retransmission only reaches those that have
requested it, but not necessarily all of them.
PGM (3)
PGM (4)
• “Repeated Retransmission”
LMS Latency
Loss at source, random topology.
LMS Latency (2)
Loss at source, random topology.
PGM Latency
Loss at source, random topology.
PGM Latency (2)
Loss at source, random topology.
PGM Latency (2)
Loss at source, random topology.
Random Loss: LMS vs. PGM
LMS picks replier that is farther than the source! Topology!!
LMS Duplicates
PGM Retransmissions
Questions? Comments?
Thank you!