Amoeba Group communicate 1-to-n Members join and leave group - - PowerPoint PPT Presentation

Amoeba Group Communication

CS294-4 P2P Systems 2003 David Ratajczak

Group Communication

Group is a collection of processes that can

communicate 1-to-n

Members join and leave group Ordering

All members see same (total) order

Broadcast messages Join/leave changes Crash/recovery

Weaker ordering possible

Unordered Causal order

Group Communication (cont’d)

Reliability

Handles message loss/duplication Handles failstop failures

Unreliable failure detection based on heartbeats Might excise non-failed nodes Multi-phase protocol for calculating new view

In Amoeba, optional support for recovery

But no built-in support for state transfer No automatic rejoining after network partitions

Amoeba GCS API

All primitives are blocking

Use threads for parallelism Claim: this simplifies programming

Question: how does an application create

a group?

Questions to think about

What about end-to-end argument? How does this apply to p2p systems? How does this compare to Castro/Liskov

BFT algorithm?

Amoeba Total Ordering

Every group view includes a designated

sequencer

If sequencer fails, must produce new view

PB method

Sender sends msg to sequencer Sequencer affixes sequence number Sequencer broadcasts to group

Amoeba Total Ordering (cont’d)

BB method

Sender broadcasts to group Sequencer receives msg and broadcasts

“accept” with sequence number to group

If messages are large, use BB

Less bandwidth and less stress on sequencer

Otherwise, use PB

Receivers only interrupted once

Amoeba Reliability

Use NACKs instead of ACKs

Fewer messages at sequencer Assumes steady stream of traffic

Sequencer stores msgs until all members have

received them (or view changes)

Members piggyback latest seq num recv’d

Settable resilience degree

Msg is tentative until sequencer hears r ACKs Sequencer broadcasts accept msg

No flow control

Latency breakdown

20 Mhz MC68030 10 Mbit/s Ethernet What is the

bottleneck?

What if no multicast?

Delay Throughput Resilience

Questions to think about

Can this scale to >100s of nodes? Will this work in a WAN? What about with resilience degree >0?