Check out the new USENIX Web site. next up previous
Next: Optimization 3: exploiting physical Up: Efficient update flooding Previous: Optimization 1: delta propagation


Optimization 2: harbingers

Flooding guarantees reliable delivery by propagating updates (deltas or full contents) over multiple links at each step of the algorithm. Thus, it consumes $m$ times the optimal network bandwidth, where $m$ is the number of edges per replica. Harbingers eliminate redundant update deliveries.

Pangaea uses a two-phase protocol to propagate updates that exceed a certain size (1KB). In phase one, a small message that only contains the timestamps of the update, called a harbinger, is flooded along graph edges. The update bodies are sent, in phase two, only when requested by other nodes. When a node receives a new harbinger, it asks the sender of the harbinger (the immediate upstream replica in the flooding chain) to push the update body. Simultaneously, it forwards the harbinger to other neighbors in the graph. When a node receives a duplicate harbinger without having received the update body, it asks its sender to retry later. This is required because the sender of the earliest harbinger may crash before sending the update body. If a node receives a harbinger after having received the update body, it tells the sender to stop sending the update. We chose the harbinger threshold of 1KB, because we found that delta sizes follow a bimodal distribution--one peak around 200 bytes representing directory operations, and a flatter plateau around 20KB representing bulk writes.4

This harbinger algorithm not only saves network usage, but also shrinks the effective window of replica inconsistency. When a user tries to read a file for which only a harbinger has been received, she waits until the actual update arrives. Since harbinger-propagation delay is independent of the actual update size, the chance of a user seeing stale file contents is greatly reduced.


next up previous
Next: Optimization 3: exploiting physical Up: Efficient update flooding Previous: Optimization 1: delta propagation
Yasushi Saito 2002-10-08